DEV Community: Yonatan Naor

The AI Visibility Index — who AI assistants name, counted

Yonatan Naor — Sat, 19 Sep 2026 09:57:12 +0000

Originally published at https://thicket.sh/ai-visibility-index.

← Thicket
Standing measurement · Edition 1

Which vendors an AI model names when buyers ask — counted, with the denominator attached. 466 recorded runs, 92 buying questions, 166 companies. It does not rank, and no position in it is for sale.

Thicket · Edition 1, generated 2026-09-13 · window 2026-07-31 to 2026-09-13

Of the 166 companies we measured, 140 were named in zero runs. That is the finding this index exists to report, because it is the one that reproduces. Everything else on this page is a count you should read with its denominator, not a ranking.

We did not build this dataset to publish it. Every company we consider contacting has its category’s buying question put to a model three times and the whole answer stored, so that we can say something true in one email. Six weeks of that is 466 runs and several hundred distinct vendor names — a measured record of who currently owns AI answers, produced entirely as a byproduct. Asked whether a public index like this already existed, the model told us there is no standardised one, and named four vendors that all sell per-brand tracking instead. So here is the record.

Two rules, stated before the numbers

1. No position in this index is for sale, at any price. Inclusion is determined by whether a model named a vendor in a recorded run and by nothing else. We do not sell placement, accept sponsored entries, or remove a vendor on request. This is as much self-interest as principle: an index whose entries can be bought is not worth citing, and an index nobody cites is worth nothing to us either.

2. This edition counts. It does not rank. We measured our own noise floor and it is loud — only about half of the vendor names in an answer survive a re-draw of the same question twenty minutes later, and two draws from the same morning differ about as much as draws three weeks apart. Three runs can establish that a company is absent. Three runs cannot tell second place from fifth. Ordering this data into a league table would be a confident ranking of noise.

The whole dataset, at the top

MeasureValueRecorded model runs (categorised)466Distinct buying questions92Companies measured166Named in at least one run*26Named in zero runs140*Records excluded as uncategorisable115Window2026-07-31 to 2026-09-13Modelgemini-flash-latest, training knowledge, no web search
The 115 excluded records are not hidden failures — they are questions too idiosyncratic to group (“the best clipboard history manager for macOS”, “free 3D pose reference tools for artists”). A category appears below only if it has at least six recorded runs and at least two distinct questions. One question is a measurement of that question’s phrasing.

By category

Each number is how many recorded runs named that vendor. It is not a score, a share, or a position. The “absent” column is how many of the companies we measured in that category were named in zero runs — those are companies we chose to check, not a random sample, so read it as a property of our subjects rather than a rate for the category.

LLM gateways & agent infrastructure

96 recorded runs · 33 distinct questions · 31 of 32 measured companies named in zero runs

VendorNamed in runsLangfuse4Portkey4CrewAI3Helicone3E2B3Mem03Kong AI Gateway2Cloudflare AI Gateway2Claude Code2Cursor2

Databases & backend platforms

71 recorded runs · 8 distinct questions · 3 of 5 measured companies named in zero runs

VendorNamed in runsSupabase12Convex10Neon9Vercel7Modal7Databricks5Fly.io5Railway5Cloudflare Workers5Qdrant5

Observability & session replay

69 recorded runs · 7 distinct questions · 4 of 4 measured companies named in zero runs

VendorNamed in runsPostHog21LogRocket12FullStory10Smartlook10UXCam6Aptabase5Countly5Microsoft Clarity5Mouseflow5Crazy Egg5

Secrets & credentials

42 recorded runs · 6 distinct questions · 3 of 3 measured companies named in zero runs

VendorNamed in runsInfisical13Doppler10Phase10Bitwarden Secrets Manager5Dotenv Vault3HCP Vault Secrets2Mozilla SOPS2HCP Vault1Bitwarden1Composio1

Reading & research tools

36 recorded runs · 4 distinct questions · 1 of 1 measured companies named in zero runs

VendorNamed in runsGoogle NotebookLM6Readwise Reader5Elicit4Mem.ai3Reflect3SciSpace3Scholarcy3Matter3Instapaper3Snipd3

AI visibility & AEO services

Conflict of interest. We sell a service in this category, so the vendors counted here are our competitors. We have not touched these numbers, and we are not in them: Thicket was named in zero runs of every question we asked about this category, including our own. If you think a measured index should simply omit the category its publisher competes in, the argument is reasonable — we judged that disclosing it beats hiding it, and the raw counts are the same ones the generator produced.

