DEV Community: Prabhash Jha

Who Owns the Ad Account? The Access Audit Every Founder Should Run Once a Year

Prabhash Jha — Sun, 27 Sep 2026 14:40:56 +0000

Most founders find out who owns their ad account on the worst possible day. Not during a quarterly review. Not while reading a contract. On the afternoon they decide to change agencies, and discover that the pixel with three years of conversion history sits inside a business portfolio they've never had a login for.

The uncomfortable part? Nobody did anything wrong. No agency set out to hold anything hostage. Someone was in a hurry during onboarding, clicked "create new" instead of "request access", and the asset was born in the wrong place. Three years later that one click is the reason a transition that should take an afternoon takes six weeks.

This is an audit, not an accusation. Run it once a year, on a quiet afternoon, whether or not you're happy with everyone you work with. The point isn't to catch anyone. Ownership questions are trivially cheap to fix while the relationship is good. Nearly impossible to fix while it's ending.

What "owning" an account actually means

Owning an account means you hold the top-level administrative role on the container that holds the asset, using credentials on an email address at a domain you control.

Three separate conditions in one sentence. Most founders only satisfy one.

The container, not the asset. In almost every ad platform, the thing that matters isn't the ad account. It's the organisational shell the ad account lives inside. Meta calls it a business portfolio. Google calls it a manager account. If your agency's shell owns your ad account, then removing the agency removes the account. Because the account was never a separate thing that could be handed back.

The top-level role. Being able to log in and see campaigns isn't ownership. Google Ads has five distinct access levels: Email-only, Billing, Read-only, Standard and Admin. Only Admin can grant access, change other people's access levels, and manage manager-account links. Everything below Admin is a permission someone else can revoke. Google's own documentation makes the same point in reverse. An account with only one administrator risks losing tag access entirely if that person becomes unavailable. Which is why two admins is the floor, not a nicety.

On an email you control. An admin role attached to founder@gmail.com is better than nothing, and worse than it looks. If the person holding that address leaves, disputes the separation, or simply stops answering, you own the account in theory and not in practice. Every administrative role should sit on an address at your own domain. Ideally one that survives an individual's departure.

The assets to check, and who should hold each

Work the list top to bottom. The order isn't arbitrary. Items near the top can be used to recover items further down. So an audit that starts at the bottom can waste a week proving things you could have taken back directly.

Asset	Who should own it	The failure mode
Domain registration	The company, at its own registrar account	Registrar account in a developer's or agency's name
DNS / nameservers	The company	Hosted inside an agency's account "for convenience"
Business email (Workspace / M365)	The company, super-admin held internally	Single super-admin who has left
Meta business portfolio	The company; agency added as a partner	Portfolio created by the agency, assets born inside it
Meta pixel / dataset	The company's portfolio	Pixel owned by agency portfolio, history non-transferable
Google Ads account	The company; agency links via manager account	Account created under the agency's manager account
Google Analytics property	The company	Property in an agency-owned account, data not portable
Google Tag Manager container	The company	Container in an agency account; site tags depend on it
Search Console property	Company holds verified ownership	Only the agency is a verified owner
Payment methods on ad accounts	The company's card or billing profile	Agency card, so spend history and credit sit with them
Social accounts	The company, with 2FA recovery internal	2FA bound to a departed employee's phone
CRM / marketing automation	The company	Seat-based tools where the admin seat is an agency seat

Two rows deserve more than a line each.

The pixel and the analytics property are the ones with real, non-recoverable value. A campaign can be rebuilt in a day. Ad creative can be re-uploaded. Three years of conversion history, audience seeds and learned optimisation signals cannot be exported and re-imported into a fresh pixel. You can move who has access to a dataset. You cannot move accumulated learning into a new one. This is why the pixel is the asset worth checking first, even though it feels like a technical detail.

Search Console has a distinction that catches people out. There is a difference between a verified owner, a delegated owner, and a full user. A delegated owner's status depends on a verified owner continuing to exist. If the verification token that made the agency an owner is removed, and you were never independently verified, you can lose the property. Verify ownership yourself, through your own DNS record. So your verification doesn't depend on anyone else's. If you want the wider picture on what that data can and cannot tell you once you hold it, your Search Console numbers are lying to you in five specific ways covers the interpretation side.

How to actually check, platform by platform

Reading a contract tells you what was agreed. Only the platform tells you what is true. Every check below takes under two minutes.

Meta. Open Business Settings in your business portfolio. Under Accounts, then Ad Accounts, look at each account and check which portfolio is listed as the owner rather than which people appear in the user list. Do the same under Data Sources for your pixel or dataset. If your portfolio isn't the owner, you have partner access to your own asset. Under Users, then Partners, you should see your agency listed as a partner with specific tasks assigned. That's the correct arrangement.

Google Ads. Go to Admin, then Access and security. Confirm at least two people with Admin access, both on your domain. Then open Account settings, then Managers, to see every manager account linked to yours. A manager account link is how an agency should be connected. It's also something you can unlink yourself, which is exactly the property you want.

Google Analytics. In Admin, check Account access management as well as Property access management. They're separate. Someone can hold power at the account level that's invisible from the property view. Confirm you hold the Administrator role at account level.

Tag Manager. Check both account-level and container-level permissions, and specifically who holds Publish rights. Publish is the one that matters. Whoever holds it can change what runs on every page of your site.

Search Console. Settings, then Users and permissions. Confirm you appear as an Owner. Check Ownership verification to see how that ownership was established.

Write down what you find before you change anything. Honestly, half the value of this audit is having last year's answer to compare against.

The order to fix things in, and the one sequencing trap

Fix in this order: domain, then email, then everything else.

The domain is first because it's the recovery mechanism for almost everything downstream. Control of DNS lets you re-verify Search Console independently, prove ownership to platform support teams, and re-establish email if you have to. If the domain sits in someone else's registrar account, nothing else you fix is durable.

Here is the trap. It's a real one. Under ICANN's Transfer Policy, registrars must apply a 60-day lock preventing transfer to another registrar following a change to the registrant's information. A domain also cannot be transferred during the first 60 days after initial registration, or after a previous registrar transfer. So if you update the registrant details to your company name first, and then try to move the domain to your own registrar account, you can find yourself locked out of the move for two months. By a rule that exists to protect you. ICANN's own guidance is explicit about the sequence. If the goal is to transfer the domain, complete the transfer first, then change the contact information.

Some registrars allow the prior registrant to opt out of that lock. They're not obliged to. And the opt-out has to be exercised before the change request, not after. Assume you won't get it.

Email comes second because password resets for every remaining platform land there. And because a super-admin account nobody at the company controls makes every other fix provisional.

How to move things without breaking a live campaign

The fear that stops most founders from fixing this is reasonable. Pausing a working funnel to satisfy a governance concern is a bad trade. It's also usually an unnecessary one.

Ownership changes and access changes are different operations. Adding your business portfolio as an owner, or adding yourself as an Admin, doesn't remove anyone. You can bring your own control up to the correct level with the agency still fully in place. Nothing stops running. Do that part immediately and independently of any conversation about the relationship.

Do the asset-by-asset transfers while everyone is still cooperating. Meta asset transfers between portfolios need action from both sides. So does unlinking a manager account. Every one of these is a two-minute task for a working relationship, and a support-ticket saga for an ended one.

Never make a tag change and a tracking change in the same week. If you move a container and re-verify a property at once, and conversions drop, you won't know which change caused it. Move one thing. Wait for a clean day of data. Then move the next.

Retire access at separation, don't just intend to. Removing a departing partner's access is a task with a date, not a sentiment. The same discipline applies to employees. If the idea of a written record of who touched what feels excessive, it's the same instinct behind what to log when an AI agent acts on your behalf. The log exists precisely for the situation where memory and goodwill are no longer available.

What to write into the contract so next year's audit is boring

The audit gets easier every year if the paperwork does part of the work. Four clauses cover most of it:

Assets created during the engagement belong to the client. Name the categories explicitly: ad accounts, pixels and datasets, analytics properties, tag containers, creative files, audience lists. Rather than relying on a general "work product" clause that was drafted with documents in mind.
Work happens under client-owned containers via partner or manager access. This is the clause that prevents the wrong-click problem at onboarding. It makes the correct setup the contractual default, rather than a preference someone has to remember.
Access is returned within a fixed number of days of termination, with a named list of what "returned" covers.
Either party can request an access review once per quarter. This is the one people leave out. It's also the one that makes the whole thing routine rather than confrontational. Asking to review access because the contract says you do it every quarter is a very different conversation from asking because you're unhappy.

A good agency will agree to all four without friction. From their side, this is simply the correct way to run an account. The sequencing that a well-run onboarding follows anyway is set out in the first 90 days of an agency account. Reluctance to put any of it in writing is itself information.

Where this connects to the harder conversations

Access is the quiet infrastructure underneath every difficult decision a founder eventually has to make.

If you're weighing whether a relationship has run its course, the honest version of that decision is much easier when you already know that leaving costs you an afternoon rather than a quarter. The signals worth acting on are covered in when to fire a client, the four signals, and why revenue is never one. The mirror image applies to vendors you're thinking of leaving.

And if the relationship is ending badly over money, the access question becomes charged in both directions. The principle I hold to when I'm the one owed money is to withhold labour, not assets. The reasoning is in the client has stopped paying: here is the sequence, in order. The reason to run this audit while everything is calm is so that you never have to find out whether the other party holds the same principle.

The annual version, compressed

Once a year, block ninety minutes and do exactly this:

Open each platform's settings page and record the owning entity for every asset in the table above.
Confirm at least two administrators per platform, both on company email addresses.
Remove every person who no longer works with you, and every partner whose engagement has ended.
Check that 2FA recovery for social accounts points at something the company controls, not an individual's personal phone.
Confirm the payment method on each ad account is the company's.
Diff this year's list against last year's, and ask about anything that moved.

Step six is the one that turns an audit into a system. A single snapshot tells you the state of things. Two snapshots tell you the direction they're drifting in. Which is the thing you actually want to know.

FAQs

Who legally owns an ad account created by an agency?

Whoever controls the container it was created in, unless a contract says otherwise. And platform terms generally govern the operational reality regardless of what a contract says. This is why the contract clause and the platform setup both matter. The contract gives you a claim. The platform setup gives you the account. A claim without control means negotiating for something you already own on paper.

Can I transfer a Meta pixel to a different business portfolio?

You can transfer ownership of a dataset between portfolios, and both sides need to act for it to complete. What you cannot do is move accumulated learning into a brand-new pixel. That history is tied to the dataset itself. This is the practical reason to fix pixel ownership early, rather than accepting a fresh pixel as a workable substitute later.

What is the difference between a Google Ads manager account and account ownership?

A manager account links to your account to administer it. It doesn't have to own it. The correct arrangement is that your company holds the ad account with Admin access on your own emails, and the agency's manager account is linked. You can unlink a manager account yourself. Which is precisely why that structure is safer than an account created inside the agency's manager hierarchy.

How many people should have admin access to an ad account?

At least two, both on company email addresses. One is a single point of failure. Google's documentation warns specifically that a sole administrator becoming unavailable can cost you tag access. More than three or four dilutes accountability without adding resilience.

What happens to my domain if I change the registrant details before transferring it?

You trigger a 60-day lock that prevents transferring the domain to another registrar. ICANN's Transfer Policy requires registrars to apply that lock after a change to registrant information. If your goal is to move the domain into your own registrar account, do the transfer first and update the contact details afterwards.

Should I run this audit if I trust my agency completely?

Yes. Trust is the reason it's easy, not a reason to skip it. Almost every ownership problem I've seen came from a rushed onboarding click rather than bad intent. It stays cheap to fix for exactly as long as both parties are willing to spend two minutes in a settings screen. The audit is insurance against accident, not against malice.

What if the agency refuses to transfer an asset?

Escalate through the platform's support process with your contract and your domain control as evidence. Take the loss on anything genuinely non-recoverable rather than letting the dispute stall the business. Then rebuild what you must under containers you own. It's a bad outcome, and it's survivable. Which is why the sequencing advice above puts domain and email first. Those are the two that make everything else recoverable.

Key takeaways

Ownership is the container plus the top-level role plus an email you control. Two out of three isn't ownership.
Check the platform, not the contract. Settings screens tell you what is true.
Fix in order: domain, email, then everything else. The top of the list recovers the rest.
Transfer a domain before changing registrant details, or eat a 60-day lock.
Raise your own access to Admin today. It removes nobody and breaks nothing.
Two admins per platform, both on company addresses. One is a single point of failure.
Put the quarterly access review in the contract, so asking is routine, not adversarial.
Diff this year against last year. The drift is the finding.

What to Log When an AI Agent Acts on Your Behalf

Prabhash Jha — Sun, 27 Sep 2026 14:40:55 +0000

The reason I know exactly what to log when an AI agent acts on your behalf is this. I once ran a job whose only record of having fired was an untracked file and a broken connection. Seventy minutes of unattended writing. A complete draft. And no evidence anywhere else that the run had ever started. If you'd asked me the next morning whether the agent had done what it was told to, I couldn't have answered honestly. The only artefact I could point to was the one produced by the agent itself. And the one produced by the agent itself is exactly the one you can't trust.

This post is the operating manual that came out of that morning. It isn't the compliance guide the platform vendors are writing. It's the honest small-shop version. The minimum log you need if you've decided to hand a machine the authority to write to your git repository, push to your live site, send an email under your name, or move money in your account. Everything below is from having run agents that do at least the first two of those things, on this website, unattended.

The morning that made me write logs

The first autonomous publishing run on this site was supposed to write three posts overnight and commit them. It died at seventy minutes. The connection that carried the agent's session to my machine dropped. The process on the far side terminated. And the local environment had exactly what a local environment has when a process dies. An untracked file, no commit, no log entry, no evidence that the job had ever started.

I found this out the next morning by looking at git status. There was a file. There was no cover for it. There were no internal links from any older post to it. There was no line in a run log saying "at 21:04 the agent picked topic X". Nothing except a Markdown file with a plausible slug and 3,141 words of prose that could have been written any time in the previous six hours.

The failure wasn't the connection. The failure was that the entire evidence chain for whether the agent had done its job lived inside the agent's own process. When that process died, so did the evidence. I could see that a thing had been written. I couldn't see why it had been written, when, what topic it had rejected on the way, or whether any of the three posts it was supposed to write had actually been attempted. I had to reconstruct all of that from the file itself. Which is the equivalent of asking a defendant to write their own witness statement.

The very next thing I did that day was change the procedure so the run log opened at the start of the run, not at the end. Every step in this post is a scar left by that morning.

Two categories: intent and effect

There are two kinds of thing you have to log when an agent acts on your behalf. Confusing them is where most audit trails go wrong.

Intent is what the agent chose to do. Which action it decided on. Why. What other actions it considered and rejected. And the field almost every vendor logging framework leaves out: what it explicitly refused to do.

Effect is what actually happened as a result. The commit sha. The API response. The file that landed on disk. The row that was inserted. The money that moved.

Intent without effect is a plan. Effect without intent is a mystery. You need both. And you need them cross-referenced by a stable identifier that survives whether the agent finished, crashed, got killed by the operator, or was rate-limited by an upstream API.

The distinction matters because one day, in the middle of the night, you'll get exactly one job that only produced half of these things. When that happens, you need to look at the log and answer, in under a minute, three questions. Did the agent get as far as deciding what to do but crash before doing it (safe, retry)? Did it decide what to do and do some of it (dangerous, investigate)? Or did it do the thing but never get to record that it had done it (worst, because the next run will probably do it again)?

The one artefact you cannot trust

The agent's own summary of its run is unreliable. For three reasons that stack.

First, the summary is written by the same actor whose behaviour it describes. This is a category of evidence a court would exclude. And there's no reason to hold your own audit trail to a lower standard than the one you'd demand of a witness in a room.

Second, the summary is written from inside the process it summarises. If the process died before the summary got written, there is no summary. If the process died after the summary got written but before the effect landed, the summary reports an effect that didn't happen. The only summary you can trust is one you would still have if the process were killed at any point in its execution. The summary written by the process itself isn't one of those.

Third, and this is the failure mode most people underestimate. The summary is optimised. The model has been trained to produce a coherent narrative of what it did. Which means it will fill in gaps to make the narrative hang together. It isn't deliberately lying. It's doing the thing it was trained to do, which is produce fluent explanation. If you rely on the agent's own after-action report as your evidence chain, you're relying on a fluent explanation. That's exactly the wrong thing to rely on when you need to know what really happened.

The rule that comes out of these three. Log the effects from outside the agent. Log the intents from inside the agent. Never let the agent be the only witness to its own actions.

The minimum log

The compliance guides are proposing thirteen-field logs with policy pointers, tenant IDs, GDPR retention markers, and a five-layer taxonomy. If you work at a bank, follow the compliance guides. If you're a founder who's given an agent access to your git repository and your Cloudflare account, you need something considerably shorter. Considerably more honest. And considerably more likely to actually get written every time. Here it is.

1. Run ID. A UUID or a timestamp-slug generated at the start of the run and written to disk before the agent does anything else. Every subsequent log line, every commit message the agent writes, every file it creates references this ID. If the process dies at second three, this file is your evidence that the run happened. If the process dies at second thirty and produces an orphan file with the run ID in its filename, you know at a glance which run to look at when reconstructing.

2. Started-at, ended-at, exit reason. Stored outside the agent's own process, in a file the wrapping script owns. The exit reason is the honest one. "Clean exit". "Operator killed". "Connection dropped". "Hit iteration cap". "Raised exception". "Completed successfully" isn't an exit reason. It's a summary, and it belongs elsewhere.

3. Inputs. The prompt the agent was given, the configuration file it read, the environment variables that were set. And critically, the versions of any tools or models it used. If the agent picks up a slightly different model tomorrow because you upgraded the endpoint, the log for today needs to say which model ran today. Because next week when you compare outputs you won't remember.

4. Intent-per-step. For each decision the agent makes, record the choice, the alternatives that were considered, and the reason the chosen one won. Pick a topic. Choose a category. Decide to refresh vs write new. Decide to commit. This is where the "the agent's own summary" concern applies most. Keep this log line short and factual, not narrative. "Chose X over Y because Y failed the four-part test" is auditable. "Considered several options and settled on X after careful thought" is prose.

5. Effects, from outside. Every write the agent commits gets recorded by the wrapping script or the platform, not by the agent itself. A git commit sha. An API call and its HTTP status. A file created on disk with its byte size. An email queued with its message ID. The record survives the agent dying. And it can't be embellished by the same process that produced it.

6. Declines. The single most useful field, and the one nobody writes. Every time the agent explicitly decided not to do something, log it, with the reason. Rejected a topic as failing a safety rule. Refused to publish a draft under the word-count floor. Backed off a commit because the build gate failed. When you audit an agent's behaviour six weeks later, the declines tell you as much about whether the safety constraints are working as the actions do.

7. Retries and failures inside the run. Not just the final exit. Every attempt that failed and was retried. A run that hit the API three times before succeeding is a different animal from a run that succeeded first go. If you never log the difference, you won't spot the day an upstream provider quietly starts failing.

Seven fields. Compliance vendors will tell you to log fifty. If you can write seven honestly, every time, the seven you write will be worth more than fifty that get logged once and skipped whenever the run is under time pressure.

Where to store it so it survives the run dying

The one lesson from the morning that started this post. The log has to live outside the process it describes. That means:

The run ID gets written to disk in the first thirty seconds of the run, before the agent does any real work, in a file whose name contains the run ID.
Every log line is flushed to disk as it's written, not buffered. Buffered logs are lost logs when the process dies.
The exit is recorded by the wrapping script, not by the agent. If the agent is the one recording that it exited cleanly, the case where the agent lied about that is exactly the case you can't detect.
Wherever practical, log to two places. A local file and something the operator can inspect from another device. On this site, that means a Markdown run log in the repo and, for the failures that matter, an issue in the repo's own issue tracker. If the machine that ran the job isn't reachable, the issue on GitHub is.

The two-places rule is the small-shop equivalent of what large systems call "log shipping" or "immutable audit sinks". You don't need Kafka. You need any two places that aren't both going to be down at the same time as the process that generated the log.

What "reconstruct why the agent did X" actually looks like

The test of whether your log is working isn't whether it has the right fields. It's whether, six weeks after a run, you can pick a specific action and answer the question: why did the agent do this? Without asking the agent.

Concretely, for this site. Pick any published post. Now, using only the log, answer four things. Which topic ID did it come from? Why was that topic chosen over the others in the same category on that day? Which alternatives were rejected during research? What commit sha carried the file to main? If any of these four can't be answered from the log without opening the post itself, the log is missing something.

For a customer-facing agent, the test is: pick any message the agent sent to a customer. Which conversation context produced it? Which policy check ran on it before it was sent? Which alternative responses were generated and rejected? Which message ID did the outgoing channel return? Same test, different verbs.

If you can't pass the test, add fields until you can. If you can pass the test easily, resist the temptation to add more fields "just in case". The fields you never use are the fields your successor won't write when the log system gets ported to a new stack in eighteen months.

