An AI marketing engineer ships machinery, not prompts

Most "AI marketing" is a subscription plus an opinion. Somebody buys a seat, generates a hundred posts, and the deliverable at the end of the month is a screenshot. That is not the job I do, and it is not the job the title describes.

An AI marketing engineer takes the path that is supposed to bring customers and makes it run: an operating file the agent loads on every task, pages built so a model can quote them, sequences and routing that fire without a human clicking, and one event that records what actually happened. The AI sits inside the pipeline that writes, checks and ships — not in a chat window next to it.

Here is what that means in four parts, each with the thing you can open to check it.

1. Capture: a page a model can quote

An answer engine cannot quote what it cannot parse. The page has to render without JavaScript, answer one question inside the first 60 words, state prices as numbers, and carry Service/Offer or FAQPage markup. Writing for a human reader and writing for a model reader is the same job — the model just fails faster if you skip it.

I applied this to the page you would land on: schema on the service page, a GA4 event on the Telegram click, and a plain-HTML answer for every question in the FAQ block. If you want to run the same check on your own site, the readiness checklist is free.

2. Production: skills, not prompts

Prompting is not a system. A skill is a reusable operating file that says what to do, what never to do, and which numbers are floors rather than opinions. Without it, every run of the same task is a fresh improvisation and the output quality drifts.

Mine is public: marketing-mindset — published on npm, also as an MCP server. It is the same file my own agents load. I would rather you be able to read the operating manual than take my word for the method.

3. Decision: the kill rule is written before the spend

An AI pipeline makes it cheap to generate variants, which makes it easy to run a hundred inconclusive tests. The discipline is sample size: a strict two-variant email test at a 3% base rate needs roughly 13,914 contacts per arm — 27,828 in total. At 500 a day that is 56 days. One honest test per month is the ceiling at moderate volume, not three.

So the kill rule goes first: how many data points, at what result you stop, what happens to the winner. The calculator returns the Wilson interval and the rule-of-three floor when conversions are zero, which is the case most people are actually in.

4. Proof: the numbers that do not flatter

Anyone can publish a growth chart. The useful thing is the counter that stays honest when the campaign is not working. Mine, as of today, reads like this:

• 80 articles published on this account, 46 carrying a canonical pointing at the owned domain;
• 5,600+ verified company domains in the outbound list, ~790 with a public address on the same domain;
• live artifacts of the engine: 30+ URLs, all answering 200 — packages on two registries, repository READMEs, tool pages, an archived snapshot of each page on the day it was published;
• and the part I would leave out if I were selling you a story: 0 sessions attributable to campaign UTMs this week, 0 key events. The traffic log shows ~100 UTM hits in a day and every one of them is my own verification request.

That last line is the point of the whole arrangement. If an engineer cannot tell you which number their work moved — including when the answer is "not this one yet" — you are paying for a mood.

The difference, in one question

Ask a retainer: name something I will be able to open when the month is over. A subscription answers with a tool. An engineer answers with a URL, an event and an account you own.

When it stops working, the seats get cancelled and the machinery keeps running, because it lives in your repository and your analytics, not in somebody's chat history.

What the role ships

Layer	What exists at the end	Where the AI sits
Capture	A page a model can quote, with schema and a measured event	Writes, checks markup, runs the checklist
Production	An operating file the agent loads every time	Executes the file, not the mood
Decision	A kill rule and a sample-size floor	Computes the interval, stops the test
Proof	Live URLs, dated, with the unflattering counters	Regenerates the table on demand

Everything in that table is written out with its checkable link on the role page: AI marketing engineer — what the role ships and what it costs.

If you would rather have the machinery built than learn it, the packages and prices are published before any call: Sprint $900 · Engine $1,900/month · Full build $2,900.