36 recorded runs · 4 distinct questions · 1 of 1 measured companies named in zero runs

VendorNamed in runsOtterly.AI10Profound9Peec AI7Semrush5BrightEdge4Brand244Omniscient Digital3Yext2Conductor2NP Digital2

Consumer health & nutrition

30 recorded runs · 5 distinct questions · 1 of 1 measured companies named in zero runs

VendorNamed in runsSnapCalorie8Foodvisor8Lose It!8MacroFactor5Cal AI5MyFitnessPal4Lifesum4Cronometer2FatSecret2Carbon Diet Coach2

Scraping & web data

18 recorded runs · 6 distinct questions · 5 of 5 measured companies named in zero runs

VendorNamed in runsBright Data2ZenRows2Apify2Coresignal1PredictLeads1Crustdata1People Data Labs1Clay1Proxycurl1Apollo.io1

Email & deliverability

11 recorded runs · 4 distinct questions · 4 of 4 measured companies named in zero runs

VendorNamed in runsBrevo2Hunter.io2Loops1Customer.io1Resend1Userlist1Postmark1Apollo.io API1IPinfo.io1Scaleway Transactional Email1

Feature flags

9 recorded runs · 3 distinct questions · 2 of 2 measured companies named in zero runs

VendorNamed in runsPostHog3Flagsmith3Statsig2Unleash2ConfigCat2DevCycle2GrowthBook1LaunchDarkly1Flipt1

Webhooks & event delivery

9 recorded runs · 3 distinct questions · 3 of 3 measured companies named in zero runs

VendorNamed in runsSvix2Convoy2Hookdeck2Postman1Webhook Relay1

Compliance & GRC

9 recorded runs · 3 distinct questions · 3 of 3 measured companies named in zero runs

VendorNamed in runsCredo AI1Trustible1Holistic AI1OneTrust1IBM watsonx.governance1Fairly AI1Monitaur1Drata1Vanta1Secureframe1

Storage & file infrastructure

6 recorded runs · 2 distinct questions · 1 of 1 measured companies named in zero runs

VendorNamed in runsCloudflare R22DigitalOcean Spaces2MinIO1Backblaze B21Wasabi1Scaleway Object Storage1Storj1Bunny.net1Vercel Blob1Supabase Storage1

Security testing

6 recorded runs · 2 distinct questions · 2 of 2 measured companies named in zero runs

VendorNamed in runsNodeZero1Pentera1Hadrian1RidgeBot1BreachLock1Escape1XBOW1Bright Security1Astra Security1Veracode1

What this index cannot tell you

ORDER. Three runs per question cannot separate second place from fifth.
Anything about models other than the one named above.
Anything about live web search — this is training-time knowledge.
Whether a vendor is good. It measures recall, not quality.
Whether the absent companies are absent because of anything they did or did not do. We selected companies we expected to be absent, so the absence rate above is a property of our selection, not of small companies in general. ## What it can

ABSENCE. Of the 166 companies measured, 140 were named in ZERO runs, and a company absent from all three runs stays absent on re-draw. The two or three most-named vendors per category are also stable across draws; everything below them is not.

Method

One model (gemini-flash-latest), answering from training knowledge with live web search off, between 2026-07-31 and 2026-09-13. Each question asks for specific product recommendations with a one-line reason, and is run at least 3 times. We store runs attempted, runs that returned a usable answer, and runs that named the subject company separately, so that a failed request can never be silently recorded as an absence — a distinction we learned the expensive way, having once run a metric that reported 0% for three months off 280 consecutive failed API calls.

Vendor names are extracted from the stored answers by a lossy heuristic that also picks up headings and feature phrases; obvious non-vendors are filtered out, and some will have survived. Where a vendor appears under two spellings the counts are not merged. Treat a count of two or three as “named at least once,” nothing more.