Retention: shorter than you think, longer than you want

The compliance frameworks give retention numbers, and they're worth citing because they're the floor. SOC 2 evidence typically covers twelve months. The EU AI Act guidance for high-risk systems points to at least six months of decision logs unless another law extends the window. If you're anywhere near either regime, meet the floor.

Under the floor, be honest about what you're trying to answer. For a solo operator running agents against their own systems, the useful retention is "long enough that any pattern of behaviour that would concern you has had time to become visible". In my experience that's roughly one to three months of full logs plus a summarised long tail. Longer than that and the logs stop being read. Shorter and you can't see the drift.

Whatever you settle on, write it into the procedure. An unstated retention rule becomes "as long as the disk lasts" by default. And disks are neither compliant nor useful.

The specific traps that catch small teams

Six things that have caught me or someone I know running agent workflows on their own infrastructure. Fix them once and never think about them again.

Trap one: logging into the same file the agent is writing. The log line and the agent's output land in the same file. The log line is what gets truncated when the agent crashes mid-write. Separate the log file from the artefact file. Always.

Trap two: logging via the same network call that's failing. A remote logging endpoint is fine until the day the network is the thing that broke. At which point the log of the network failure is the log line that couldn't be written because the network was broken. Log to disk first, ship to remote later.

Trap three: logging only in the happy path. The try block logs "action succeeded". The except block logs nothing because the developer assumed the exception itself was the log. The exception isn't the log. The exception is the reason for the log line that also needs to say what the agent was trying to do when the exception fired.

Trap four: logging the agent's chain-of-thought verbatim. Model chain-of-thought can be enormous, unstable across model versions, and often contains material that shouldn't be persisted. Customer data quoted back to the model. Tokens. Error messages containing secrets. Log the intent as a short structured field. The decision, the alternatives, the reason. Not the whole reasoning trace.

Trap five: logging without a run ID that links intent to effect. If your intent log lives in one file and your effect log lives in another and they share no identifier, you have two logs, not one audit trail. Every log line, everywhere, carries the run ID.

Trap six: assuming the model's summary of the run is the log. It isn't. The model's summary is a piece of evidence produced by the actor being audited. It's useful. It's not sufficient. Read it and then read the log the wrapping script wrote. If the two disagree, believe the log.

The five-minute test for any agent you run today

If you already run agents on your own systems, do this. Pick a run from any time in the last week. Pick one action inside that run. One commit. One email. One API call. Now, without opening the agent's output, answer four questions from your logs.

Which run did this action belong to?
What was the agent trying to do at that moment?
What did it consider doing instead, and why did it choose this?
Did the action succeed, and what was the effect on the outside world?

If you can answer all four in under two minutes, you have a log. If you can't, you have hope. And hope doesn't stand up in a review.

Fix the missing field this week. Not next quarter. The next agent run is either going to prove the log works or expose that it doesn't. It's far cheaper to be the one who discovers the gap than the one whose customer discovered it first.

Where this fits with the rest of the work

The reason to log at all is that the space of failures an agent can produce is larger than the space of failures a human can. A human who forgets to send an email leaves an inbox. An agent that forgets to send an email leaves a log line that isn't written. Which is a different, and easier to miss, kind of evidence. The specific case where the workflow reports success and does nothing is covered in The automation failure nobody catches: the workflow that runs green and does nothing.

If your problem is upstream of "what do I log", you haven't yet decided which tasks to hand to an agent at all. Start with How to automate your work with AI, which frames the choice.

If your interest is broader, where the whole direction of unattended AI systems is heading, and what that means for jobs and organisations built around human sign-off, the further-out picture is in AI is heading toward autonomy: what dark factories tell us about what's next.

Logging is one half of running unattended agents. The other is deciding what they're allowed to touch at all. The audit trail lets you reconstruct why the agent did what it did. The permissions boundary is what limits what it could have done in the first place. The two are designed together. That decision framework, blast-radius classification, standing vs just-in-time tokens, and the three artefacts every Class 2+ permission needs, is in deciding what an AI agent is allowed to touch.

FAQ

How long should I keep agent logs before deleting them?

If you're under SOC 2 or the EU AI Act, meet the floor those regimes set. Twelve months for SOC 2 evidence and at least six months for high-risk system decision logs. For a solo operator running agents on their own systems, one to three months of full logs plus a summarised long tail is usually enough to see drift without the logs going unread.

What if the agent crashes before it writes any log at all?

That's the case the run ID is designed to catch. Write the run ID to disk in the first thirty seconds, before the agent does any real work. Even a three-second crash leaves an evidence file behind. If you can't see a run ID file, the wrapping script didn't start properly. Which is a different bug from an agent failure and worth fixing separately.

Should I log the model's full chain-of-thought or just the decision it made?

Just the decision, the alternatives it considered, and the reason. As a short structured field. Chain-of-thought traces are enormous, unstable across model versions, and often contain material that shouldn't be persisted. Storing them creates more risk than value for a small-team audit trail.

How many log fields do I actually need to start with?

Seven. Run ID, timestamps and exit reason, inputs, intent-per-step, effects captured from outside, declines, and retries. Compliance vendors will push fifty. Seven you write honestly every run is worth more than fifty that get skipped whenever a job is under time pressure.

What if I already run agents and have no logs at all, where do I start?

Add the run ID and the wrapping-script exit reason first. Those two survive the process dying and unlock everything else. Then run the five-minute test on a recent action and add whichever field you needed to answer the four questions but couldn't.

Should I use a remote logging service or just write to a local file?

Write to a local file first and ship to remote later. A remote-only setup fails exactly when you need it most. The day the network is the thing that broke is the day the log line about the network breaking can't be written.

What if the agent's summary and the wrapping-script log disagree?

Believe the log. The agent's summary is written by the actor being audited from inside the process being described. Any conflict with an outside-observed log is resolved in favour of the outside observer. Same standard you'd apply to any other witness.

The one-sentence version

Log intent from inside the agent, log effect from outside the agent, cross-reference both by a run ID that survives the process dying, and treat the agent's own summary of its run as evidence to be checked against the log rather than as the log itself. Everything else is variation on those four rules. The day you skip any of the four is the day you find out which one you skipped.

What Is Retargeting? A Beginner's Guide That Actually Works

Prabhash Jha — Sat, 26 Sep 2026 13:45:05 +0000

Retargeting is showing ads to people who've already touched you in some way. Visited your site. Watched a video. Added something to a basket. Opened an email.

The logic behind it is straightforward, and it's sound. Most people don't buy on a first visit. Somewhere between 95% and 98% of visitors leave without converting. If you have no way to reach them again, that traffic is just gone. Retargeting is how you get a second conversation.

It's also the most over-credited tactic in digital marketing. And that's worth understanding before you spend anything on it.

How it actually works

Three steps. The middle one is where most implementations fall apart.

1 · Something records the visit. A pixel or tag on your site (Meta Pixel, Google tag) fires when someone loads a page. That person gets added to a list.

2 · That list becomes an audience. "Everyone who visited in the last 30 days." "Everyone who reached checkout but didn't buy." This is the part people get wrong. The section below is about why.

3 · You show ads to that audience across whatever inventory the platform can reach. Instagram. Facebook. YouTube. The Display Network. Other apps.

One important change: third-party cookies are being restricted across browsers. Safari and Firefox already block much of this. In practice, browser-based retargeting windows are shorter than they used to be, and your audiences are smaller. First-party approaches (your email list, uploaded customer lists, server-side tracking) have become materially more reliable than pixel-only setups. If someone's advice on this predates about 2023, treat it as out of date.

The audiences worth building

The default is "everyone who visited the site". That's the worst one. It treats a person who bounced in three seconds identically to someone who spent ten minutes reading your pricing page.

Segment by how much interest they showed. And spend accordingly.

Audience	Intent	What to show
Bounced in under 10 seconds	Near zero	Nothing. Exclude them.
Read one article	Low	More content, not an offer
Visited a product or pricing page	Medium	The product, with a specific reason to return
Started checkout, didn't finish	Very high	A reminder, and address the objection
Bought already	n/a	Exclude, or show something different

Two of those rows are exclusions. They matter as much as the inclusions.

Excluding bouncers stops you paying to chase people who never engaged. Honestly, in most accounts they're the largest segment by a mile. Advertising to them is where retargeting budgets quietly die.

Excluding buyers stops the single most irritating experience in online advertising. Being followed for three weeks by an ad for something you already own. It also wastes money on people who cannot convert again. Every retargeting campaign needs a purchaser exclusion. A surprising number don't have one.

The honest problem: retargeting steals credit

This is the part most guides skip. It changes how you should judge the numbers.

Retargeting targets people who already showed intent. A meaningful share of them were going to come back anyway. They bookmarked you. They were comparing options. They were waiting for payday. When they return and buy, the retargeting ad claims the conversion.

So a retargeting campaign showing a spectacular 12× ROAS is often not generating 12×. It's taking credit for sales that would have happened without it. The dashboard cannot tell the difference. All it sees is that an ad was shown, and a sale followed.

This doesn't make retargeting useless. It makes the reported number unreliable. Three practical consequences flow from that.

Judge it on incremental lift, not reported ROAS. The real question isn't "how much revenue did this campaign report". It's "how much more did I sell than I would have without it".

Test it by holding out. Exclude a random slice of your retargeting audience, say 20%, for a month. Compare their conversion rate against the group that saw ads. The difference is the actual contribution. Most platforms support this. Almost nobody does it. It's the single most valuable test in this article.

Be sceptical of very high ROAS. A retargeting campaign reporting far better returns than everything else usually indicates credit-stealing, not brilliance. Genuinely incremental retargeting produces good numbers. Not absurd ones.

Frequency: the thing that turns it negative

Retargeting is the only channel that can actively damage your brand. That's because the same small audience sees your ads over and over.

Without a cap, someone who visited once gets served the same creative for weeks. Past a point, additional impressions don't just stop working. They generate irritation. Which is a real cost you never see measured.

Sensible defaults:

Cap frequency at roughly 3–5 impressions per person per week.
Cap duration. A 30-day window suits most businesses. Someone who hasn't returned in 30 days isn't going to because of a sixth ad.
Rotate creative every two to three weeks. Performance decay in retargeting is mostly fatigue. Not audience exhaustion.
Exclude converters immediately, not at the end of the campaign.

A useful test. Would you find this frequency acceptable if you were the one being followed? People are unreasonably tolerant of their own campaigns. And unreasonably annoyed by everyone else's.

What to actually say

The most common mistake is showing the same message they already ignored.

They saw your offer. They didn't act. Repeating it louder rarely works. Better approaches, by audience.

For content readers: more of what they came for. The goal is a second visit. Not a sale. Trying to close someone who read one blog post is why "retargeting doesn't work for us" gets said.

For product viewers: show the specific thing they looked at. Add the information that might have been missing. Delivery time. Returns policy. A review. The size guide. Most abandonment is an unanswered question, not a pricing objection.

For abandoned checkouts: the highest-value audience by a distance. Be direct. Remind them what's waiting. Address the usual blockers. Shipping cost. Payment options. A guarantee.

On discounts: the reflex is to offer one. Be careful here. If you routinely discount to abandoners, people learn to abandon. Try answering the objection first. Hold the discount for a later touch if at all.

Where retargeting isn't the answer

When your traffic is too small. Retargeting needs volume. With a few hundred visitors a month, you'll have an audience of dozens. The platform won't deliver efficiently. The money is better spent getting more people in.

When the product is a one-off, low-value purchase. The economics rarely justify the setup.

When the real problem is the landing page. If people arrive and leave immediately, retargeting pays twice to bring back people the page will fail again. Fix the page first. Retargeting multiplies whatever your site already does.

When you have an email list. Email costs almost nothing per contact and reaches people directly. If someone gave you an address, use that before paying to advertise to them.

The privacy side, and why it's not optional

Retargeting is the tactic people find most unsettling. The regulation around it has tightened accordingly. This section is short, but skipping it creates real exposure.

Consent before tracking. Under GDPR and similar regimes, dropping an advertising pixel requires consent obtained before it fires. Not a banner that loads after the tag has already run. A cookie notice that says "by continuing you agree" while tracking from the first millisecond does not meet the standard. That pattern is extremely common.

Say what you're doing. Your privacy policy should state plainly that you use advertising cookies. Name the platforms. Explain how someone opts out. Vagueness here is what turns a technical breach into a complaint.

Don't retarget on sensitive inference. Health conditions. Financial distress. Sexuality. Religion. Both major platforms prohibit targeting built on these, and the reputational damage from getting it wrong far exceeds any campaign gain. The classic failure is retargeting people who viewed a medical or debt-related page. Technically easy. A serious problem.

Honour deletion requests. If someone asks to be removed, they need removing from your uploaded audiences too. Not just your CRM. Custom audiences persist independently and are routinely forgotten.

The practical framing: retargeting is the most visible thing you do. People notice being followed in a way they never notice a search ad. Doing it carelessly costs you more in trust than it earns in conversions. Which is the same argument as the frequency cap, arriving from a different direction.

Why your audiences are smaller than they used to be

If you've run retargeting before and your audience sizes look wrong, they probably are. The ground moved. And it moved on purpose.

Retargeting was originally built on third-party cookies. Browsers have spent years dismantling them. Safari's WebKit announced that "cookies for cross-site resources are now blocked by default across the board" in March 2020, alongside "deleting all of a website's script-writable storage after seven days of Safari use without user interaction on the site" (WebKit: Full Third-Party Cookie Blocking and More). Other browsers have moved in the same direction at their own pace.

Three practical consequences. None of which are a reason to stop.

Your real audience is smaller than the pool you imagine. A meaningful share of visitors will never enter it. Plan spend against the audience the platform actually reports, not against your traffic.
Long windows are largely fiction now. A 180-day audience built on browser storage is not a 180-day audience. Short windows were always the better performers. Now they're also the more honest ones.
First-party signal is the fix. Server-side conversion APIs, logged-in identifiers and email lists survive what the browser deletes. Same argument as the email point above, arriving from the technical side.

A sensible setup

If you're starting from nothing.

Install the pixel or tag properly and verify it fires. On every page. And specifically on checkout and purchase events. Broken tracking is the most common failure, and it fails silently.
Build three audiences: product/pricing viewers (30 days), checkout abandoners (7 days), and purchasers (180 days, used as an exclusion).
Run one campaign to abandoners first. Highest intent. Clearest message. Fastest read on whether any of this works for you.
Cap frequency and duration before launch, not after complaints.
After a month, run a holdout test. Learn what it's genuinely contributing before you scale it.

Step 5 is the one that separates people who use retargeting from people who are used by it.

FAQs

What is retargeting in simple terms?

Showing ads to people who've already interacted with you. Visited your site. Watched a video. Abandoned a basket. Since most people don't buy on a first visit, it's a way to reach them again rather than losing that traffic entirely.

What's the difference between retargeting and remarketing?

In practice, nothing. They're used interchangeably. Historically "remarketing" leaned toward email follow-up and "retargeting" toward paid ads. But Google uses "remarketing" for its ad product and the distinction has effectively disappeared.

Does retargeting still work with cookie restrictions?

Yes, but less well than it used to. Safari and Firefox already block much third-party tracking and windows are shorter. First-party approaches (your email list, uploaded customer lists, server-side tracking) are now materially more reliable than pixel-only setups.

How long should a retargeting window be?

Around 30 days for most businesses. Shorter for impulse purchases. Longer for considered ones. Someone who hasn't returned in 30 days is unlikely to return because of another ad, and continuing to serve them mostly generates irritation.

Why does my retargeting show such a high ROAS?

Usually because it's claiming credit for sales that would have happened anyway. Retargeting targets people who already showed intent, and many were returning regardless. Test it with a holdout. Exclude 20% of the audience for a month and compare. That'll show you the actual incremental contribution.

Should I offer a discount to people who abandoned their basket?

Cautiously. Routine discounting to abandoners teaches people to abandon. Most abandonment is an unanswered question. Shipping cost. Delivery time. Returns. Answer the objection first and hold any discount for a later touch.

Key takeaways

Retargeting reaches people who already showed interest. Which is why it looks so effective, and why the numbers mislead.
Segment by depth of interest. Exclude bouncers and buyers. The exclusions matter as much as the audiences.
Reported ROAS overstates the contribution. Run a holdout test to find the real lift.
Cap frequency (3–5 a week) and duration (about 30 days). Rotate creative every few weeks.
Don't repeat the message they already ignored. Answer the objection instead.
If the landing page is the problem, retargeting just pays twice to fail twice.

What Is Programmatic Advertising? A Plain-English Guide

Prabhash Jha — Sat, 26 Sep 2026 13:45:04 +0000

Programmatic advertising is the automated buying and selling of ad space. Instead of a human negotiating with a publisher over a rate card, software decides in the time it takes a page to load which ad to show, to whom, and what that impression is worth.

That's the whole idea. Everything else is plumbing.

Plumbing is worth understanding anyway, because programmatic is the only major ad channel where a meaningful share of your budget can disappear without anybody doing anything wrong on purpose. In search or social you're buying from one company that owns the inventory. In programmatic you're buying through a chain of intermediaries, from publishers you've never heard of. Difference between a well-run buy and a badly-run one is usually not the creative or the targeting. It's whether you know where the money went.

What actually happens when a page loads

An auction runs, and it finishes before the page does. Here's the sequence:

The publisher's ad server notices it has an ad slot to fill and sends a request out.
That request reaches an ad exchange (a marketplace) carrying details about the slot. Site. Ad size. Roughly where the reader is. What device they're on.
Advertisers' systems evaluate the request and decide whether this particular impression is worth bidding on. And how much.
Highest bid wins, the ad is returned, and it renders.

All of it inside roughly 100 milliseconds. This is real-time bidding, or RTB. It's the mechanism most people mean when they say "programmatic".

Two things follow from that speed. Both matter operationally.

Nobody looks at the impression before it's bought. There's no human in the loop deciding this page is appropriate and that one isn't. Whatever rules you set before the campaign started are the only judgement being applied. If your exclusion list is empty, you haven't made a decision. You've declined to make one, and the auction will fill the gap on your behalf.

Every decision is made on the data in the bid request. Request carries what it carries. If a publisher passes a vague or spoofed domain, your platform is bidding on a description rather than a page. That's why the transparency plumbing further down this article exists at all.

The four acronyms worth knowing

Most programmatic explainers drown you in initialisms. You only need four.

DSP, Demand-Side Platform. The buyer's tool. Where an advertiser sets budgets, targeting and bid rules. When you "run programmatic", you're almost always sitting in a DSP.

SSP, Supply-Side Platform. The publisher's side. Puts their inventory up for sale and works to get the best price for each impression.

Ad exchange. The marketplace where the two meet and the auction runs.

DMP / CDP, the data layer. Where audience information lives, so a bid can be based on who is likely seeing the ad rather than only where it appears.

A useful way to hold it: the DSP is the buyer, the SSP is the seller, the exchange is the market floor, and the data layer is what tells the buyer whether to raise their hand.

The thing worth noticing is that every one of those is a business with its own margin. Not a scandal. Market makers charge for making a market. But it's the reason a rupee of budget and a rupee of publisher revenue are different numbers. And why "what did we pay" and "what did the publisher receive" are two separate questions you should be able to answer.

Ask which auction you are actually in

Before you tune a single bid, find out whether the exchange is running a first-price or a second-price auction. Correct bidding behaviour is opposite in each.

In a second-price auction the winner pays one increment above the second-highest bid, so bidding your true maximum is safe. You rarely pay it. In a first-price auction the winner pays exactly what they bid, so bidding your true maximum means paying your true maximum every single time. Industry moved largely to first-price. But "largely" isn't "entirely", and different supply paths inside the same DSP can behave differently.

This is a question your platform rep can answer in one sentence. Ask it. Second-price bidding habits applied to a first-price auction never surface as an error anywhere. They surface as a CPM that's quietly higher than it needed to be, indefinitely. Which is honestly the most expensive kind of mistake because nothing ever flags it.

Where the money actually goes

This is the part that gets skipped. And it's the part that costs the most. There are four leaks. All four are knowable, and none of them are reported to you by default.

1. Not every impression is a person

Invalid traffic is a real line item, and the industry has a formal vocabulary for it that's worth borrowing, because it tells you what you can and can't fix yourself.

Media Rating Council splits it in two. General Invalid Traffic (GIVT) is the routine, identifiable kind. Crawlers and spiders. Straightforward bot traffic. Data centre traffic. Detectable through ordinary filtration. Sophisticated Invalid Traffic (SIVT) is the kind built to evade detection: hijacked browsers and devices, adware, incentivised or deceptive clicking, malware-driven traffic using obfuscation (MRC Invalid Traffic Detection and Filtration Guidelines Addendum).

Practical consequence: GIVT filtering is table stakes and you should assume your platform does it. SIVT is where the money actually goes. It requires a vendor or an exchange that invests in detection. And it's why verified inventory costs more per impression and usually costs less per outcome. When you choose between a cheap CPM on an unverified path and a higher CPM on a verified one, you aren't choosing between expensive and cheap. You're choosing between a known price and an unknown one.

2. Viewability is not delivery

An impression can be served and never scroll into view. "Served" and "seen" are different numbers, and the gap between them is money.

There's an actual standard here, which is useful because it gives you something to hold a vendor to. A display ad counts as viewable when 50% of its pixels are in the browser window for a continuous one second. For creatives larger than 242,000 pixels the threshold drops to 30% of pixels. In-stream video needs 50% of pixels for a continuous two seconds (Google Ad Manager: Overview of Viewability and Active View).

Read that definition carefully. It's a floor rather than a promise. Half an ad, visible for one second, on a page nobody was reading, counts. Viewability is a hygiene metric. It tells you the ad had a chance. Doesn't tell you it took it. Buy against it, because unviewable inventory is wasted by definition, but don't report it upward as though it were a performance number. It's the absence of one specific failure, not the presence of a result.

3. The supply chain takes a cut at every step

Between the advertiser's budget and the publisher's revenue sit the DSP, the exchange, the SSP and often several more parties. You can find out who they are. Most advertisers never ask.

Three specifications exist precisely so you can. ads.txt lets a publisher declare who is authorised to sell their inventory. sellers.json "provides a mechanism to enable buyers to discover who the entities are that are either direct sellers of or intermediaries in the selling of digital advertising", and the OpenRTB SupplyChain object "enables buyers to see all parties who are selling or reselling a given bid request" (IAB Tech Lab: sellers.json). Together they let you trace a single impression back through every hand it passed through.

Operational version of this is supply path optimisation. Noticing that the same publisher's inventory is reachable through four routes at four different costs, and buying it through the shortest one. It's unglamorous work that nobody will thank you for. And it's one of the highest-return things available to a mid-sized programmatic buyer, because it improves your economics without touching your creative, your targeting or your offer.

Question to put to your platform or your agency, in writing: what share of our spend reaches the publisher, and through which paths? You're entitled to an answer. Quality of the answer tells you a great deal about who you're working with.

4. Frequency compounds quietly

Without a cap, the same person sees the same creative repeatedly. You pay each time. Past a certain point each additional view does less than nothing. It irritates.

Reason this leak is so persistent is that frequency is invisible in every report you're likely to be looking at. Impressions go up, cost goes up, the ratio between them looks stable, and nothing on the dashboard says this is a small group of people seeing it dozens of times rather than a large group seeing it once. You have to go and ask for the frequency distribution specifically. And the first time you do that on an uncapped campaign is usually memorable.

Cap it before launch. Uncapping later is a decision. Never capping is an oversight.

The order I actually work in

If I were opening a programmatic account tomorrow, this is the sequence. Order isn't decorative. Each step is close to worthless without the one before it.

#	Step	Why it sits at this position
1	Fix what a conversion is worth	The system optimises toward whatever number you give it. Everything downstream inherits this decision.
2	Set the frequency cap	The cheapest single protection available, and impossible to apply retroactively.
3	Build the exclusion list	An empty list isn't neutrality. It's an unmade decision the auction makes for you.
4	Choose verified supply and name the paths	Decides what you're actually buying, before any budget optimises against it.
5	Run flat, learn, then optimise	Optimising during the learning period means optimising against noise.
6	Reconcile spend to outcomes weekly	The only step that catches problems the platform isn't designed to report.

Step 1 is the one people skip. And it's the one that determines whether the other five matter. Automated buying will optimise toward whatever you tell it to, at speed and at scale. Point it at the wrong number and it will hit that wrong number very efficiently. If you can't yet say what a conversion is worth to you, the sheet that tells you whether your marketing makes money is the prerequisite, not the follow-up.

Step 5 is the one people violate without noticing. A campaign that has been live for four days has produced a small sample, and a small sample produces dramatic swings that are mostly randomness. Reacting to each swing feels like management and is actually noise amplification. Decide in advance how long you'll leave it alone, write that date down, and honour it.

The weekly check that catches most of it

Once a week, pull the placement report. The list of actual domains and apps your ads ran on. Read it. Not skim it. Read it.

You're looking for three things.

Sites you can't identify. Open the domain. If it's a content farm with forty ad slots and no discernible reader, it's a made-for-advertising site. Exists to convert your budget into its revenue. Exclude it and move on.

Sites that surprise you. Sometimes pleasantly. A niche publication you'd never have thought to buy, performing well. That's the genuine upside of programmatic and it's worth harvesting deliberately. Take the good surprises and buy them directly next quarter.

Concentration. If a small number of domains are absorbing most of your spend, your "programmatic reach" is really a handful of sites bought expensively through an auction. That may be fine. But you should know it. And you can often buy those sites for less by going straight to them.

That report is where the four leaks above stop being abstractions and turn into a list of domains you can act on. Fifteen minutes a week. Honestly the highest-yield fifteen minutes in the channel.

The three things that have to be present

Programmatic only works when marketing, technology and data are all doing their job.

Marketing decides who's worth reaching and what to say to them.

Technology executes the buy at a speed and scale no human can match.

Data tells you whether any of it worked, and feeds that back into the next bid.

Weaken any one and the other two stop compounding. Good data with no offer is a well-measured failure. A great offer with no measurement is a guess you keep repeating. And strong marketing with weak technology is a good idea being bought at a bad price. Which is the failure mode programmatic specialises in, because everything still runs, the reports still populate, and nothing announces that it's going wrong.

When programmatic is the wrong tool

Programmatic rewards scale and patience. It's a poor fit when:

Your addressable audience is small enough to reach by hand. If your entire market is a few hundred companies, a list and an email client beat an auction. Auction's advantage is finding people at a scale you can't enumerate. If you can enumerate them, you're paying for machinery you don't need.

You need results this week. Learning period is real, and cutting it short means paying for data you then throw away. Programmatic is a channel you commit a quarter to, not a fortnight.

You can't yet measure what a conversion is worth to you. Covered above, and the most common of the four.

Demand for your category doesn't exist yet. Programmatic is very good at putting a message in front of people. It isn't good at making people want something they've never considered. That's a brand problem, and the point at which it becomes the binding constraint is when performance marketing stops working and brand is the only lever left.

If you're early, comparing channels and trying to decide where to start, programmatic is almost never the answer. Google Ads and Meta Ads are a better first classroom. One counterparty, a much shorter supply chain, and mistakes that surface faster and cost less.

The honest summary

Programmatic isn't a channel that works or doesn't. It's an execution layer. It makes whatever you already do faster and larger, including the mistakes. Get the offer, the measurement and the audience right first. Then automation is leverage rather than an expensive way to be wrong at scale.

The four leaks aren't a reason to avoid the channel. They're the reason to treat it as an operational discipline rather than a setting you switch on. Invalid traffic, viewability, supply path and frequency are all knowable. And the buyers who bother to know them are buying the same inventory as everyone else at a materially better price. That gap is the entire opportunity.

If you want the practical version of how this fits into a wider paid strategy, the performance marketing playbook covers how the channels sit together, marketing metrics explained covers the numbers you'll be judged on, and what advertising actually is and why digital took over explains how an auction ended up deciding all of this in the first place.

FAQs

What is programmatic advertising in simple terms?

It's the automated buying and selling of ad space. Instead of a person negotiating with a publisher over a rate card, software decides in the time it takes a page to load which ad to show, to whom, and what that impression is worth. Everything else, including the platforms and the acronyms, is plumbing around that one idea.

What is the difference between a DSP and an SSP?

A DSP, or demand-side platform, is the buyer's tool: where an advertiser sets budgets, targeting and bid rules. An SSP, or supply-side platform, is the publisher's side, putting their inventory up for sale and working to get the best price for each impression. The ad exchange is the marketplace where the two meet and the auction runs.

Is programmatic the same thing as real-time bidding?

Not quite, although the terms get used interchangeably. Real-time bidding is the auction mechanism: a request goes out, advertisers' systems evaluate whether the impression is worth bidding on, the highest bid wins and the ad renders. All inside roughly 100 milliseconds. RTB is what most people mean by programmatic, but programmatic is the broader term for automated buying, and it also covers direct deals executed through the same pipes without an open auction.

Where does programmatic budget actually leak?

In four places. Not every impression is a person. Invalid traffic, both the routine kind and the kind built to evade detection, is a real line item. An impression can be served and never scroll into view, so "served" and "seen" are different numbers. The supply chain between your budget and the publisher's revenue takes a cut at each step, and you can trace those steps using ads.txt, sellers.json and the SupplyChain object. And without a frequency cap you pay repeatedly to irritate the same person.

What counts as a viewable impression?

For display, 50% of the ad's pixels visible in the browser window for a continuous one second. For creatives larger than 242,000 pixels, 30% of pixels. For in-stream video, 50% of pixels for a continuous two seconds. Treat that as a floor rather than a promise. It means the ad had a chance to be seen, not that anybody saw it.

Should I bid differently in a first-price auction?

Yes, and it's the first question to settle. In a second-price auction the winner pays just above the second-highest bid, so bidding your true maximum is safe. In a first-price auction you pay exactly what you bid, every time, so your maximum becomes your average. Ask your platform which one applies on which paths before you tune anything else.

When is programmatic the wrong tool?

When your addressable audience is small enough to reach by hand. When you need results this week (the learning period is real, and cutting it short means paying for data you throw away). When you can't yet measure what a conversion is worth. Or when demand for your category doesn't exist yet. Automated buying optimises toward whatever number you give it, so a wrong target gets hit very efficiently.

When a Competitor Undercuts You by Half, What Is the Brand Actually Worth?

Prabhash Jha — Fri, 25 Sep 2026 14:35:31 +0000

Search for what to do when a competitor undercuts you and you will get the same four sentences from every result. Do not start a price war. Compete on value, not on price. Emphasise quality and service. Differentiate. All of it is true, none of it is wrong, and not one line of it tells you what to do on Monday when a real prospect forwards you a quote at half your number and asks whether you would like to revise.

Two of the results on that first page are price-monitoring software companies, which explains the shape of the advice. The rest are written at the altitude of strategy, where price wars are an abstraction. At the altitude where you actually live, this is not a strategic question at all. It is a sequence of specific checks, run in a specific order, and the order matters because doing the last one first is how people talk themselves into a discount they cannot reverse.

Brand-building writing asserts that a strong brand lets you charge more and then stops, as though the assertion were the useful part. The useful part is the arithmetic underneath it: which buyers will pay the premium, what they are actually paying for, and how you find out before you have given anything away.

First: do nothing to your price for two weeks

Change nothing about your pricing until you have established what the cheaper competitor is actually selling, because most "half your price" quotes are not the same scope.

This is the whole first step and it is skipped almost universally, because the arrival of a cheap competitor feels like an emergency and doing nothing feels like negligence. It is not. Price changes are close to irreversible — a client who has been at the lower number for one cycle now treats it as the price, and moving it back is a fresh negotiation you will probably lose. You get one shot, and spending it in week one on incomplete information is how a temporary competitive event becomes a permanent margin problem.

So spend the two weeks establishing four things:

What is actually in their scope. Get the competing quote if a prospect will share it, and read it line by line against yours. In services the gap between two prices is usually a gap in what is included: fewer rounds, no strategy time, a junior team, reporting monthly instead of weekly, no account manager, a twelve-month lock-in, or payment terms that are much better for them than yours. A quote at half your price with two-thirds of your scope removed is not undercutting. It is a different product, and the correct response is a product decision rather than a price one.

Whether the price is even sustainable. New entrants routinely price below their own cost to buy logos, and that is a temporary condition with a natural end. You cannot know their numbers, and guessing at them is a good way to convince yourself of something comfortable — but you can observe behaviour. A price that arrives with a heavy launch push, applies to any prospect, and comes with unusually short commitments is behaving like customer acquisition. A price that is quietly consistent, applies selectively, and comes with a structurally leaner delivery model is behaving like a real cost advantage. The second one is the one that changes your business.

Who exactly is asking. Which of your prospects and clients have raised it? Write the names down. This list is the actual dataset and it is much smaller than the noise suggests — usually two or three accounts and a handful of prospects, which is a very different problem from "the market has repriced."

Whether you are losing on price or losing and being told it was price. These are extremely hard to tell apart and almost everybody gets it wrong. Price is the socially easy reason to give. Nobody enjoys telling a supplier that the reporting was confusing or that their account manager was slow to reply, and "we found someone cheaper" ends the conversation politely. This is exactly why the first sales conversation is one of the five things a founder should never fully delegate — the real reason surfaces in the wording of a live conversation and never survives into a summary.

The question that decides everything: which buyers are you losing?

Before you consider any response, sort the accounts you are losing into ones you wanted and ones you did not, because a cheap competitor taking your worst accounts is a favour and should not be resisted.

This is the single most useful move in the whole sequence and it is almost never the first instinct, because losing feels bad regardless of what is being lost. But the accounts most likely to leave for a materially cheaper option have a recognisable profile, and it is largely the same profile as your least profitable work: heavy on requests, slow to approve, price-focused from the first conversation, no interest in strategy, high service load relative to fee. If you have never done the exercise, which client is actually profitable when everyone works on everything is the method, and you want that answer before you decide how hard to fight.

Run every at-risk account through three questions:

Question	If yes	If no
Is this account above your average margin?	Worth defending	Losing it improves the business
Does it generate referrals, reference value or category credibility?	Worth defending beyond its margin	Its value is only the fee
Would you take this client again today at the same fee?	Defend it	You have your answer

An account that fails all three and leaves for a competitor at half price has done something you should have done yourself and did not have the nerve to. The related judgement — that revenue is never on its own a reason to keep an account — is the subject of when to fire a client.

The genuinely bad outcome is different and much narrower: it is losing accounts that pass all three. If your high-margin, well-run, strategically engaged clients are the ones asking about the cheaper option, you have a real problem and it is probably not a pricing problem. It means those clients cannot see what they are paying the difference for, which is a value-communication failure, and discounting will not fix it — it will confirm it.

Which segments are actually price-elastic

Assume price sensitivity is a property of the buyer and the purchase, not of your category, because the same service sold to two different buyers has completely different elasticity.

The mistake is treating "our market has become price sensitive" as a fact about the market. Within any category the elasticity varies enormously by buyer, and the variation is predictable. Buyers become much less price-sensitive when:

The cost of the thing going wrong is large relative to the fee. Nobody shops hard on price for the supplier who touches their payment flow or their compliance filings. The fee is small next to the failure.
They cannot easily evaluate quality in advance. Where the buyer cannot tell good from bad before purchase, price becomes a quality signal rather than a cost, and cutting it actively hurts you. This is why the discount sometimes loses you the deal. It is also one of the better-evidenced findings in the marketing literature rather than a piece of agency folklore: Rao and Monroe's integrative review, The Effect of Price, Brand Name, and Store Name on Buyers' Perceptions of Product Quality, pooled the experimental work to that point and found a positive and statistically significant relationship between price and perceived quality — strongest, unsurprisingly, where the buyer had little else to go on.
Switching has a real cost in time or risk. Integration, onboarding, retraining, institutional knowledge.
The buyer is spending someone else's money and will be asked to justify the choice. A procurement-defensible choice is worth a premium to the individual making it.
The purchase is infrequent and consequential. Rare, high-stakes purchases get evaluated on confidence; frequent, low-stakes ones get evaluated on price.

And they become highly price-sensitive when the opposite holds — a routine, easily-compared, low-consequence, easily-switched purchase made by someone spending their own money. If your offer sits there, brand is not going to save you, and the honest answer is that you need to either change what you sell or accept the market clearing price.

Most businesses serve a mix and have never separated them. Do that first. You will typically find the cheap competitor is only actually competitive in one segment, and the panic was about the whole business.

What a discounter structurally cannot deliver

Charge for the things a lower-priced operator cannot provide because of how they are built, not for the things they merely claim less loudly.

This is the part brand writing gets wrong. "We care more" and "we have better people" and "we're more experienced" are not differentiators; they are claims every competitor makes, including the cheap one, and the buyer has no way to verify any of them before purchase. They cost nothing to copy, so they are worth nothing.

What is worth something is a constraint the competitor's own model imposes on them. A business priced at half of yours has made real structural choices to get there, and those choices have necessary consequences:

Time with anyone senior. A price at half requires more work per person, which mathematically means less attention per account and more junior delivery. They cannot give you both the price and the seniority.
Saying no. A business buying logos cannot afford to turn down bad-fit work or tell a client their request is a mistake. Willingness to disagree with a client is a direct function of not needing them.
Continuity. Low-price models run higher staff turnover, so the person who knows the account is more likely to be replaced. Continuity is a genuine asset and clients feel its absence in month eight, not month one.
Absorbing an emergency. Slack capacity costs money. A lean operation cannot drop everything when something breaks, because there is nothing to drop.
Being right about something unpopular. Related to saying no, but sharper: advising against the spend, the campaign, or the rebrand the client already wants.

Two rules for using these. Be specific rather than general — "you will get a senior person on every call, and here is who" beats "we offer senior attention." And be honest that they are trade-offs rather than universal superiority, because the buyer who genuinely just needs the cheap version should go and buy it. Trying to keep them is how you end up with the unprofitable account and the discount.

The delivery side of this matters as much as the claim. A premium price with an inconsistent experience is the weakest position available — worse than being cheap, because it invites the comparison and loses it. Your brand is whatever your worst touchpoint is is the operational constraint on everything in this section.

The response sequence, in order

Work through the responses in ascending order of irreversibility, and stop at the first one that holds.

Change nothing, and say nothing. For most of the at-risk list, the answer is that they are not going anywhere and raising it makes them think about it. Do not pre-emptively defend a price nobody has challenged. Volunteering a discount to a client who was content is a self-inflicted wound and it happens constantly in nervous quarters.
Improve what the price buys, without moving the price. Add something with high perceived value and low marginal cost to you — usually access, speed, or visibility. A quarterly session with someone senior. A faster response commitment. Reporting that answers the question they actually ask. This changes the ratio without touching the number, and it is reversible.
Restate the scope explicitly. Often the client has genuinely lost track of what they are getting, especially on a retainer that has been running a while and quietly expanded. Put the full inventory in front of them next to the competitor's scope. This is not a sales trick; the drift is real, and it is the same drift that makes a profitable retainer quietly become an unprofitable one.
Introduce a genuine lower tier, deliberately built. A stripped version at a lower price, with real things removed, sold as a different product. This is legitimate and often correct — it lets you compete for the price-sensitive segment without repricing the rest. The discipline is that things must actually be removed. A "lower tier" that quietly delivers the same service is just a discount with extra steps, and every full-price client will find out.
Cut the price. Last, rarely, and never quietly. See below.

Anything at steps 1 to 3 you can undo. Step 4 is a product commitment. Step 5 is effectively permanent. Most businesses under pressure start at step 5 because it is the only one that feels like action.

When matching the price is actually correct

Match the price when the competitor has a real structural cost advantage in a segment you need to keep, and be honest that you are changing your business rather than running a promotion.

There are three conditions, and all three have to hold:

Their advantage is structural, not promotional. They have a genuinely leaner model — offshore delivery, heavy automation, a narrower service that they do repeatedly, a different cost base. If it is structural, it does not expire, and waiting it out is not a strategy.

The segment is one you cannot afford to lose. It carries your fixed costs, or it is where your category credibility lives, or losing it makes you sub-scale.

You can find the cost to fund it. Matching a price with your existing cost structure just relocates the problem to your margin, and a business running the same operation at half the price is a business quietly going out of business. If you match, something in delivery has to change in the same decision — scope, seniority mix, automation, the number of accounts per person.

If all three hold, match, and restructure in the same move. If only the first two hold, you do not have a pricing decision, you have a cost problem you have to solve first. And if the first does not hold — the price is promotional — do not match at all, because you will still be at the low price long after they have raised theirs.

The one thing never to do is match invisibly, account by account, as each one asks. That is how you end up with a rate card that is fiction, a realised price nobody can explain, and clients comparing notes. If you are going to reprice, reprice deliberately and structurally; the mechanics of moving a price across a book of clients without losing the good ones are in how to raise your prices without losing the accounts you actually want to keep, and they run the same in both directions.

Three mistakes that make it worse

Discounting to keep an account, then resenting it. The discount is granted, the margin goes, and the account quietly drops down the priority list because it is now the least rewarding work in the building. The client feels it within two months. You have paid for the privilege of delivering worse service and will lose them anyway, later, with less money.

The reason a discount hurts more than it feels like it should is that price is the most leveraged line in the business. Marn and Rosiello's Managing Price, Gaining Profit made the point that for a typical company a given percentage improvement in price moves operating profit further than the same percentage improvement in volume, variable cost or fixed cost — because a price change carries no offsetting cost at all. That leverage runs in both directions, and it is why a concession that sounds modest in a client conversation is rarely modest by the time it reaches the bottom line. Their other contribution is the more useful one here: the pocket price waterfall, the gap between the price on your rate card and the money you actually keep once every discount, concession, extra round and unbilled hour has been taken out of it. Most services businesses have never drawn theirs, and the number that comes out of it is usually well below what everyone in the room believes they are charging.

Attacking the competitor to the client. Explaining why the cheap option is bad positions you as the incumbent defending territory, which is a weak posture and reads as anxiety. Describe what you do and what it costs. Let the comparison happen in the buyer's head. If your differentiation only works when you are in the room disparaging someone, it is not differentiation.

Concluding you have a brand problem and rebranding. A cheaper competitor prompts a surprising amount of logo work. It is almost never the answer — the diagnostic for whether you are actually looking at a brand problem or something else entirely is the rebrand that should not have happened. Repositioning may well be correct. A new visual identity, on this timeline, for this reason, is displacement activity.

There is a fourth, subtler one: assuming the answer is more performance spend. Buying your way past a price disadvantage through paid channels works until the channel gets expensive, and then you have two problems. That failure mode is when performance marketing stops working, and brand is the only lever left.

What to look at ninety days later

Judge the response on retained margin and the profile of who stayed, not on how many accounts you kept.

The count is the wrong measure and the tempting one. Keeping every account by conceding on each is a worse outcome than losing three and holding your price, and only the margin line shows the difference. Look at:

Realised average rate, not rate card. If it moved and you did not decide to move it, the discounting happened account by account and you have the invisible-repricing problem.
Who left, against the three-question table. If the departures were accounts failing all three, the competitor did you a favour and the correct conclusion is that your qualification was too loose — that is a pricing floor question, not a competitive one.
Whether any account that passed all three left. Even one is the signal worth acting on, and the action is understanding what they could not see, not lowering the price.
New business at full price. The cleanest evidence. If you are still winning new work at the old number, the price is fine and the noise was a small number of loud conversations.

The honest summary is that a competitor at half your price is mostly a sorting event rather than a pricing event. It sorts your client base into people buying an outcome and people buying an hour, and it does it faster and more decisively than you would have managed on your own. The businesses that come out of it in better shape are the ones that let the sort happen and adjusted who they sell to. The ones that come out worse are the ones that tried to keep everybody, and paid for it out of margin for the next two years.

What Is Marketing Automation? A Beginner's Guide

Prabhash Jha — Fri, 25 Sep 2026 14:35:30 +0000

Marketing automation is software that runs marketing actions triggered by what someone does, so you're not doing them by hand each time.

Someone downloads a guide. A sequence of emails begins. Someone abandons a cart, and four hours later a reminder goes out. A lead crosses a scoring threshold, and your sales team gets pinged. You set the rules once. The software runs them at whatever scale you eventually reach.

That's the definition. What follows is what it's really for, which workflows to build first, and the mistakes that make automated marketing perform worse than doing nothing at all.

What it is actually for

The name suggests the point is saving time. Time savings are real but secondary. Two things matter more.

Consistency. A person handling follow-ups does the job well on a quiet Tuesday. Badly during a busy launch week. Software does it the same way every time. For anything that runs on a follow-up sequence, consistency usually beats brilliance.

Timing. The gap between someone showing interest and you responding is one of the strongest predictors of whether anything happens next. Manual processes can't respond in four minutes at 11pm. Automation can.

Then the scale benefit. The same nurture sequence serves 50 leads or 50,000 with no extra work.

So the honest framing isn't "automation replaces marketers." It's that the software handles the reliably repeatable parts. People are left doing the parts that need judgement. The offer. The positioning. Deciding what's worth saying at all.

The vocabulary, briefly

Four terms and you can follow any conversation about this.

Trigger is the event that starts a workflow. A form submission. A page visit. A purchase. A date. A period of inactivity.

Workflow (or journey) is the sequence that runs after the trigger. Wait two days, send email, check if opened, branch.

Segmentation is dividing your audience by attribute or behaviour so the same message doesn't go to everyone. This is where most of the performance difference lives, actually.

Lead scoring assigns points for actions like opening an email, visiting pricing, requesting a demo. That way you can tell genuine interest from idle browsing. Useful once you have real sales volume. Premature below that.

The five workflows worth building first

In this order. Each is genuinely useful on its own, so you get value before the whole system is finished. Which matters, because most automation projects get abandoned half-built.

1 · Welcome sequence. Someone joins your list. Three to five emails over two weeks. What you do, the most useful thing you've ever made, some proof, then a soft offer. This is your highest-engagement window. Most businesses waste it entirely with a single confirmation email.

2 · Abandoned action. For e-commerce, an abandoned cart. For services, a form started but not finished. Or a pricing page visited twice with no enquiry. One reminder, timed in hours not days. Usually the highest-ROI automation a business will ever run.

3 · Post-purchase. Confirmation. Then how to get the most from what they bought. Then a review request. Then a relevant next purchase. Existing customers are cheaper to sell to than new ones by a wide margin, and this sequence is almost always the one missing.

4 · Re-engagement. Someone hasn't opened anything in 90 days. Two emails, then remove them. Emailing people who never open damages your deliverability for everyone else. Cleaning the list is a performance improvement, not a loss.

5 · Internal alerts. Not customer-facing but consistently underrated. When a high-value action happens, a demo request, a pricing page visit from a known account, notify a person immediately. Automation's job here is speed. A human does the rest.

Notice that none of these need AI. They're rules. Worth holding onto that distinction: most of the value in marketing automation is deterministic logic, and adding a language model to a rule makes it slower, more expensive and less predictable. The general principle is here.

Where AI genuinely adds something

AI earns its place where the input is unstructured or the output needs language.

Classifying inbound enquiries written in free text, so they route correctly.
Drafting personalised variants of a message at a scale you couldn't write by hand.
Summarising a lead's activity into a sentence a salesperson can read in three seconds.
Predictive scoring based on patterns rather than points you assigned by guessing.

Everything else in the list above is a rule and should stay one.

Choosing a tool

Ignore feature lists. Almost every platform does the five workflows above. Three questions decide it.

1 · Does it connect to what you already have? Your store. Your CRM. Your forms. An automation platform that can't see your data is decoration. This eliminates most options straight away.

2 · What does it cost as your list grows? Pricing usually scales with contacts, and often steeply. Model the cost at ten times your current list before committing, because migrating later is genuinely painful.

3 · Can you build a workflow yourself? If every change needs a developer or an agency, you'll stop making changes. And unmaintained automation degrades quickly.

Rough guidance by stage. An email platform with basic automation is enough at the start. A mid-market platform once you have real segmentation needs. Enterprise suites only when the complexity genuinely justifies the price and the implementation project.

Start smaller than you think. Almost everyone buys too much platform, uses 10% of it, and pays for the rest for years.

The mistakes that make automation worse than nothing

Automating a broken funnel. Automation multiplies whatever you already have. If the offer doesn't convert, you now fail faster and at scale. Fix the offer first.

Sending more because sending is free. The marginal cost of another email is nothing. Which is exactly why people over-send. The real cost isn't the send. It's the unsubscribe and the deliverability damage. More frequency is not more revenue.

"Hi {FirstName}" as personalisation. A merge tag is not personalisation. And a broken one, "Hi ,", is worse than none. Real personalisation is different content for different segments, and that's work.

Never checking it again. Automations break silently. An integration expires. A form field gets renamed. A link rots. A sequence can run wrong for months while looking fine on the dashboard. Every workflow needs a failure alert and a quarterly review. This is the most common and most expensive mistake.

No exit conditions. Someone buys and keeps receiving the "why haven't you bought" sequence. Every workflow needs to know when to stop. And this is the error customers notice most.

Building everything before launching anything. Ship workflow one, watch it for a fortnight, then build workflow two.

A worked example: the welcome sequence

Abstract advice about workflows is hard to act on. So here is one in full. The first thing most businesses should build, and the one most often reduced to a single confirmation email.

Trigger: someone joins the list.
Exit condition: they buy, or they reply. Both should stop the sequence immediately.

Email 1, immediately. Deliver whatever they signed up for, in the first line, with no preamble. Then one sentence on who you are and what to expect. Nothing else. This email has the highest open rate you will ever get. Do not spend it on a newsletter template.

Email 2, day 2. The single most useful thing you have ever published, sent as a gift with no ask attached. The job of this email is to make opening the next one feel worthwhile.

Email 3, day 5. Proof. A case, a result, a specific story of the problem being solved. Still no ask. You're establishing that you know what you are talking about before requesting anything.

Email 4, day 9. A soft, specific offer. Not "buy now". Something like a call, a free audit, a trial. Specific beats general. "Reply with your biggest problem in X and I'll tell you how I'd approach it" outperforms "get in touch".

Email 5, day 14. The graceful close. Recap what you sent, restate the offer once, tell them the frequency they'll now be on. Give them a genuine easy exit here. The people who leave at this point were never going to buy, and removing them protects your deliverability.

What to measure: not opens. Measure how many people who complete this sequence take the offer, against people who joined and got nothing. That difference is the sequence's actual value. Usually large enough to make everything else on this page worth doing.

When you are not ready for it yet

Automation gets sold as something every business needs immediately. It isn't. Buying it early is a common and expensive mistake.

You are not ready if you have no list. Automation acts on people who have given you permission to contact them. With two hundred subscribers, the entire welcome sequence is worth building. But a platform with lead scoring and multi-branch journeys is not. Build the audience first. The tooling is trivial to add later.

You are not ready if you do not know what converts. Automation encodes decisions. If you can't yet say which message moves people, you'll encode a guess and then run it thousands of times. Send a few things manually and watch what happens first.

You are not ready if nobody owns it. Automation is not set-and-forget. A system with no owner drifts into a state where nobody knows what is running or why. That state is worse than manual work, because it looks like it's handled.

You are ready when the same message is being sent by hand repeatedly, and the delay in sending it is costing you. That's the actual trigger. It usually arrives well before people think it does, often at a list of a few hundred rather than a few thousand.

The right first step for most businesses is smaller than the sales pitch suggests. One email platform. One welcome sequence. One abandoned-action reminder. Reviewed monthly. That combination outperforms an expensive suite used at 10% of its capability, and it can be running this week.

How to know if it's working

Not opens. Not clicks. Those are diagnostics, not outcomes.

Revenue per subscriber, the whole point, and rarely measured.
Conversion rate by segment, where segmentation earns its cost.
Time from trigger to conversion, automation should shorten it; if it hasn't, something is wrong.
List health, unsubscribe rate, spam complaints, deliverability. A rising unsubscribe rate is the earliest signal you're over-sending.

Judge each workflow separately. An average across all of them hides the two that are doing everything and the six that are doing nothing.

FAQs

What is marketing automation in simple terms?

Software that performs marketing actions automatically when someone does something. Joins your list, abandons a cart, stops engaging. You define the rules once and the software runs them every time, consistently and immediately, at any scale.

Is marketing automation just email?

Email is the most common channel because it's cheap and permission-based. The same logic drives SMS, push notifications, in-app messages, ad audience membership and internal alerts to your sales team. The trigger-and-workflow model is the same regardless of channel.

Do I need AI for marketing automation?

No. Most of the value is deterministic rules. If this happens, do that. AI helps specifically where input is unstructured (classifying free-text enquiries) or output needs language (drafting variants, summarising activity). Adding a model to a rule makes it slower, costlier and less predictable.

What should I automate first?

A welcome sequence for new subscribers and an abandoned-action reminder. Between them they cover the highest-intent moments in most businesses, and they work independently so you get value before the system is complete.

How much does marketing automation cost?

Entry-level email platforms with basic automation start very cheap and scale with list size. The trap is that pricing usually scales steeply. Model your cost at ten times your current list before committing, because migration later is painful.

Why is my marketing automation not working?

Most often one of four things. The underlying offer doesn't convert and automation is scaling the failure. You're over-sending because sending is free. A workflow broke silently and nobody checked. Or there are no exit conditions, so people receive sequences that no longer apply to them.

Key takeaways

Automation's real value is consistency and timing. Time saved is a bonus.
Build in order: welcome, abandoned action, post-purchase, re-engagement, internal alerts.
Most of it is rules, not AI. Use a model only for unstructured input or language output.
Choose a tool on integrations, cost at scale, and whether you can edit it yourself.
Automation multiplies what you already have. Fix the offer before scaling it.
Every workflow needs an exit condition, a failure alert, and a quarterly review.

What Is Advertising? And Why Digital Marketing Took Over

Prabhash Jha — Thu, 24 Sep 2026 14:12:00 +0000

Advertising is paid communication intended to persuade someone to do something. Buy a product. Use a service. Vote. Donate. Change their mind about anything. The "paid" part is what defines it. If you earn the attention through word of mouth or coverage, that's publicity. If you bought the space, that's advertising.

An advertisement is anything you place in that bought space to draw attention toward what's being sold.

That definition hasn't moved in four hundred years, and you'll find it stated more or less identically on a hundred pages. What's actually worth your time is the second half of the question. Why digital took the budget. The usual answer stops at four advantages and never mentions that each of them arrived with a bill attached. Advantages are real. So are the bills. And the bills are what you'll actually spend your time managing.

A short, accurate history

Paid advertising is older than most people assume. Printed advertisements appear in English newspapers in the 17th century, not long after commercial printing made newspapers viable at all. Through the 18th and 19th centuries, print was the dominant paid channel, and the first advertising agencies emerged in the 1800s. Initially as brokers reselling newspaper space.

The 20th century added the broadcast channels. Commercial radio in the 1920s. Television advertising from the 1940s. Outdoor and direct mail grew alongside them.

Digital came late. The web became publicly usable in the early 1990s, and the first banner ads ran on 27 October 1994, on HotWired. Fourteen clickable 468×60-pixel ads. Advertisers included AT&T, MCI, Volvo, Club Med and Zima (Guinness World Records: first banner ad). Search advertising followed toward the end of that decade.

So the entire digital advertising industry is roughly thirty years old. Younger than most of the people running it. That timeline is worth getting right, honestly, because it explains the shape of the industry. Digital didn't gradually replace traditional media. It arrived suddenly and took the budget in about two decades. An industry that grew that fast never had a settled period. Which is why the tooling changes under you every eighteen months, and why nobody in it has a career's worth of stable precedent to draw on.

Why digital took the budget

Not because it's inherently better. Because it fixed four specific weaknesses of traditional channels.

You can reach a defined group, not a broad one. A newspaper sells you its whole readership. Digital lets you reach people by location, device, interest, or the fact that they visited your site last week and left. That's the entire basis of retargeting.
The cost of entry collapsed. A television campaign has a production floor and a media floor before anything runs. A digital campaign can start with a small daily budget.
You can see what happened. In print or broadcast, you infer results from sales lift and surveys. In digital you can, imperfectly, trace a click through to a purchase.
You can change it while it runs. A print ad is fixed once it goes to press. A digital campaign can be paused, re-targeted, rewritten or re-budgeted the same afternoon.

Every one of those is true. Now the part the vendor pages leave out.

The bill for targeting: everyone is aiming at the same people

Precision targeting stopped being an advantage the moment it became universal. What remains is a cost.

Logic that identifies your best prospects is available to every competitor buying the same inventory. And it identifies the same people. So the most valuable audiences are the most contested ones, and the price of reaching them reflects that. You aren't paying for precision. You're paying for the right to compete for a smaller, more expensive group. Broad targeting wastes impressions on people who won't buy. Narrow targeting wastes money on people everyone else also wants. There's no setting that avoids both. Choosing between them is most of the job.

Second cost is subtler and does more damage. Narrow targeting shrinks the pool you're learning from, and small pools produce noisy results. A tightly targeted campaign will show you dramatic week-to-week swings that are mostly randomness. And the natural response, react to every swing, makes performance worse. Much of what gets discussed as "the algorithm behaving strangely" is honestly a sample-size problem wearing a costume.

The bill for cheap entry: the auction now contains everyone

The barrier that kept you out of television also kept your competitors out. Removing it let you in. And let all of them in with you.

This isn't a metaphor about a crowded market. It's mechanically what happens, because most digital advertising is sold by auction rather than by rate card. Whether your ad shows, and where, is decided by Ad Rank. Google describes this as combining your bid, the quality of your ad and landing page, the Ad Rank thresholds, the competitiveness of the auction at that moment, and the context of the search (Google Ads Help: how the Google Ads auction works). Note the third and fourth items. Two of the inputs to your cost are things other advertisers do.

Two consequences follow. Neither is optional.

Your costs rise for reasons that have nothing to do with you. A well-funded competitor entering your category raises your prices without you changing a setting. Rising cost per click isn't automatically a sign that something is broken on your side. And treating it as one leads to a lot of pointless rebuilding of campaigns that were fine.

Quality is a real discount, not a platitude. Same documentation is explicit that even if a competitor bids higher, you can win a better position at a lower price with higher-quality ads and landing pages. That's the one input in the formula that's entirely yours. Which makes it the one worth working on when the auction gets more expensive. Also why the page you send traffic to is an advertising cost decision and not just a design one. And why the whole funnel fails after the click at least as often as before it.

If you want the machinery underneath all this, programmatic advertising is the same auction running in about a hundred milliseconds while a page loads.

The bill for measurement: budgets drift toward what is countable

This is the most expensive bill of the four. And almost nobody bills it to the right account.

Digital didn't make marketing measurable. It made some of marketing measurable, cheaply and immediately, and left the rest exactly as hard to measure as it always was. Clicks and conversions are easy. The effect of someone seeing your name for the eleventh time and finally trusting it enough to search for you isn't.

What happens next is a structural drift rather than anybody's mistake. Every budget review compares options. The measurable option always has a number attached. The unmeasurable one has an argument. The number wins, quarter after quarter, until the entire budget sits on the activities that can be counted. Which are mostly the ones that harvest demand that already exists rather than create any. Then demand stops growing. Harvesting channels get more expensive because you're competing harder for a fixed pool. And the numbers on your best-measured channel slowly deteriorate for reasons that are invisible inside that channel. That's the situation where brand becomes the only lever left. And it's usually diagnosed about a year late.

There's a second measurement bill. Numbers you do get are less precise than they appear. Platform figures are credit assigned by an attribution model over a conversion window, not a count of sales caused, and the gap between what the platform reports and what your bank received is normal rather than exceptional. The sheet that tells you whether your marketing makes money goes through how to reconcile the two. Short version: use platform numbers to compare campaigns against each other, and money that actually arrived to judge whether the channel deserves a budget.

Anyone who tells you their attribution is exact is selling something. Anyone who concludes from that that measurement is worthless is selling something else.

The bill for control: the ability to change it is the thing most abused

Being able to change a campaign the same afternoon is genuinely valuable. It's also the advantage most consistently turned into a liability by the people who have it.

Modern bidding is a system that learns from your data, and learning takes both time and volume. Google states that after a change to a bid strategy it can take up to three weeks, or one to two conversion cycles, for the strategy to calibrate to the new objective. Length depends on how many conversions the campaign is getting (Google Ads Help: duration of the learning period). Changes to the strategy, its settings, or the campaign's composition all put it back into learning.

Now put that next to how an anxious advertiser actually behaves. Results look poor on day three, so the target gets adjusted. Day six is still poor, because it's measuring a strategy that restarted on day three, so the targeting changes too. Each intervention is reasonable on its own. Campaign never completes a calibration. So it never produces the stable result that would have justified leaving it alone. The advertiser concludes the channel doesn't work for their business. Channel was never allowed to finish.

Discipline that fixes this is unglamorous and hard to follow. Decide the observation window before you launch. Write it down. And don't touch the campaign inside it unless you're losing money faster than you agreed to. Set that budget cap at something you can genuinely afford to lose, precisely so the decision to wait costs you nothing you need. In print, patience was enforced by the printing press. In digital nothing enforces it. So you have to.

Same principle applies to how often you change bids at all. Small, infrequent adjustments with a week between them behave very differently from daily fiddling. Even when the total adjustment is identical.

What digital did not fix

Being honest about the limits is more useful than the sales pitch.

Attention is harder to hold, not easier. Cheaper access to people came with far more competition for the same seconds. Cost of reaching someone fell. Cost of being remembered by them didn't.

The channel doesn't fix the offer. No amount of targeting rescues a product people don't want at a price they won't pay. This is the most common root cause of a campaign that "isn't working". And the one people are least willing to check first, because every other explanation is cheaper to act on.

Nothing about digital changed the arithmetic. You still have to be worth more to a customer than they cost to acquire. Precision changed how quickly you find out. It didn't change what you find out.

Owned channels still beat rented ones over time. Everything in the auction is rented, priced by competitors, and repriced without your consent. Search traffic you earn and an email list you own are slower to build and don't have that property.

So where does a budget actually go now?

Start from what the four bills imply rather than from a channel list. A channel list is how you end up spending money in five places badly.

If you have no idea what a customer is worth to you, that's the only job. Everything downstream is guesswork without it, and the guesswork is expensive in the auction because the auction charges you for it in real time.

Pick one paid channel and stay in it long enough to learn it. Not because the others are bad. Because the learning period is real, attention is finite, and a small budget divided across three platforms produces three underpowered campaigns and no conclusions.

Spend something on the part you can't measure well, deliberately and as a policy. Not because it's fashionable. Because the drift toward countable things is structural and won't correct itself. Decide the split when you're calm, not during a quarter when the measurable number looks better.

Treat rising costs as weather, not as failure. Your job isn't to stop the auction getting more expensive. It's to have better margins, a better landing page and a longer customer relationship than the advertiser next to you. So that you can survive a price you both have to pay.

Where to start if you're starting

If you're new to this, the order that actually works:

Know what a customer is worth to you. Everything downstream is guesswork without it. Cash flow vs profit is a useful companion here.
Learn the vocabulary you'll be billed in. CPC, CPM, CTR, CPA, ROAS. Marketing metrics explained covers these in plain English.
Pick one channel and stay there long enough to learn it. Google Ads vs Meta Ads covers how to choose the first one.
Agree your observation window and your maximum loss before you launch. And hold to both. This is the step that separates people who learn something from their first campaign from people who only learn that advertising is confusing.

Advertising hasn't fundamentally changed in four hundred years. Pay for attention, and try to be worth it. Digital only changed how precisely you can aim, how quickly you find out whether you missed, and how many other people are aiming at exactly the same spot.

FAQs

What is advertising, exactly?

Paid communication intended to persuade someone to do something. Buy a product, use a service, vote, donate, change their mind. The "paid" part is what defines it. If you earn the attention through word of mouth or press coverage, that's publicity. If you bought the space, it's advertising, and the advertisement is whatever you place in that bought space.

What is the difference between advertising and marketing?

Advertising is one activity inside marketing. Marketing covers what you sell, to whom, at what price, and how it reaches them. Advertising is specifically the part where you pay for attention. Distinction matters practically, because a lot of problems that get handed to advertising (a price the market rejects, an offer nobody wants) can't be solved by buying more attention for them.

How old is digital advertising?

Roughly thirty years. The first banner ads ran on 27 October 1994 on HotWired, and search advertising followed toward the end of that decade. Printed advertisements, by contrast, appear in English newspapers in the 17th century, and the first advertising agencies emerged in the 1800s as brokers reselling newspaper space.

Why did digital take the budget from traditional media?

Not because it's inherently better, but because it fixed four specific weaknesses. You can reach a defined group rather than a whole readership. Cost of entry collapsed, so a small daily budget can run a campaign. You can trace a click through to a purchase, imperfectly. And you can pause, re-target or rewrite a campaign the same afternoon.

Why do my ad costs keep rising when I haven't changed anything?

Because most digital advertising is sold by auction, and two of the inputs to your price (how competitive the auction is and what other advertisers are doing) are outside your control. A well-funded competitor entering your category raises your costs without you touching a setting. The input that is yours is quality: better ads and landing pages can win a higher position at a lower price than a higher bid does.

How long should I leave a campaign alone before judging it?

Longer than feels comfortable. Bid strategies need time and conversion volume to calibrate, and changing the strategy, its settings or the campaign's composition restarts that process. Practical approach is to decide the observation window and the maximum you're willing to lose before launching. Then leave the campaign alone inside that window unless the loss limit is hit.

Is digital advertising better than traditional?

For most small and medium businesses, yes. Mainly because of the cost of entry and the ability to stop paying for something that isn't working. But "better" hides a trade. Digital gives you precision, speed and partial measurement, and charges you in auction competition, noisy small samples, and a persistent pull toward whatever is easiest to count.

Key takeaways

Advertising is paid communication intended to persuade. If you bought the space, it's advertising. If you earned it, it's publicity.
Digital advertising is only about thirty years old. The first banner ads ran in October 1994. That's why nothing in it stays settled for long.
Digital won on four things: defined audiences, a collapsed cost of entry, traceable results, and the ability to change a campaign mid-flight.
Each of the four came with a bill: contested audiences, an auction containing every competitor, budgets drifting toward what is countable, and the freedom to interfere with your own campaigns.
Rising costs are frequently caused by other advertisers, not by anything you did. Quality is the one input in the auction that's entirely yours.
Measurement made some of marketing countable and left the rest as hard as ever. That asymmetry, not anyone's bad judgement, is what quietly defunds demand creation.
Bid strategies need time and volume to calibrate, and every significant change restarts the clock. Decide the observation window before you launch.
The channel never fixes the offer, and no amount of targeting rescues a product people don't want at a price they won't pay.

What Is a Marketing Funnel? A Simple Guide With Real Examples

Prabhash Jha — Thu, 24 Sep 2026 14:11:59 +0000

Most guides to the marketing funnel try to describe how a customer moves through it. That description is wrong. And knowing it's wrong is what makes the funnel actually usable.

Nobody has walked through a funnel. Nobody, ever. People search for a comparison query and land halfway down. They read three pages, disappear for two months, come back through a branded search on a different device, and buy in four minutes. They buy because a colleague told them to, and no analytics tool anywhere ever records that conversation. If the funnel were a map of real buying behaviour, it would be a badly drawn one.

So here's the plain definition first, then the argument. A marketing funnel is a set of counting points between "someone found you" and "someone paid you". At each point you count how many people got that far. The gaps between the counts are where your money leaks out. That's the whole thing.

That's why serious marketers still use it. It isn't a map. It's a ruler. And its whole job is to convert a vague feeling, "marketing isn't working", into a specific answerable question: which of my three step rates is worst right now, and how much room does it actually have to move?

The funnel is a ruler, not a map

A thermometer doesn't describe the weather. It measures one property of it, accurately and repeatably, in a unit you can compare against yesterday and last week. That limitation is exactly what makes it useful. Nobody complains that a thermometer fails to capture wind chill or humidity.

The funnel works the same way. It takes an unobservable process, a hundred strangers deciding whether to give you money, and installs four counting posts along it. Each post asks one question, and only one. How many people got this far?

You don't need the funnel to be psychologically true. You need it to be countable. Once you accept that, the whole "is the funnel dead" argument turns into noise, and a more useful question replaces it. What is the conversion rate between each pair of posts, and which one has the most room to move this quarter?

The four stages of a marketing funnel, and where they came from

Awareness. The person now knows a solution to their problem exists. They've seen your ad, found your article, watched your video, heard your name.

Interest / Consideration. They're evaluating. They want to know what the options are, whether you're credible, and whether this category of solution is even right for them.

Desire / Intent. They've narrowed it down. Now they're checking fit for their situation. Price. Terms. Whether it works for a business their size. Whether they'll look stupid for choosing it.

Action. They do the thing. Buy, book, submit, sign.

The lineage runs back to advertising writing around the turn of the twentieth century, usually attributed to E. St. Elmo Lewis, who set out a hierarchy along the lines of attention, interest and desire. "Action" got appended later. The acronym AIDA is a later coinage. And the funnel shape was bolted on decades after that. The cone diagram is commonly credited to William W. Townsend's Bond Salesmanship in 1924. Lavidge and Steiner formalised a similar "hierarchy of effects" model in 1961. Anyone who tells you a single person invented the marketing funnel in 1898 is compressing three separate ideas from three separate decades.

Why does the history matter? For one reason. These stages were never derived from data about how people decide. They were a copywriter's convenient sequence. Treat them as labels for measurement checkpoints, not as psychological states you can observe. You can't see desire. You can see a click on a pricing page.

Why the funnel is wrong as a customer journey, and why that doesn't matter

The honest counter-models are worth knowing, because they stop you designing your marketing around a fiction.

McKinsey's Consumer Decision Journey (McKinsey Quarterly, 2009) argued something buyers do not do. They do not narrow a set of brands steadily downwards. What they do instead is often expand their consideration set partway through, add brands they'd never thought about, and loop between evaluating and exploring. Google's "messy middle" research (2020) described the same territory. A repeated back-and-forth between exploration and evaluation that continues until something tips. No fixed number of steps. No fixed order.

Both are more accurate than the funnel. Neither replaces it, because neither hands you a number you can trend week over week.

Take a business with 10,000 monthly visitors. A chunk of them arrive already halfway down. They typed a comparison query, so they know the category, know two competitors, and skipped "awareness" entirely as far as you're concerned. Another chunk bounce, return three weeks later on a different device, and get counted as two people. Some buy on the first visit because a colleague had already sold them before they even arrived on your site.

None of that breaks the ruler. Honestly. The funnel doesn't claim these people moved in some tidy order. What it claims is that of everyone who reached checkpoint two, a certain proportion reached checkpoint three. That proportion is real. It's measurable. It's comparable to last month's figure. And when it drops you've learned something worth acting on. The measurement survives intact even though the story it sits inside is false.

Marketing funnel stages mapped to content, channel and one metric each

The right-hand column is the one people skip, and it's the one that matters. Each stage has exactly one number that tells you whether it's working.

Stage	What the person is actually thinking	Content that fits	Channel that fits	The one metric
Awareness	"I have a problem", or doesn't yet know it's a problem	Explainers, short video, problem-phrased search content, PR	Organic search on informational queries, social, YouTube, broad paid reach	% of arrivals who take any next step, not impressions
Interest / Consideration	"What are my options, and is this legit?"	Comparisons, category guides, case studies, reviews, product and service pages	Comparison search results, communities, retargeting, email capture, review sites	% who reach a product, solution or pricing page
Desire / Intent	"Is this the right one for my situation?"	Pricing, demos, trials, spec detail, objection FAQs, testimonials	Branded search, email nurture, retargeting, sales conversation	% who start the conversion action (cart, demo request, trial)
Action	"Can I do this without friction or risk?"	Checkout, order form, guarantee and refund copy, proposal	Site checkout, cart-abandon email, sales rep, transactional email	Completion rate of the started action
Loop (after the sale)	"Was it worth it? Would I tell anyone?"	Onboarding, help content, referral prompts, community	Lifecycle email, in-product messaging, support	Repeat rate, referral rate, lifetime value

Note what the awareness metric is not. Impressions and reach measure delivery, not progression. If you can't say how many people took a next step, you've measured spend. Not awareness.

A worked example: 10,000 visitors and three step rates

One month. One business. Invented numbers chosen to be easy to check. Every figure below is arithmetic you can redo yourself.

10,000 visitors arrive
2,000 view a product or service page: step rate 20%
400 start the conversion action: step rate 20%
100 complete the purchase: step rate 25%

Overall conversion: 0.20 × 0.20 × 0.25 = 1.0%. At an average order value of ₹2,000, that's ₹2,00,000 a month.

Two contradictory pieces of advice circulate about what to fix first. One says start at the top, because the numbers up there are bigger. The other says start at the bottom, because it's closest to the money. The arithmetic below says position in the funnel is not the variable at all.

Where an equal improvement lands: nowhere in particular

Apply the same 25% relative improvement to each stage, one at a time.

Fix applied	Resulting chain	Sales	Revenue
Stage 1: 20% → 25%	10,000 → 2,500 → 500 → 125	125	₹2,50,000
Stage 2: 20% → 25%	10,000 → 2,000 → 500 → 125	125	₹2,50,000
Stage 3: 25% → 31.25%	10,000 → 2,000 → 400 → 125	125	₹2,50,000

Identical. Each gives +25 sales, +₹50,000 a month, +₹6,00,000 a year. This falls straight out of the structure. The funnel is a product of rates, and multiplication doesn't care about order. A 25% lift anywhere multiplies the whole chain by 1.25.

That's not a revelation. But it kills both pieces of folk wisdom in one line. Neither "top first" nor "bottom first" is a rule. Whatever makes one stage the better bet has to come from somewhere other than its position.

Where the room actually is: low base rates

Two things vary between stages, and both track the base rate rather than the position.

First, marketers don't think in relative lifts. They think in percentage points. "I'll get the pricing-page rate up five points" is how the conversation actually goes. So apply +5 percentage points to each stage instead.

Fix applied	Sales	vs baseline
Stage 1: 20% → 25%	125	+25
Stage 2: 20% → 25%	125	+25
Stage 3: 25% → 30%	120	+20

The third stage underperforms. Not because it's at the bottom, but because it starts from a higher base. To match a five-point move at a 20% stage, the 25% stage needs a 6.25-point move (25% × 1.25 = 31.25%). A fixed percentage-point gain is a bigger relative gain when the base rate is lower.

Second, the same fact sets a hard ceiling. Each stage's maximum possible upside is 1 ÷ its current rate. Perfect that stage and change nothing else, and that's your multiplier.

Stage 1 at 20% → perfect gives 500 sales: 5× headroom
Stage 2 at 20% → perfect gives 500 sales: 5× headroom
Stage 3 at 25% → perfect gives 400 sales: 4× headroom

Now take a second business with the same 100 sales but a strong close. 10,000 visitors → 800 reach pricing (8%) → 200 enquire (25%) → 100 buy (50%). Overall conversion is still exactly 1.0%.

Perfecting the close here, 50% to 100%, meaning literally everyone who enquires buys, caps out at 200 sales. Two times headroom, and you can never reach it. Meanwhile, moving the pricing-page rate from 8% to 12% is a routine four-point improvement. Better internal linking. A clearer navigation label. A service page that loads in under two seconds. That gives 10,000 → 1,200 → 300 → 150 sales.

So a routine fix delivers +50 sales, and an impossible one delivers +100. Push that pricing-page rate from 8% to 16% and you get 10,000 → 1,600 → 400 → 200 sales, matching the impossible outcome exactly. The reason to leave the close alone in this business isn't that it sits at the bottom. It's that it has no room left.

How fast you can learn: volume decides

There's a second constraint that rarely appears in funnel articles. To know whether a change worked, you need enough people through the test. A standard back-of-envelope sample size for roughly 80% power at 95% confidence, two-sided, is:

n per variant ≈ 16 × p̄(1 − p̄) ÷ δ²

where p̄ is the average of the two rates you're comparing and δ is the absolute difference between them. It's an approximation rather than an exact power calculation, but it's close enough to plan with.

The 400-person stage, testing 25% → 30%: p̄ = 0.275, δ = 0.05. That gives 16 × 0.275 × 0.725 ÷ 0.0025 ≈ 1,280 per variant, or 2,560 in total. Only 400 people reach that step each month. Time to a readable result: 6.4 months.

The 10,000-person stage, testing 20% → 24%: p̄ = 0.22, δ = 0.04. That gives 16 × 0.22 × 0.78 ÷ 0.0016 ≈ 1,720 per variant, or 3,440 in total. Ten thousand people arrive each month. Time to a readable result: about ten days.

Same statistical standard. Similar relative gain. And one takes nearly twenty times longer to read than the other.

Be precise about what this does and doesn't say, though. The speed advantage belongs to steps with high volume passing through them, usually on-site, page-to-page steps. It does not belong to top-of-funnel marketing. New ad campaigns, SEO and content are the slowest things in the business to evaluate, because the response arrives over months and the traffic mix shifts underneath you while you wait. Volume through the step is the variable. Not the stage label.

So which leak should you fix first?

The rule that survives all three tests: attack the lowest step rate that has real volume behind it.

Lowest rate, because a fixed percentage-point gain is worth more on a low base, and a low rate strongly implies unclaimed headroom sitting there. Real volume, because volume is what buys you a readable test in weeks rather than in full seasons or entire quarters.

This also explains a very common trap I see. Checkout is instrumented, immediate and satisfying to test, so teams pile into it by default and stay there. But measurability is not the same thing as leverage. If your close is already at 50%, you have at most a 2× ceiling there. And the low traffic reaching that final step means you'll wait half a year to learn whether your new button copy did anything at all.

How to find your own funnel leak in an afternoon

Five steps. A spreadsheet is enough. You don't need new software.

1. Define exactly one countable event per stage. Not a feeling. A specific loggable action. Page view on any service page. Click on "Request a quote". Form submitted. Order confirmed. If you can't name the click that counts, you don't have a funnel. You have a diagram.

2. Pull 30 to 90 days of raw counts for each event. Long enough that weekday and payday cycles wash out. Analytics funnel exploration reports, CRM stage counts and ad platform conversion columns will each hold part of it. Put the four numbers in one column.

3. Compute the three step rates. Stage two ÷ stage one. Stage three ÷ stage two. Stage four ÷ stage three. Three numbers. That is the whole diagnostic.

4. Compare each rate to its realistic ceiling rather than to a published benchmark. Ask what this rate would be if the step were as good as it could plausibly get for a business like yours. A 6% pricing-page rate has obvious room. A 55% checkout completion rate does not.

5. Segment by traffic source before concluding anything. This is the step people skip, and it's the one that most often reverses the answer.

Why step rates beat overall conversion rate and raw counts

Your overall conversion rate, 1.0% in the example, tells you a problem exists. It cannot tell you where. The two businesses above had identical 1.0% overall rates, completely different problems and completely different correct actions.

Absolute counts are worse. "We got 400 enquiries last month, up from 350" is unreadable, because you can't tell whether the funnel improved or you simply bought more traffic. Counts move when traffic moves, so they can't be trended as performance. Rates hold traffic volume constant by construction.

And blended rates hide broken segments. Take the same 10,000 visitors, now split by source:

Organic search: 6,000 visitors → 90 sales = 1.5%
Paid social: 4,000 visitors → 10 sales = 0.25%
Blended: 100 ÷ 10,000 = 1.0%

The blended number is technically correct and practically useless. It conceals a paid channel converting four times worse than the site average and six times worse than organic. Which is almost certainly where the money is going.

This is straightforward aggregation, not yet Simpson's paradox. The paradox is the sharper case, where a change wins in every single segment and still loses in the aggregate because the mix between segments shifted underneath it. Both problems have the same fix. Segment before you optimise, not after. Segment by source first, then by device, since mobile and desktop nearly always behave differently. If you want the underlying definitions straight before you build the sheet, the plain-English guide to CPC, CPM, CTR, CPA and ROAS covers the vocabulary this section assumes.

Marketing funnel vs sales funnel, and the handover in between

A marketing funnel usually covers everything up to a qualified lead or an online purchase. A sales funnel covers what happens once a human is involved. Qualification. Proposal. Negotiation. Close. And its stages are CRM pipeline stages. In self-serve ecommerce they're one instrument. In B2B they're two instruments with a handover point, and the handover is where the numbers usually break. Marketing counts an MQL. Sales counts an opportunity. Nobody reconciles the two. Both teams report a good quarter while revenue is flat.

The second B2B problem is timing, and it produces false alarms constantly. Say the median time from first visit to closed deal is 60 days. This month you had 10,000 visitors and closed 100 deals. But most of those deals came from the cohort that arrived two months ago, when traffic was 6,000.

Naive same-month rate: 100 ÷ 10,000 = 1.0%
Cohort rate: 100 ÷ 6,000 = 1.67%

Nothing about the funnel changed. Traffic grew, and the naive calculation reported a 40% collapse in conversion because it divided mature sales by immature traffic. Growing businesses see this constantly and go looking for a broken landing page that doesn't exist.

Two rules follow. Your measurement window has to be at least as long as the time from first touch to purchase for most buyers. And cohorts must be allowed to mature. A cohort's conversion rate only stabilises after roughly one full sales cycle has passed, so the most recent month or two on your chart is always understated and should never be used to trigger a decision.

What content and channel suits each stage

Read the table above as a matching problem rather than a content list. A mismatch is almost never a bad asset. It's a good asset delivered at the wrong step.

A pricing page shown to someone who doesn't yet know the category exists will convert at close to zero. Not because the pricing is wrong, but because "how much" isn't the question they're holding. A 3,000-word explainer served to someone with a card in their hand and a competitor's tab open is equally useless, for the mirror-image reason. You answered a question they finished asking last week.

Three practical consequences.

The same channel appears at multiple stages doing entirely different jobs. Search is the clearest case. "Why is my ad account not spending" is an awareness query. "Meta ads agency Delhi pricing" is an intent query. Same channel. Same tool. Opposite content requirements. And you should measure them separately. The SEO basics guide covers how query intent maps to page type.

Retargeting is a stage-two and stage-three instrument. Its entire mechanism is that the person was already counted at an earlier checkpoint. So using it to build awareness is a category error. There's nobody in the audience yet. If that distinction is new, start with the beginner's guide to retargeting.

Email is the only asset that lets you re-enter the funnel on your own terms. Every other channel requires you to pay again or wait for the person to come back. That's why building an email list from zero is worth more than its stage-two conversion rate suggests. It changes what the second attempt costs.

Choosing between the two big paid platforms is a stage-one and stage-three decision more than a general one. Google Ads versus Meta Ads for beginners is the comparison to read before you split a budget.

Funnel vs flywheel: why retention is a top-of-funnel weapon

A funnel treats the sale as the terminal event. The person exits at the bottom and the model stops caring. A loop, HubSpot popularised the flywheel framing at INBOUND in 2018 though the underlying idea is older than the label, treats the customer as an input to the next turn. They buy again. They refer. They leave a review that shortens someone else's evaluation.

"Retention adds revenue" is true and obvious. The consequence worth understanding is that retention also changes what you can afford to pay for new customers.

Back to the numbers. One hundred customers at ₹2,000 gives ₹2,00,000, and under the funnel view that's the end. Now suppose 30% of them buy once more. Lifetime value becomes ₹2,000 + (0.30 × ₹2,000) = ₹2,600. The same 100 acquisitions now produce ₹2,60,000. 30% more revenue with zero additional acquisition spend.

Here's the leveraged part. Use the common venture heuristic of spending at most a third of lifetime value to acquire a customer. A rule of thumb, not a law, but a widely used one. Your allowable cost per customer rises from ₹666.67 to ₹866.67. You can now bid 30% higher than you could last quarter for the exact same traffic your competitor is bidding on. Nothing about your ads changed. You improved the back end and it bought you the front end.

That's an argument about auction economics rather than sentiment. One caution though. Check that the extra revenue arrives before the acquisition cost is due. A healthier lifetime value with a longer payback period can still starve you, and the distinction between cash flow and profit is exactly where this goes wrong for growing businesses.

Three mistakes that make a funnel useless

Treating the funnel as literal. This produces a content calendar built for a linear path nobody walks, followed by genuine confusion when "top-of-funnel" traffic buys in the first session and "bottom-of-funnel" traffic disappears for a quarter. The stages are checkpoints. Not a script. Build for the questions, and expect people to arrive at checkpoint three without ever passing checkpoint one.

No defined conversion event per stage. By far the most common failure. Someone shows you a funnel diagram with four coloured bands and no numbers, or numbers pulled from four incompatible tools counting four different things. If you can't point at a specific logged action for each stage, and confirm that the same visitor identity carries between them, you have a slide. Not an instrument.

Changing the definitions while you trend the numbers. Subtler, and very common. You widen "reached a product page" to include a CTA click from the blog. The rate jumps four points. Someone reports a win. But the jump was definitional, not behavioural. Version your event definitions with the date they changed, and treat any change as a break in the series rather than a data point in it. Same applies to attribution windows and to any switch of analytics platform.

What to do on Monday

Define four events. Pull four numbers for the last 90 days, or longer if your sales cycle demands it. Compute the three step rates. Split each by traffic source, and then by device. Circle the lowest rate that has meaningful volume behind it. That's your leak.

Then ship exactly one change against that rate. Not five, not three. Because five changes at once means you learn nothing about any of them individually. Estimate the sample size with the formula above before you start, so you know in advance whether you're waiting ten days or six months for the answer. If the answer is six months, you've chosen the wrong stage to work on. Re-measure at the end of that window, not before it, and not halfway through. Then repeat the whole exercise with the next-worst rate.

If you want the financial view alongside it, what each of these numbers is worth to you after costs, the profitability sheet is the companion to this exercise.

FAQs

What is a marketing funnel in simple terms?

It's a set of counting points between "someone found you" and "someone paid you". At each point you count how many people got that far, and the gaps show where you're losing them. It's a measuring tool rather than a description of how customers behave. Real buying is messy, non-linear, and frequently starts halfway down.

What are the 4 stages of a marketing funnel?

Awareness (they know you exist). Interest or Consideration (they're comparing options). Desire or Intent (they're checking fit for their situation). And Action (they buy). The wording traces back to early-1900s advertising writing, usually credited to E. St. Elmo Lewis. The funnel shape was added decades later. Many teams add a fifth retention stage.

What is a good conversion rate between funnel stages?

There is no universal good number. It varies enormously by category, price point and traffic source. The useful comparison is your own previous period, segmented by source, measured over a window at least as long as your sales cycle. A rate improving against your own matched history tells you more than one beating a stranger's average.

What is the difference between a marketing funnel and a sales funnel?

A marketing funnel usually covers everything up to a qualified lead or an online purchase. A sales funnel covers what happens after a human gets involved. Qualification, proposal, negotiation, close. Its stages are CRM pipeline stages. In self-serve ecommerce they're the same thing. In B2B they're two connected instruments with a handover point where numbers commonly break.

How do I know where my marketing funnel is leaking?

Define one countable event per stage. Pull 30 to 90 days of counts. Then divide each stage by the one before it to get three step rates. Segment those by traffic source. The lowest rate with meaningful volume behind it is your leak. Absolute counts can't tell you this, because counts move whenever traffic volume moves.

Is the marketing funnel dead?

No. But it was never alive in the way people claim. As a description of buying behaviour it has always been wrong, and McKinsey's decision journey and Google's messy-middle research describe reality better. As a measurement instrument that turns "marketing isn't working" into "stage two is at 8%", it remains the most practical tool available.

Key takeaways

The funnel fails as a description of buying behaviour and succeeds as a measurement instrument, because it converts a vague complaint into three step rates you can act on.
Position in the funnel decides nothing. An identical 25% relative lift at any stage took the worked example from 100 sales to 125 every time.
What varies is base rate and volume. A fixed percentage-point gain is worth more on a low base, a stage's maximum upside is 1 ÷ its current rate, and a 400-person step needs 6.4 months to read a test that a 10,000-person step reads in ten days.
Fix the lowest step rate that has real volume behind it, which is often neither the top nor the bottom, and remember that the easiest stage to instrument is frequently the one with the least room left.
Blended rates conceal broken segments. A paid channel converting at 0.25% sat inside a healthy-looking 1.0% site average, so segment by source before you optimise anything, and cohort by first-touch month whenever your sales cycle runs longer than your reporting window.
Retention buys you front-end firepower. A 30% repeat rate lifted lifetime value from ₹2,000 to ₹2,600 and allowable acquisition cost from ₹666.67 to ₹866.67, letting you outbid competitors for identical traffic.

The Skills That Actually Matter in the Age of AI

Prabhash Jha — Wed, 23 Sep 2026 14:13:25 +0000

The standard advice for surviving AI is "learn to code". It's close to backwards.

Code is text. It has a machine-checkable output. It compiles or it doesn't. The tests pass or they don't. The type checker screams or stays quiet. And there's an enormous public corpus of it, written in public, with the answers attached. Text in, verifiable result out, mountains of training data. That combination describes exactly the shape of work machines get good at fastest. Telling someone to learn programming because it's safe from automation is like telling someone in 1995 to become a typist because computers were coming.

That doesn't mean programming is a bad thing to learn. It's a superb way to build precise thinking. Someone has to direct the machines. What's wrong is the reason people give for learning it, and the wrong reason leads you to the wrong version of the skill. Memorising syntax instead of learning to specify a system, argue about trade-offs, and own the outcome when it breaks in production at 2am.

So the real question isn't "which job title survives?" It's narrower and more useful: which parts of my work have a checkable answer, and which parts require someone to carry the consequence of being wrong? The first part is getting cheap. The second part is where your income is going to come from.

Why "learn to code to stay safe from AI" gets it backwards

Work automates in a predictable order. Difficulty isn't the variable people think it is. Chess fell before laundry-folding. Radiology image classification advanced faster than plumbing. The pattern isn't "easy tasks first". It's "verifiable tasks first".

Three conditions make a task cheap for a machine to learn:

The output can be scored automatically. A test suite, a compiler, a scoring function, a right answer at the back of the book. Without a signal, there's nothing to train against.
The input and output are both text or pixels. No physical world, no permissions, no meeting three people to find out why the last attempt failed.
The corpus is public and huge. Millions of examples, freely scraped, with the reasoning included.

Software engineering ticks all three harder than almost any other white-collar craft. Most legal drafting ticks two. Financial modelling ticks two and a half. Managing a supplier relationship in Chandni Chowk ticks none.

Honestly, the corollary is uncomfortable. Some of the most expensive credentials in India, the ones families optimised entire childhoods around, sit in the highest-verifiability zone. An entry-level services role whose value is "convert a written spec into working code, correctly, on time" is defined by the exact loop a model closes cheaply. Meanwhile a diagnostic electrician who works out why an old building trips its breaker only on humid evenings has none of the three conditions working against them.

Task	Auto-scoreable output?	Purely text/pixels?	Large public corpus?	Automation pressure
Writing a CRUD API endpoint	Yes (tests, compiler)	Yes	Yes	Very high
Summarising a 40-page report	Partly	Yes	Yes	Very high
Building a campaign forecast model	Partly (backtest)	Yes	Some	High
Deciding which of two strategies to fund	No	No	No	Low
Telling a client their brief is wrong	No	No	No	Very low
Diagnosing an intermittent electrical fault	No (world is the test)	No	No	Very low
Signing off on a regulated filing	No (liability attaches)	Partly	No	Very low

Read the last column as "how fast the price of this falls", not "does this disappear". Endpoints still get written. They're just no longer what you get paid for.

What actually protects a skill from automation

Apply one test to any task you do: if it goes wrong, who gets the phone call?

If the answer is "nobody, we just regenerate it", the task is already commoditised whether or not your employer has noticed yet. If the answer is a named human who loses money, standing or a licence, that task is protected. Not by difficulty. By the fact that a model can't be a defendant, can't be fired, and can't be trusted by a counterparty who wants someone to blame.

Five things pass that test. They're the spine of the rest of this piece.

Accountability. Being the person whose name is on the decision.
Judgement under ambiguity. Choosing when the data is incomplete and will stay incomplete.
Relationships and trust. Access, reputation, the benefit of the doubt.
Physical skill in an unstructured world. Hands, sites, machines, other people's premises.
Taste. Knowing which of ten adequate options is the right one, and why.

The original version of this argument is the one worth keeping. As execution gets cheap, the bottleneck moves to deciding what to make, why, and whether it's any good. Judgement, taste and communication become the scarce inputs. What follows is how to build each one on purpose rather than hoping it accumulates.

Accountability is the skill people are actually paid for

Most senior salaries aren't paid for output. They're paid for someone to hold the bag. A model can propose the budget reallocation. It can't be the person who tells the board why the quarter missed.

You build accountability the same way you build credit. Take small amounts of it early and never default. Concretely:

Put your name on forecasts, in writing, before the result. "I expect this to land between X and Y by the 20th." Then publish the comparison. Include the misses. Two quarters of that is worth more than any certificate.
Own the number, not the task. "I ran the campaign" is a task. "I own blended CAC for this line" is accountability. Ask to be measured on an outcome even when nobody offers.
Escalate early and specifically. The people who become trusted are the ones who say "this will miss by 15%, here are the two options" on day three. Not the ones who say nothing until day thirty.
Never launder a decision through the tool. "The model said so" is the fastest way to destroy the thing you're trying to build.

Judgement under ambiguity, and how to build it deliberately

Judgement isn't opinion. It's the ability to make a defensible call when the information is missing and will stay missing, and to say what would change your mind.

The reason models struggle here isn't intelligence. Ambiguity has no ground truth to train on. Two competent people can look at the same numbers and reasonably choose differently, because they're weighting risks the data doesn't contain. Cash runway. A supplier's temper. A regulator's mood. A co-founder's tolerance for a bad quarter.

Build it like this:

Keep a decision log. One line per real decision. What you chose, what you expected, what would prove you wrong, review date. Nothing improves judgement faster than being confronted with your own past reasoning.
Write the pre-mortem. Before committing, write the paragraph explaining why this failed, dated six months out. If you can't write it, you don't understand the decision.
Force the reversibility question. Is this a one-way door or a two-way door? Two-way doors deserve speed, not analysis. One-way doors deserve the opposite. Most people apply exactly the wrong effort to each.
Use AI as the opposing counsel, not the advocate. Ask it to make the strongest case against your plan. Then check whether its objections survive contact with facts you know and it doesn't.

Relationships and trust: why the network is the real moat

A model can draft the email. It can't be the reason someone replies.

Trust is the one asset that appreciates while everything around it deflates. When drafting costs nothing, inboxes fill with competent, well-formatted, entirely ignorable messages. The marginal value of being known goes up. Distribution beats production. And distribution, at the individual level, is relationships.

This is buildable and most people build it wrong. They network when they need something. The compounding version is narrow and boring. Pick fifteen to thirty people whose work you genuinely find interesting. Be useful to them at a low, consistent rate. An introduction. A document. A heads-up on something they were about to get wrong. Don't ask for anything for a year. The register you're creating is "this person is worth knowing", which is not a favour you can call in but an asset that pays out in unplanned ways.

Publishing is the leveraged version of the same thing. Writing in public about how you actually work compounds into inbound trust. People arrive already convinced, which is the most expensive step in any sale. The mechanics of that are covered in networking: the one lesson I wish I had learned earlier and how to grow your brand using digital channels.

Physical skill and the world that resists text

The safest work in the next decade is disproportionately work that happens in a place, with hands, in conditions that vary. Not because it's hard to reason about. Because it's expensive to instrument. Every physical task requires a robot, a supply chain, service infrastructure and someone to accept liability when it drops something on a customer's floor. Text costs a few paise per thousand tokens.

This matters for two groups. If you already do physical work, the protected part is diagnosis and site judgement, not the routine execution. Become the person who works out why, not just the person who fits the part. If you're a knowledge worker, the lesson isn't "retrain as a plumber". It's that any part of your job which involves being physically present, running a shoot, walking a warehouse, sitting across from a distributor, is worth more per hour than the part you do at a keyboard. Shift your time accordingly.

Taste is the only defence against infinite mediocre output

When anyone can produce forty variants in a minute, the scarce skill is picking. Taste is the compressed memory of a thousand comparisons. Knowing that the second headline is better and being able to say why in one sentence that generalises.

It's trainable, and the method is unglamorous. Consume a lot of work in your field with the intent to judge it. Write down the judgement. Then check it against reality. In advertising that means keeping a swipe file with a note attached to every entry. Not "nice ad" but "this works because the objection is handled in the first four words". Then run things and find out how often you were right. Taste that has never been marked against an outcome is just preference.

The AI-era version of taste has a second component. Knowing what not to make. Cheap production creates enormous pressure to publish more, and more of the same thing is how a brand becomes invisible. Someone has to say no forty times an hour.

Directing AI without becoming dependent on it

Briefing, reviewing and combining model outputs is now a core skill in itself. The failure mode isn't using it too much. It's using it in the zone where you can't check the answer.

The rule I'd hold to: use AI heavily where you can verify, cautiously where you can only sample, and not at all where you cannot tell right from wrong. In a domain where you have no expertise, the output isn't knowledge. It's a confident-sounding claim you are unable to audit. And you'll find out it was wrong at the worst possible time. Why ChatGPT gives wrong answers covers the failure mechanics; how to actually use ChatGPT and Claude for real work covers the working patterns.

Two habits keep dependency in check. First, keep one hard skill you still practise unassisted. Write one thing a week with no tool open. Or build one thing by hand. Second, never let the tool make the choice. Let it make the options. The moment you find yourself unable to explain why you shipped what you shipped, you've handed over the part that was yours.

A worked example: what accountability is worth in rupees

Abstractions don't move people. So here is the arithmetic on why "someone who checks" is a paid role.

Take an account running ₹25,000 a day in media spend with a target cost per acquisition of ₹600. An automated bidding change, or an AI-suggested reallocation nobody reviewed, pushes the actual CPA to ₹1,200. The spend continues either way. Only the efficiency drops.

Over three days, ₹75,000 is spent:

At the target CPA of ₹600, ₹75,000 should buy 125 conversions.
At the actual CPA of ₹1,200, it buys 62.5 conversions.
Value delivered at target rates: 62.5 × ₹600 = ₹37,500.
Wasted spend: ₹75,000 − ₹37,500 = ₹37,500.

Now suppose someone catches it on day one instead of day three. Only ₹25,000 has been spent at the bad rate. 20.83 conversions instead of 41.67. So the waste is ₹25,000 − (20.83 × ₹600) = ₹12,500. Catching it two days earlier saved ₹25,000 on a single account, in a single incident.

That's the whole argument in one calculation. The model made the change in a second and cost nothing. The value was created entirely by a human who looked, understood what "normal" looks like on this account, and had the standing to reverse it without a meeting. Run twelve accounts and one such incident a quarter, and the review function pays a salary by itself. If you want the sheet that surfaces this kind of drift, see your marketing either makes money or it doesn't and the definitions in marketing metrics explained.

What to stop investing in

Being honest about the other side of the ledger is more useful than another list of virtues.

Stop investing in	Why it is deflating	Invest instead in
Tool certifications as the main asset	Interfaces change; the certificate proves familiarity, not judgement	A public record of decisions and outcomes
Speed at routine production	Marginal cost of the first draft is heading to zero	Speed at reviewing and killing bad work
Memorised syntax and formulas	Recall is the cheapest thing a model does	Specification, trade-offs, system design
Being the person who "knows the process"	Processes get documented and then automated	Being the person who decides when the process is wrong
Volume of output as a personal brand	Everyone's volume just went up	Being right in public, on the record, repeatedly
Generic "learn to code because it's safe"	The safety premise is false	Coding as a way to think precisely, if you want it

None of these are worthless. They're just no longer scarce. Scarcity is what you're paid for.

How to build these in the next 90 days

A plan you could actually start on Monday, sized so it survives a busy week.

Weeks 1–2. Start the decision log. Write down every decision that cost more than an hour or ₹10,000, with your expected outcome and a review date. Ten entries is enough to begin.
Weeks 3–4. Pick one number in your work and formally ask to own it. Put a forecast in writing to whoever cares.
Weeks 5–8. Publish four pieces about how you actually do the work. Mechanisms, not opinions. This builds the writing skill and the trust asset at the same time.
Weeks 5–12, running in parallel. Contact two people a week from your fifteen-to-thirty list with something useful and no ask attached.
Weeks 9–12. Automate one recurring task of your own end to end. Then write down where the automation fails and what you had to check. That document is your verification skill, made explicit. How to automate your work with AI is the practical starting point.
Throughout. One deliberate unassisted rep per week. Keep the muscle.

FAQs

Should I still learn to code in 2026?

Yes, but for the right reason. Learn it because it teaches precise specification, debugging and systems thinking. Skills that transfer everywhere. Don't learn it as insurance against automation. Code is the most verifiable, most publicly documented text there is, which is exactly what models absorb fastest. Learn to direct and review code, not just to type it.

What jobs are safest from AI?

Roles where someone must carry liability, work physically in unstructured environments, or hold relationships that grant access. Regulated sign-offs. Skilled trades and diagnostics. Senior client ownership. Anything where being wrong has a named consequence. Safety comes from accountability and physical presence, not from technical difficulty. Difficult desk work automates faster than routine physical work.

Will AI make my skills useless?

It will make some tasks worthless and simultaneously raise the price of the skills around them. Drafting, summarising and routine production are deflating. Judging, deciding, verifying and being trusted are appreciating. The risk isn't owning no skills. It's owning only the deflating half and not noticing until your rate falls.

How do I build judgement when I am junior?

Volume of small reversible decisions plus a written record. Make calls, write down what you expected, review honestly. Ask to own one measurable outcome rather than a list of tasks. Sit in on decisions above your level and privately write down what you would have chosen. Then compare. Judgement is compressed feedback, so shorten the feedback loop.

Is prompt engineering a real skill worth learning?

The narrow trick version isn't. Models get better at interpreting sloppy input every release. The durable version is: writing a brief that is unambiguous, specifying what "good" looks like, and reviewing output critically. That's just clear thinking and clear communication applied to a new interface. It transfers to briefing humans too.

How do I stop AI from making me worse at my job?

Use it heavily where you can verify the answer, cautiously where you can only spot-check, never where you can't tell right from wrong. Keep one weekly unassisted rep in your core craft. And never let it make the final choice. Let it produce options and make the call yourself, so you can always explain why you shipped what you shipped.

Key takeaways

Automation follows verifiability, not difficulty, which is why "learn to code for safety" inverts the actual risk.
The durable test for any task is simple. If it goes wrong, does a named human get the phone call? If not, it's already commoditised.
Accountability, judgement under ambiguity, relationships, physical skill and taste are the five categories that resist automation. All five are deliberately buildable.
Directing AI is a genuine skill, but only in domains where you can verify the output. Outside those, you are accumulating unauditable claims, not knowledge.
Stop investing in certifications, memorised syntax and raw production speed. Start investing in a public record of decisions, outcome ownership and review capability.
A single caught error on a ₹25,000-a-day account saves ₹25,000 in two days. The review function pays for itself. That's the job.

The Problems Every Millennial and Gen Z Hits, and What Actually Fixed Them

Prabhash Jha — Wed, 23 Sep 2026 14:13:24 +0000

Every generation thinks its struggles are unique. Millennials and Gen Z actually share a surprising number of the same ones. Burnout. Comparison. Money stress. Not knowing what to do with our lives.

Here's what almost every article on this gets wrong. It takes problems that are structural, the price of a flat relative to a salary, income that arrives in lumps instead of monthly, a job ladder whose middle rungs were removed, and prescribes mindset fixes for them. Gratitude journals for housing costs. "Stop comparing yourself" for a market where the comparison is the product.

The reverse mistake is just as common and just as expensive. Treating genuinely behavioural problems, never deciding, never finishing, never letting anything compound, as if they were structural, and therefore not yours to fix.

So the useful move isn't more advice. It's a sort. Every problem on the list below belongs in one of two columns. Change the constraint. Or change the behaviour. Applying the wrong tool to the wrong column is why so much of this advice bounces off.

Why the usual advice fails: it treats structural problems as character flaws

A structural problem has a price, a market, or a rule behind it. You don't fix it by trying harder. You fix it by changing your exposure to it. Moving. Negotiating. Restructuring. Or deciding to accept the cost with your eyes open. A behavioural problem has a habit behind it. You fix it with a system, and no amount of relocating will help.

The sort matters because the two failure modes look identical from the outside. Someone who can't save is either paying 55% of take-home in rent (structural) or leaking money in a hundred small unexamined places (behavioural). The advice for those two people has almost nothing in common. Generic "budget better" content serves neither.

Problem	Mostly structural	Mostly behavioural	The lever that actually moves it
Rent eating half your income	Yes	No	Location, flatmates, negotiating on renewal, employer flexibility on remote
Income arriving unevenly	Yes	Partly	Buffer sized to the trough, retainers over one-off projects, staggered invoicing
Feeling behind your peers	Partly	Yes	Deciding what you are actually measuring, and against whom
Never finishing anything	No	Yes	Fewer commitments, a defined finish line, a review date
Career path disappeared	Yes	No	Skill stack with visible output, not a title hunt
Exhausted all the time	Partly	Yes	Sleep and load management before productivity systems
No idea how money works	No	Yes	One afternoon of reading, once

Only two rows in that table are fixed by attitude. The rest need a decision about where you live, what you sell, or how you get paid. That's a very different kind of work. Not a discipline problem. A design problem.

The reason this matters in practice: effort spent in the wrong column doesn't compound. You can journal for a year and your rent doesn't drop by a rupee. You can relocate to a cheaper city and still leak money if the habit isn't there. Sort first. Act second.

The problems we (quietly) all share

These are the five that come up most often. Worth stating plainly before we get to the fixes.

Comparison overload. Social media turns everyone's highlight reel into your baseline. A fast way to feel behind while you're actually doing fine.
Burnout from hustle culture. We were told to grind non-stop, then wondered why we felt exhausted and creatively empty.
Money stress with no money education. Rising costs, no financial literacy, and a lot of quiet guilt.
Analysis paralysis. Infinite options, infinite advice, and a fear of choosing the wrong path.
Impatience. We want the result now, and quit good things right before they'd have worked.

Two of those, comparison and impatience, are behavioural. A system genuinely fixes them. The money one is a mix. The education part is behavioural and cheap to fix. The cost part is structural and isn't. Burnout is usually load, not mindset. Let's take the structural ones properly, because that's where the honest arithmetic lives.

Housing: is renting really throwing money away?

This is the single most repeated piece of bad advice aimed at anyone under 40. It collapses the moment you run the numbers.

Take a ₹1.2 crore flat in an Indian metro. You put down 20%, ₹24,00,000, and borrow ₹96,00,000 over 20 years at 8.5%. The EMI works out to about ₹83,300 a month. Over 240 months that's roughly ₹2.00 crore in payments, plus the ₹24 lakh you already put in. About ₹2.24 crore of outlay against a ₹1.2 crore asset.

Now the part people skip. In year one, of the roughly ₹9,99,700 you pay, only about ₹1,91,000 goes to principal. The other ₹8,08,700 is interest. About ₹67,400 a month that leaves your account and never comes back. Exactly as "dead" as rent.

What would the same flat rent for? At a 3.5% gross rental yield, which is normal for Indian metros, ₹1.2 crore of property rents for about ₹35,000 a month.

So in year one you're comparing ₹35,000 of rent against ₹67,400 of interest, plus the running costs of ownership. Society charges. Property tax. Insurance. Repairs you can no longer ring a landlord about. Budget about 1% of the property's value a year for that, which is ₹10,000 a month here. Then add the opportunity cost of ₹24 lakh sitting in the down payment rather than earning anything (at 8%, about ₹16,000 a month). All-in, roughly ₹93,400 versus ₹35,000.

Line item	Buying (year 1, monthly)	Renting (monthly)
Interest / rent	₹67,400	₹35,000
Maintenance, property tax, insurance	₹10,000	₹0
Opportunity cost of down payment	₹16,000	₹0
Recurring cost of shelter	₹93,400	₹35,000
Principal repaid (yours, not a cost)	₹15,900	—

Put it as a rate instead of rupees and it gets clearer. Your blended cost of capital is about 8.4% (8.5% on the 80% you borrowed, ~8% opportunity cost on the 20% you put in). Add roughly 1% a year for maintenance, tax and insurance: about 9.4%. Subtract the 3.5% rental yield you're no longer paying. The property has to appreciate around 5.9% a year just to break even against renting. Before stamp duty and registration, which run 5-7% in most states and, spread over a seven-year hold, add roughly another 0.7-1% a year.

That isn't an argument against buying. Buying is genuinely worth it for things the spreadsheet can't price. You can't be asked to vacate. Your housing cost stops inflating. And an EMI is forced saving for people who wouldn't otherwise save. Those are real and they matter.

It is an argument against buying because renting feels like failure. That's the structural problem being misdiagnosed as a personal one. And it's the version that traps people in a 20-year commitment in a city they were going to leave in three.

The honest test I use: if the house wouldn't feel like the right decision on a spreadsheet alone, ask what the shame is doing there. Somebody sold you that shame. It wasn't your parents. It wasn't the market. It was a slow drip of "everyone your age has one by now", which is a comparison problem wearing a housing problem's clothes.

Income volatility: why an average salary and an uneven one are not the same job

The second structural problem is that a rising share of work, freelance, contract, commission, retainer, creator, pays unevenly. And the standard advice, "save three months of expenses", is sized for a salary. Wrong shape for a lumpy income.

Here's a year of variable income against fixed monthly costs of ₹70,000.

Month	Income	Surplus / deficit	Running cash
1	₹1,40,000	+₹70,000	₹70,000
2	₹60,000	−₹10,000	₹60,000
3	₹95,000	+₹25,000	₹85,000
4	₹1,80,000	+₹1,10,000	₹1,95,000
5	₹45,000	−₹25,000	₹1,70,000
6	₹30,000	−₹40,000	₹1,30,000
7	₹60,000	−₹10,000	₹1,20,000
8	₹1,25,000	+₹55,000	₹1,75,000
9	₹2,10,000	+₹1,40,000	₹3,15,000
10	₹80,000	+₹10,000	₹3,25,000
11	₹1,50,000	+₹80,000	₹4,05,000
12	₹1,45,000	+₹75,000	₹4,80,000

Total income: ₹13,20,000. Average: ₹1,10,000 a month against ₹70,000 of costs. On paper this person is comfortable. A ₹4,80,000 annual surplus, a 36% saving rate.

Now look at months 4 to 7. Cash peaks at ₹1,95,000 and falls to ₹1,20,000. A ₹75,000 drawdown over three consecutive months, in a year that ended well. If they'd started the year at zero, or spent the month-4 windfall the week it landed, months 5 to 7 are a credit card and a bad decision.

Two practical consequences follow, and neither is a mindset fix.

Size the buffer to the trough, not the average. For volatile income, the number that matters is the worst peak-to-trough gap you've actually experienced, plus a normal emergency fund on top. Six months of costs is a reasonable floor here rather than three, and the emergency fund guide covers where that money should actually sit.

Your tax bill doesn't care when clients pay. The Income Tax Department's advance tax schedule falls due on 15 June, 15 September, 15 December and 15 March, at cumulative thresholds of 15%, 45%, 75% and 100% of estimated liability. Miss them and interest accrues under sections 234B and 234C. If your gross receipts cross ₹20 lakh in services in most states, GST registration and periodic filing become obligations too. If you're a specified professional, section 44ADA presumptive taxation may apply up to ₹50 lakh of receipts (₹75 lakh where cash receipts stay within 5%). Worth checking with a chartered accountant, because it changes both your paperwork and your cash timing.

This is the same distinction that kills small businesses. A profitable year and a solvent year are different things. Cash flow versus profit is the version of this problem at company scale, and it behaves identically at personal scale.

Comparison: what social media actually does to your decisions

Comparison is the one problem on the list that's genuinely behavioural. It's also the one where the mechanism is usually described too vaguely to act on.

The damage isn't that you feel bad. It's that the feed quietly supplies your reference class. You don't compare yourself to a random sample of people your age. You compare yourself to a sample selected for having something worth posting. That sample then sets your sense of what's normal. Which house, which salary, which milestone, by which age.

Then it gets worse, because the same feed sells you the fix. A large amount of Indian money content comes from people giving specific buy-and-sell recommendations without being registered. Under SEBI rules, giving investment advice for consideration requires registration as an investment adviser, and unregistered recommendations aren't a grey area. Education about how instruments work is legal and useful. "Buy this stock on Monday" from an unregistered account is neither. Same for referral income: earning commission on insurance requires the appropriate IRDAI registration, so an affiliate link under a "best term plan" video is a regulatory question, not just an ethics one.

The practical fix is unglamorous. Change the input, then change the scoreboard. Unfollow accounts whose main product is aspiration. Write down the three numbers you're actually trying to move this year, a savings rate, a skill, a client count, and check those instead. A scoreboard you chose beats a scoreboard the algorithm chose.

The career ladder that no longer exists, and what replaced it

The old path was legible. Join, do well, get promoted every two or three years, retire senior. Large parts of that ladder have been removed, particularly in the middle, where a lot of coordination work used to sit.

The common answer, "build a personal brand", is mostly wrong. Or at least badly ordered. Attention without a demonstrable skill decays fast and converts badly. What has actually replaced the ladder is a skill stack with visible output. Two or three capabilities that are individually ordinary and jointly rare, plus artefacts a stranger can inspect.

The practical test: can someone who has never met you verify your claim in under five minutes? A repository. A live campaign. A dashboard you built. A post that explains a real problem. A spreadsheet that does something. Not a title. Titles are portable only within the company that granted them.

The AI question sits inside this and is worth being unsentimental about. The tasks most exposed are the ones with clear inputs, clear outputs and no accountability attached. Exactly the middle-rung work that was already thinning. The response isn't to avoid the tools. It's to move up the accountability curve and use them. Automating your own work is the cheapest version of this, and a practical way to think about AI and your job is the longer argument.

One more thing that's structural and rarely named. Most good work still arrives through people who already know what you can do. That's not a personality trait. It's a distribution channel, and it can be built deliberately. The networking lesson is about exactly that.

Delayed milestones: which ones can wait and which ones cannot

Marriage later. First home later. Children later. Financial independence later. The standard framing is that this is a failure of the generation. It's more accurately a repricing. The assets that constituted the old milestones cost more relative to income than they did, so they arrive later.

What's genuinely useful here is sorting milestones by whether delay is cheap or expensive.

Delay is cheap for anything whose price you can re-enter later at roughly the same terms. A house. A car. A title. A wedding of a particular size. These have no deadline attached other than a social one.

Delay is expensive for anything driven by compounding or biology. Money invested at 25 has forty years to work. The same money at 35 has thirty. That gap isn't recoverable by working harder later. It's arithmetic. Fertility has its own timeline and its own costs, and that's a medical conversation rather than a financial one, but it belongs in the same honest sort.

So the practical sequencing: start the things that compound as early as you can afford to, at whatever amount is genuinely sustainable, and let the purchase-type milestones happen when the price and your income line up. The failure mode is doing it backwards. Buying the milestone that could have waited, on borrowed money, and postponing the one that couldn't.

Burnout: the difference between being tired and being depleted

Tired is recoverable with a weekend. Depleted isn't. The tell is that rest stops working. You take the break, and come back the same.

Burnout is usually a load problem before it's a mindset problem. Too many open commitments. Too little control over which ones. No defined finish line on any of them. Sleep, workload and autonomy do more than any productivity system will, and no morning routine survives a genuinely unmanageable job.

Where behaviour does come in is what happens after you notice. The instinct is to add. A new system, a new app, a new discipline. The move that works is subtraction. Cut the number of things in flight until the remaining ones can actually be finished, then reintroduce slowly. Fewer bets, finished. Not more bets, half-done.

What actually fixed them (for me)

Play your own game. Comparison dies the moment you define success on your own terms. Measure against last month's you, not a stranger's feed.
Trade hustle for systems. Sustainable beats intense. Small, repeatable habits outperform heroic bursts every time. In fitness, money and work.
Learn the boring basics once. An afternoon learning how money actually works removes years of quiet stress. Same for almost any skill.
Decide, then adjust. A decent decision you act on beats a perfect one you delay. You can course-correct. You can't steer a parked car.
Let things compound. Most good outcomes, savings, skills, audiences, relationships, are boring for a long time and then suddenly not. Patience is the cheat code.

Each of those is a behavioural lever. That's precisely why they work. They're aimed at the column where behaviour is the binding constraint. Point them at a rent-to-income ratio and they won't do anything, which is worth knowing before you blame yourself for the result.

What genuinely helps versus what is just advice

What is usually said	What it is worth	What to do instead
"Stop comparing yourself"	Low — it is a feeling, not an action	Change the inputs, then pick three numbers you actually track
"Rent is throwing money away"	Negative — it is arithmetically wrong in year one	Run interest plus opportunity cost against rent before deciding
"Save three months of expenses"	Partial — sized for a salary	For lumpy income, size the buffer to your worst peak-to-trough gap
"Build a personal brand"	Low on its own	Build a skill stack with output a stranger can verify in five minutes
"Follow your passion"	Low — passion follows competence more often than it leads	Get good at something with demand, then choose within it
"Just work harder"	Negative when the issue is load	Cut commitments until the remaining ones can be finished
"Learn how money works"	High, and it is a one-time cost	Spend one afternoon on it, once, and stop paying the stress tax

The one lesson under all of them

Nearly every fix comes back to the same idea. Stop optimising for how things look, and start optimising for how they compound. Comparison, hustle, impatience. They're all short-term games. The people who seem to have it figured out mostly just played the long one, quietly.

The structural version of that same sentence. The costs that compound against you, interest, an unaffordable location, a commitment you can't exit, deserve the same seriousness as the returns that compound for you. Most of the difference between people at 40 isn't intensity. It's which of those two lists got longer.

FAQs

Why do millennials and Gen Z feel so behind?

Mostly because we compare our real lives to everyone else's edited highlights, at a scale no generation faced before. The feed sets your reference class, and it's selected for people with something worth posting. Muting the comparison and defining your own scoreboard fixes more than any productivity hack. But check first whether the problem is comparison or an actual cost structure.

Is it better to rent or buy a house in India right now?

Depends on the gap between the rental yield and your loan rate. At a 3.5% yield and an 8.5% loan, the property needs roughly 6% annual appreciation just to match renting, before stamp duty. Buy for stability and forced saving if you'll stay seven years or more. Don't buy because renting feels like failure.

How much emergency fund do I need if my income is irregular?

More than the standard three months. Look at your last twelve months, work out the largest peak-to-trough fall in your cash balance, and hold that on top of a normal buffer. Six months of fixed costs is a sensible floor for freelance or commission income, kept somewhere liquid rather than invested.

How do I stop burning out?

Swap intensity for consistency. Build small systems you can repeat on a bad day, protect your rest, and stop treating exhaustion as a badge. But check the load first. If you have more open commitments than you can finish, no routine will fix that. Cut the list before you optimise the schedule.

Is a career ladder still worth chasing, or should I freelance?

Neither is automatically better. Employment buys predictable cash flow and someone else's client acquisition. Freelancing buys optionality and pays unevenly. What survives both is a skill stack with output a stranger can verify. Choose the structure that matches your tolerance for a ₹75,000 cash drawdown in a bad quarter.

Should I follow finance influencers for money advice?

Use them for education, not recommendations. Under SEBI rules, giving investment advice for consideration requires registration, and specific buy-or-sell calls from unregistered accounts aren't permitted. Earning insurance commission needs IRDAI registration too. Learn how instruments work from anyone. Take actual recommendations only from someone registered and accountable.

Key takeaways

Most advice on this topic fails because it applies mindset fixes to structural problems and structural excuses to behavioural ones. Sort each problem before you spend effort on it.
Renting isn't throwing money away. In year one of a ₹96 lakh loan at 8.5%, roughly ₹67,400 a month is interest against ₹35,000 of equivalent rent.
Buying makes sense for stability, rent-inflation immunity and forced saving over a long hold. Not because renting feels like failure.
Irregular income needs a buffer sized to your worst peak-to-trough drawdown, not to your average month, because tax deadlines and rent don't move when clients pay late.
The replacement for the vanished career ladder is a skill stack with output a stranger can verify in five minutes, not a personal brand.
Delay the milestones you can re-enter at the same price. Start early on the ones driven by compounding, because that gap is arithmetic and can't be recovered later.

What Actually Breaks When You Put a Language Model in a Customer-Facing Flow

Prabhash Jha — Tue, 22 Sep 2026 14:01:23 +0000

Search for what goes wrong with a language model in production and you get a remarkably consistent answer: hallucination, drift, latency, cost, observability gaps. The list is correct. It is also written almost entirely by companies selling observability platforms, which is why every version of it ends at the same place — a dashboard.

The dashboard is not the problem. The problem is that the list is organised by what the model does, and the things that actually hurt you are organised by what the customer does next. Those are different taxonomies. A model that hallucinates into a log file is a quality issue you fix on Thursday. The same hallucination said to a customer, in your brand voice, on a page with your logo on it, is a commitment somebody may hold you to.

This is the version of the list I would want before shipping: the failure modes in the order they actually turn up, what each one looks like from the outside, and the specific thing that catches it. It assumes you have already decided the automation is worth building — should you automate this is the arithmetic for that decision, and it is a different question from this one.

What breaks first when a language model starts talking to customers?

The first thing that breaks is not the model's general knowledge — it is its account of your own policy, stated confidently, to somebody who then acts on it.

This is the failure that surprises people, because it is the opposite of what the testing predicted. In testing you ask the model hard general questions and it does well. In production, customers ask it narrow, specific, high-stakes questions about your terms: can I get a refund after 30 days, does this plan include that, will you waive the fee, can I apply for the discount retroactively. Those answers are not in the model's training data in any reliable form. They are in your policy document, your CRM, and three exceptions your support lead carries in her head. The model does not know it does not know. It produces the most plausible-sounding policy, and a plausible-sounding refund policy is a very close relative of a real one.

The reason this matters commercially rather than just technically is that a statement made by your chatbot on your website can be treated as a statement made by you. That is not a hypothetical. In Moffatt v. Air Canada, 2024 BCCRT 149, decided by British Columbia's Civil Resolution Tribunal in February 2024, a customer asked the airline's website chatbot about bereavement fares and was told he could apply retroactively. He could not — the airline's actual policy said otherwise. Air Canada's position was that it should not be held liable for the information the chatbot provided. The tribunal disagreed, found the airline had not taken reasonable care to ensure its chatbot was accurate, held that this amounted to negligent misrepresentation, and awarded damages of $650.88.

The money is trivial and the money is not the point. The point is the principle the case settles: the chatbot is not a third party you can stand behind. It is a channel, and everything on it is yours. Whatever you would not let a new support hire say unsupervised in week one, the model should not be saying either.

The practical consequence is a scoping rule I would apply before writing a line of the prompt: separate what the model is allowed to explain from what it is allowed to state. Explaining is safe — here is roughly how our pricing works, here is what this feature does, here is where to find that setting. Stating is not — your refund will be processed, you qualify for that rate, we can do that for you. Anything in the second category should come out of a system that knows the answer, with the model doing the wording rather than the deciding.

Why does it get worse after launch instead of better?

It gets worse because the questions change. The model is the same on day 60 as on day 1; the input distribution is not.

Every pre-launch test set is written by people who already know how the product works. Real users do not, and the gap shows up in three specific ways within the first two months:

The questions get shorter and worse-formed. Internal testers write full sentences. Customers write "refund???" and "not working". A model that was reliable on well-formed questions has to guess at intent, and guessing at intent is where it starts inventing.
The long tail arrives. The 20 scenarios you tested cover most of the volume and almost none of the risk. The edge cases — the customer who is angry, the one asking about a plan you discontinued, the one describing a situation your policy genuinely does not cover — arrive at a low rate and a high cost.
Your own product changes and the prompt does not. You change a price, retire a plan, alter a delivery window. The people who make that change do not think of the prompt as a place where the old value is written down, because it does not look like a config file. So the model keeps confidently quoting a price you no longer charge.

That third one is the same class of defect as the automation failure nobody catches — the workflow that runs green and does nothing. Nothing errors. Every dashboard is green. The system is doing its job perfectly against an instruction that stopped being true in March.

The fix is boring and it works: treat the prompt as production configuration, not as copy. It goes in version control, it has an owner, and any change to pricing, policy or plan structure has "update the assistant's context" on the same checklist as "update the pricing page". If your release checklist does not name it, it will not happen.

What happens when the provider updates the model?

Your system's behaviour changes underneath you, at a time you did not choose, and nothing in your monitoring says so.

This is the failure mode that is most specific to building on someone else's model and least like anything in normal software. You did not deploy. Your code did not change. But the thing your product's tone, refusal behaviour and formatting all depend on has been retrained, and the prompt you tuned against the old behaviour is now slightly mistuned against the new one. Usually it is small. Occasionally it is not — a model that used to refuse a category of request now answers it, or one that used to answer now hedges, or the output format your parser depends on gains a preamble.

Three things make this survivable:

Pin the version. Use a dated or pinned model identifier rather than the floating alias that always points at the latest. The floating alias is convenient exactly until the day it is not. Pinning turns an unannounced change into a change you schedule.

Keep a golden set. Thirty to fifty real questions with the answer you consider correct, run on demand, diffed by hand. Not an evaluation framework, not a score out of ten — a list you can re-run in ten minutes and read with your own eyes. Most teams skip this because it feels unscientific. It catches more than the scored version does, because you notice the thing you were not measuring.

Read the deprecation notices. Providers publish them. Somebody has to be subscribed, and if nobody is named, nobody is.

Why did the bill triple in a month?

Because the cost of a language model feature scales with conversation length and retry behaviour, not with the number of customers — and both of those grow quietly.

The budget you build before launch is almost always per-request: tokens in, tokens out, multiplied by expected volume. The bill you get is per-conversation, and a conversation resends the entire history on every turn. A support thread that goes eight turns instead of three does not cost you 2.6× more, it costs considerably more than that, because each turn carries everything before it. Add a retrieval step that pastes in documents, and the context grows faster still.

Then there is the failure case nobody budgets: a request errors, something retries it, and the retry also errors. Retry loops are cheap when the unit cost is a database read and expensive when the unit cost is a model call. The first surprising bill I would expect any team to get is not from traffic. It is from a loop.

What actually controls this:

Control	What it prevents	Where it goes
Hard cap on turns per conversation	Unbounded thread growth	Application layer, before the call
Retry limit with backoff, and a dead end	Retry-loop spend	The client wrapper, not the retry library's default
A per-day spend ceiling with an alert below it	Discovering the problem on the invoice	Provider console, plus your own counter
Truncating or summarising old turns	Context growth on long threads	Conversation state, before it is sent
A cheaper model for classification steps	Paying premium rates to decide "is this a billing question"	Routing, not the main flow

The last row is the one with the most money in it and the least attention. A lot of production spend is a frontier model doing work that a smaller model does adequately — routing, classification, extraction. The model that talks to the customer should be your best one. The model that decides which queue the customer belongs in does not need to be.

What stops a customer talking your assistant into something?

The thing that stops it is not a better prompt. It is not giving the model the ability to do the damaging thing in the first place.

Two related risks sit here and they are worth naming separately. The first is prompt injection: a user, or a document the model reads, contains instructions that the model follows as though they came from you. The OWASP Top 10 for LLM Applications keeps prompt injection at the top of its list, and the reason it stays there is structural rather than a matter of engineering effort — instructions and data arrive through the same channel, the context window, and there is no equivalent of the parameterised query that cleanly separates them. You cannot prompt your way out of an architectural property. You can only limit what a successful injection is able to reach.

The second is excessive agency: the model has more tools, permissions or autonomy than the job requires, so an influenced model can take a harmful action rather than merely say a wrong thing. OWASP's mitigations are the ordinary ones from access control — least privilege on every tool, human approval for consequential actions, spending and rate limits — which is the correct and slightly deflating answer. This is not an AI problem with an AI solution. It is a permissions problem, and I have written the longer version of it in deciding what an AI agent is allowed to touch.

The test I would apply to any tool before connecting it: assume the model has been successfully persuaded to call this with the worst plausible arguments. What is the damage? If the answer is "reads a public FAQ", connect it. If the answer is "issues a refund", "changes an address", "sends an email as us", or "cancels the order", it does not get called directly — it gets proposed, and something else confirms.

Why the customer-facing part is a brand decision, not a support decision

Because the failure is not evenly distributed across your customers. It is concentrated on the ones already having a bad day.

People do not open a support chat when things are going well. The population your assistant talks to is skewed towards confusion, frustration and urgency, which means a wrong answer lands on the exact customer least able to absorb it, at the exact moment they are deciding what they think of you. That asymmetry is why I treat this as a brand question. Your brand is whatever your worst touchpoint is, and an assistant that is excellent 95% of the time and confidently wrong to an angry customer 5% of the time is not a 95% system in the customer's memory.

Two design consequences follow, and both cost conversion rate on paper:

The escalation path has to be visible from the first message, not offered after three failures. Hiding the route to a human raises containment and lowers the thing containment is a proxy for. A customer who wanted a person and was made to negotiate with a bot for four turns has already had the experience you were trying to prevent.

The assistant should be able to say it does not know. Explicitly, in the prompt, with examples. Models default to answering because answering is what the training rewards, and "I am not certain — let me get someone who is" is a sentence you have to ask for. It is also, in a support context, a genuinely good answer, and customers respond to it far better than product teams expect.

The order I would instrument it in

Before launch, in this order, because each one catches something the previous one cannot:

Log the full exchange, with an identifier. Input, output, model version, prompt version, tools called, timestamp. Without this you cannot investigate anything — you are relying on a customer screenshot. The general principle and the specific fields are in what to log when an AI agent acts on your behalf.
A one-click way for a support human to flag an answer as wrong. The people who spot the failures are the ones cleaning up after them. If flagging is a form, it will not get used.
A daily read of a sample. Twenty conversations, read by a person, every day for the first month. Nothing automated finds what this finds, and it is the single highest-value habit in the list. It ends when you stop being surprised.
Alerting on the shape of the traffic, not just errors. Escalation rate, average turns per conversation, spend per day, refusal rate. All four move before anything errors.
A kill switch that a non-engineer can reach. One toggle, documented, that routes everything to the existing human path. If turning it off requires a deploy, it will stay on through the incident.

Item five is worth being blunt about. The reason to build it is that at some point you will want it at an inconvenient hour, and the decision to use it should not depend on who is awake.

What to do in the first hour after it says something wrong

Assume the answer it gave is now a commitment, and work backwards from that.

The sequence I would follow, in order:

Honour it if it is small. If the model promised a discount, a waiver or a delivery date that you could reasonably absorb, absorb it. Arguing with a customer about whether your own website meant what it said is a losing position commercially and, per the Air Canada decision, not a strong one otherwise. The cost of honouring one wrong answer is almost always less than the cost of the argument.

Find out how many others got the same answer. This is what the logging is for. One customer with a wrong answer is an incident; four hundred with the same wrong answer is a policy change you did not make. Search the logs for the topic, not for the exact wording — the model will have said it differently every time.

Turn off the specific capability, not the whole system, if you can. Narrow the blast radius before you debug.

Fix the source, not the symptom. The instinct is to add a line to the prompt saying "never say X". That works for the exact phrasing and fails for the next one. If the model got the policy wrong, the question is why the policy is not being supplied to it as a retrievable fact rather than a remembered one.

Write down what class of question caused it, and add it to the golden set. The failure you already had is the cheapest test case you will ever get.

The summary table

Failure mode	How it shows up	What catches it
Confident wrong policy answer	Customer holds you to it	Retrieval for facts; model words, system decides
Prompt goes stale after a product change	Model quotes an old price	Prompt in version control, on the release checklist
Provider updates the model	Tone, format or refusals shift with no deploy	Pinned version + a golden set you re-read
Cost growth	The invoice, a month late	Turn caps, retry limits, spend ceiling, cheaper model for routing
Prompt injection	Model follows instructions that were not yours	Least privilege on tools; assume the prompt can be beaten
Excessive agency	It does the harmful thing, not just says it	Propose-and-confirm for anything consequential
Containment optimised over resolution	Falling escalation rate, rising complaints	Visible human path; read a daily sample

FAQs

Is a retrieval setup enough to stop the model getting our policy wrong?
It removes the largest cause and not the whole problem. Retrieval fixes the case where the model was remembering instead of reading. It does not fix a policy document that is ambiguous, out of date, or contradicted by an exception your team applies in practice but never wrote down. Retrieval makes the model as accurate as your documentation, which is usually a large improvement and rarely a complete one.

Should the assistant be able to complete transactions or only answer questions?
Answer questions first, and add actions one at a time, each with a confirmation step, in increasing order of blast radius. The reason is not that the model is incapable — it is that an action taken wrongly is discovered later and costs more than a sentence said wrongly. If you want a general rule for which actions qualify at all, the classes in deciding what an AI agent is allowed to touch apply unchanged.

How do we measure whether it is working?
Not by containment. Containment rises when you hide the human option, which is the failure you are trying to avoid. Better measures are resolution without a follow-up contact within a week, escalation rate read alongside satisfaction rather than instead of it, and the number of answers your support team flags as wrong per hundred conversations.

Does a smaller or cheaper model make the failures worse?
For the failures in this post, mostly not — they are architectural rather than capability-limited. A stronger model hallucinates a policy slightly less often and is not meaningfully harder to talk into something, because prompt injection is a property of how instructions and data share a channel. Model choice is a cost and latency decision. Scope, retrieval and permissions are the safety decisions.

Why do these systems seem fine in testing and degrade in production?
Because your test set was written by people who know how the product works, and it stops representing the real traffic within weeks. The questions get shorter, the long tail arrives, and your own product changes without the prompt changing with it. The related question of why the model sounds equally confident whether it is right or wrong is covered in why ChatGPT gives wrong answers.

The honest position

A language model in a customer-facing flow is a good idea more often than the sceptics say and a much larger commitment than the vendors say. The technology is not the risky part — the risky part is that you have connected a fluent, confident, non-deterministic system to the channel where your customers form their opinion of you, and then measured it on containment.

Most of what makes it work is unglamorous and predates all of this: scope it to what it can be right about, give it the facts rather than trusting it to remember them, restrict what it can do as opposed to what it can say, read the transcripts, and keep a way to turn it off. The parts that fail are rarely exotic. They are the same parts that fail in any automation, which is why the discipline that keeps the automation you should delete from quietly rotting applies here without modification.

If you are deciding whether to ship one, the useful question is not "is the model good enough". It is: what is the worst sentence this can produce, and are we willing to honour it? If the answer is no, the scope is wrong, and no amount of prompt engineering fixes a scope problem.

The Handover Document That Decides Whether You Can Take a Week Off

Prabhash Jha — Tue, 22 Sep 2026 14:01:22 +0000

Search for how to hand over before a holiday and you will find a large amount of sensible, generic advice: brief the team, write an overview of ongoing projects, set an out-of-office, start two weeks early, name a point person. All of it is true. None of it explains why founders who do all five still spend the week answering messages from a hotel.

The reason is that the standard advice is written for an employee going on leave, and an employee's handover problem is coverage — someone has to do the tasks. A founder's handover problem is different and mostly invisible in the advice: it is authority. Your team usually knows perfectly well what to do. What they do not know is whether they are allowed to do it without you, and in the absence of that knowledge the safe move is to wait, or to ask. So the messages arrive, and the week is not a week off, it is the same week with worse wifi.

This is the document that fixes that, in the order the items actually matter — which is the order in which their absence breaks the week, not the order they appear on a checklist.

Why do handovers fail even when everything is documented?

They fail because the document answers "what happens" and the team's real question is "what am I allowed to decide".

The distinction is easy to miss because both look like the same sentence. "Client X's report goes out on Wednesday" is coverage. "If client X asks for a change to the report, Priya decides, up to two hours of extra work, without checking" is authority. The first one is satisfied by a project list. The second one is the entire difference between a real week off and a remote week.

Oncken and Wass gave the underlying dynamic its lasting name in Management Time: Who's Got the Monkey?, which opens on the paradox that managers run out of time while their subordinates run out of work. Their metaphor is that responsibility — the "monkey" — jumps from the subordinate's back to the manager's during ordinary conversations, because the manager takes ownership of solving the problem rather than leaving it where it belongs. A founder's holiday is the highest-intensity version of this. Every unresolved question in the business is looking for a back to sit on, and yours is the default one. It stays the default one unless something has been written down that assigns it elsewhere in advance.

Which is why the useful mental model for the document is not "handover notes". It is a decision-rights document, with the project list attached at the end because it is the least important part.

The three questions that produce the whole document

Before writing anything, answer these three, in this order, because the document is just their answers written out:

What decisions arise in a normal week? Not emergencies — normal ones. Approving an invoice, agreeing a scope change, replying to a complaint, signing off creative, letting someone start early.
Who decides each one while I am away, and up to what limit? A named person, and a number or boundary. "Use judgement" is not an answer; it is the thing that generates the message to you.
What is genuinely worth interrupting me for? A short list, written in advance, that both of you have agreed. If you do not write it, the list becomes "anything the person is nervous about", which is unbounded.

That third one is where most founders quietly sabotage themselves. Saying "call me if anything comes up" feels generous and is the opposite — it transfers the judgement of what counts back to a person who has every incentive to be cautious. Two or three specific triggers, written down, is both kinder and more effective.

What goes in the document, in order of what breaks first

1. The decision-rights table

This is the whole document if you only write one thing, because it is the only part that prevents the messages rather than answering them.

One row per recurring decision. A named owner. A limit. It looks like this:

Decision	Who decides	Limit / boundary	Escalate if
Scope change on a live project	Delivery lead	Up to ~half a day of extra work	It changes the deadline or the fee
Invoice approval	Ops	Up to your normal supplier ceiling	New supplier, or above the ceiling
Client complaint	Account lead	Any response, any goodwill up to a set value	The client uses the word "contract" or "notice"
Discount on a live quote	Nobody	—	Always waits
Hiring: offer, start date, salary	Nobody	—	Always waits
Press, public statement, anything on the record	Nobody	—	Always waits

Two things about that table matter more than its contents.

The "nobody decides" rows are as important as the rest. A handover that grants no authority produces a paralysed week; one that grants unlimited authority produces a week you spend undoing things. Naming the small number of decisions that simply wait until you are back is what makes granting the others safe, and it is a much easier conversation to have in advance than in the moment. Most of those rows will be the same functions in the five things a founder should never fully delegate — pricing, the first-time complaint, the hire — which is not a coincidence. The permanent version of that list and the one-week version have the same logic behind them.

Every limit is a number or a boundary, never an adjective. "Small changes" is not delegable. "Up to half a day" is. The number does not have to be right. It has to exist, because a wrong number gets corrected once and an absent number generates a question every time.

2. The escalation path, with a named human at each step

Write down who the covering person calls when the thing is above their limit and you are unavailable — and make sure it is a person, not you.

This is the item that most founder handovers skip entirely, on the assumption that the answer is obviously "the founder". If that is the answer, you have not handed over; you have simply added latency. Google's SRE practice makes the same point about people carrying a pager: the chapter on being on-call lists the most important on-call resources as clear escalation paths, well-defined incident-management procedures, and a blameless postmortem culture. Note what is not on that list — deep expertise, or the presence of the person who normally does the job. What makes cover work is knowing where the next decision comes from.

For a small business the path is usually short and that is fine: covering person → second named person → you, with the second name being someone who can make a call rather than someone who merely knows things. It has to be written, because the point is to remove the moment where someone has to decide whether their problem is big enough to justify contacting you. That moment is where holidays are lost.

3. Access, before you go, tested

Everything the covering person will need to open, opened by them, while you are still in the country.

Access failures are the most common way a handover collapses, and they are almost entirely preventable, because they are discoverable in advance. The pattern is familiar: the ad account is under your personal login, the domain renews to your card, the client portal has one seat, the payment gateway requires an OTP to your phone, the bank needs your token. Nothing here is a documentation problem. It is an ownership problem that becomes visible only when you are unreachable — which is exactly the argument in who owns the ad account, and the reason to run that audit is not really the holiday, it is what happens if you are ever unreachable without planning it.

The test is the whole point, and it is one line: have the covering person actually log in to each system while you are still available, rather than confirming they have the credentials. Those are different claims and only one of them is checkable.

Give particular attention to anything that sends a code to your phone. Two-factor tied to a founder's personal device is the single most common blocker, it fails silently until the exact moment it matters, and the fix — shared authenticator, a second registered device, or a delegated account — takes an afternoon.

4. The money view

One paragraph: what is due in, what goes out, what happens if either does not.

A week is short enough that most cash questions can wait and long enough that a few cannot. What the covering person needs is small: which receipts are expected and from whom, which payments go out automatically, what the balance floor is, and what to do if a large expected receipt does not land. That last one is the only genuinely tricky case, and it deserves a written instruction because the instinct — chase immediately and firmly — is not always right early in a relationship. The sequence for a client who has stopped paying is a good default to point at, and pointing at an existing document beats inventing a policy in an airport.

If you do not know your own floor, that is a bigger finding than the holiday: how much runway a services business actually needs is the calculation, and it is worth doing before you go, not because the week is risky but because the answer changes how much authority you are comfortable granting.

5. The client-facing sentence

Decide what your clients are told, and write the exact wording rather than the policy.

Two decisions here, and both are yours rather than the team's. First, whether clients are told at all — for a founder-led business, "he is away this week" is heard by some clients as "the person I actually buy from is not available", which is a service message whether you intend it or not. Second, if you say nothing and a client discovers you are away by getting a slow reply, that is worse than having said it. The workable middle is usually: tell the accounts that have a live deliverable this week, say nothing to the rest, and give the covering person the exact sentence to use.

Write the sentence. Not a guideline about the sentence. This is a brand touchpoint like any other — your brand is whatever your worst touchpoint is — and the version an anxious team member improvises under time pressure is rarely the version you would have chosen.

6. The project list

Last, because it is the part everyone writes first and the part that causes the fewest problems.

Your team already knows what the work is. Keep this short: what is due, to whom, and by when, with anything unusual flagged. The only rows worth writing carefully are the ones where something is not yet agreed, because those are the ones that will need a decision, and a decision needs a row in table one.

The rehearsal that actually tells you whether it works

Take a Wednesday off, unreachable, before you take the week.

This is the highest-value item in the entire post and it costs one day. A full day of genuine unavailability — phone off, no email, agreed in advance — surfaces every gap in the document in a way that no amount of reviewing it does. You will come back to a list of the things that were waiting for you, and that list is the missing part of your handover. Nothing else produces it. A document reviewed at a desk always looks complete, because you read the gaps and fill them in from memory without noticing.

Do it two or three weeks before the real trip so there is time to fix what it finds. And be strict about the unavailability, because a rehearsal you break at 3pm produces no information at all.

Reading what happened when you get back

The week produces a data set about your business that nothing else produces. Read it before it evaporates.

Ask the covering person for three things — not "how did it go", which gets you "fine":

What did you wait on that you would rather have decided? These become new rows in the decision table, with limits. This is the list that makes next time shorter.
What did you decide that you were not sure you were allowed to? These are the rows where your document was unclear rather than absent, which is a different fix.
What broke that nobody had thought of? Occasionally this is a genuinely new failure mode and it belongs in the document permanently.

Then read the list of things that got escalated to you against the trigger list you agreed. Anything that reached you and was not on that list is either a gap in the document or a signal that the covering person did not believe the authority was real. Those two look identical from a distance and have opposite fixes, so ask which it was.

The uncomfortable version of this finding is worth naming. If a week away reveals that nothing can move without you, that is not a documentation failure and no document will fix it. It is a structural fact about how the business is built, and it has a cost you are already paying — in the ceiling on what you can take on, in what the business is worth to a buyer, and in what happens if you are unavailable for a reason you did not choose. That is a longer piece of work than a handover, and it starts with knowing which client is actually profitable when everyone works on everything, because concentration of work in one person usually tracks concentration of margin in a few accounts.

What not to do

Do not write it the night before. The value is in the decisions, and decisions made at 11pm while packing default to "ask me".

Do not make the document long. A decision table, an escalation path, an access list, a money paragraph, a client sentence, a project list. Two pages. Anything longer will not be read at the moment it is needed, which is the only moment that counts.

Do not check in "just briefly" on day two. One check-in re-establishes you as the decision-maker for the rest of the week, and everyone reverts to waiting. If you genuinely cannot be unreachable, schedule one fixed window — a specific hour on a specific day — and be unavailable outside it. A predictable window is a workable compromise. Ambient availability is not.

Do not skip the debrief. The week is the most honest audit of your operational dependency you will ever run, and its findings have a shelf life of about three days.

Do not treat the document as single-use. Nearly all of it is true every week, not just holiday weeks. The decision table in particular is a permanent artefact that happens to have been written for a trip.

FAQs

How far in advance should I start?
Two to three weeks, and the reason is the rehearsal rather than the writing. The document itself takes an afternoon. Finding out what it missed takes a test day plus time to fix what the test day surfaces.

What if I have no one to hand over to — it is just me?
The document changes shape but does not disappear. A solo founder's version is: what genuinely cannot wait a week, what a client is told, who has emergency access if something goes badly wrong, and what is scheduled to happen automatically while you are away. Most solo businesses can be unavailable for a week; what they cannot do is be unavailable silently. Say it in advance and almost nothing breaks.

Should the covering person get extra pay?
If you are granting real authority, you are asking for something materially different from their normal job, and treating it as a favour is how it becomes resented. Whether that is money, time back, or the explicit developmental framing depends on the person — but decide it before you ask, not after they raise it.

What if a client escalates specifically because I am not there?
That is worth knowing about and it is usually a relationship-design problem rather than a holiday problem: the client has been trained that you personally are the escalation route. The fix is not available this week. It is the same work as the first 90 days of an account — establishing early who the client's actual counterpart is, before the habit forms.

Is this the same as a business continuity plan?
It is the readable half of one. A continuity plan covers being unavailable without notice, which needs a couple of things this document does not — mainly, someone other than you having emergency access, and a written record of what is where. If you are writing the holiday version anyway, adding those two turns it into the version that also covers the situation nobody plans for.

The point of the whole exercise

A founder who cannot take a week off does not have a scheduling problem. They have a business where every decision routes through one person by default, and that default is expensive long before it becomes a holiday question — it caps how much you can take on, it makes the business fragile, and it makes it worth less to anyone who might buy it.

The document is worth writing for the trip. It is worth keeping for what it tells you: read your decision-rights table and you are looking at an honest map of how much of your company currently runs on your presence rather than on its systems. Most founders have never seen that written down. Writing it is the useful part, and the week off is really just the deadline that forces it.