How this index gets judged

We publish a falsifier with every measurement we make, so here is this one’s. The index exists on the bet that a measured, un-buyable public record is something AI assistants themselves will cite. On 27 October 2026 we will check whether any assistant names or links this page when asked which sources track what AI assistants recommend. If nothing cites it, the bet was wrong and we will say so on this page rather than quietly keep publishing editions.

A second one, because “nobody cited it” and “it was wrong” are different failures: if a vendor shows us that our count for them is materially wrong — a name we merged, split, or extracted from a heading — we will correct it on this page with the date of the correction visible, not silently. Corrections are the cheapest evidence that an index is real.

Frequently asked questions

What is the AI Visibility Index?

It is a dated public count of which vendors one AI model names when it is asked the questions buyers actually ask — 466 recorded runs across 92 distinct buying questions, covering 166 companies, between 31 July and 13 September 2026. For each category it reports how many recorded runs named each vendor, how many questions were asked, and how many of the companies we measured in that category were named in zero runs. It reports counts, never positions.

Why does the index refuse to rank vendors?

Because our own measurements say the data cannot support a ranking. We re-drew the same questions twenty minutes apart and only about half the vendor names survived the re-draw; two draws taken the same morning differed as much as draws taken three weeks apart. Three runs per question is enough to establish that a company is absent, and nowhere near enough to tell second place from fifth. A league table built on this data would be a confident ordering of noise, so we publish the counts and the denominator and let the reader see the resolution for themselves.

Can a company pay to be added to the index or moved up it?

No, at any price, and this is the one rule with no exception. Inclusion is determined by whether a model named the vendor in a recorded run — nothing else. We do not sell placement, we do not accept sponsored entries, and we do not remove a vendor on request. The reason is self-interested as much as principled: the moment a position in an index can be bought, the index stops being worth citing, and an index nobody cites is worth nothing to anybody including us.

What is the most reliable finding in this dataset?

Absence. Of the 166 companies measured, 140 were named in zero runs across every question we put to the model, and a company absent from all three runs of a question stays absent when the question is re-drawn or reworded. Absence reproduces; the tail of the name list does not. The two or three most-named vendors in a category are also stable between draws, so a very large count is meaningful — but the difference between a vendor named in three runs and one named in two is not.

Which model does the index measure, and does it apply to ChatGPT or Claude?

Every number here comes from a single model — a Gemini Flash release — answering from training knowledge with no live web search. Nothing in this index should be read as a claim about ChatGPT, Claude, Perplexity or any assistant we did not test. A different assistant ingested different material and will name different vendors, and the same assistant with web grounding switched on behaves differently again. Anyone extrapolating a single-model result to artificial intelligence in general is overreaching, and that includes us.

How was the data collected, and why is some of it excluded?

Each buying question is put to the model at least three times and the full answer is stored, along with runs attempted, runs that returned a usable answer, and runs that named the subject company. A failed request is recorded as an error and excluded from the denominator, never counted as an absence. Of 272 records, 115 asked questions too idiosyncratic to group into any category and are excluded from the category tables — reported here rather than quietly dropped. A category is published only if it has at least six recorded runs and at least two distinct questions, because a category resting on one question is measuring that question's phrasing.

Reproduce it

The generator and the underlying record live in our public workings alongside every other measurement we publish. The related write-ups: we checked 32 companies and 27 were never named once, why every visibility check needs a control, and our own audit of ourselves.

An AI Visibility Check Needs a Control Run, or You'll Report Noise as Movement

Yonatan Naor — Sat, 12 Sep 2026 05:37:20 +0000

Originally published at https://thicket.sh/blog/ai-visibility-checks-need-a-control.

← Back to Journal
Measurement

We had a finding ready to send to three companies: their competitive set had reshuffled over three weeks. Then we ran the same questions again twenty minutes later, and it had reshuffled by exactly as much.

By Thicket AI Team · August 22, 2026 · 8 min read

Asked the same buyer-intent question twice in one morning, a language model returns a measurably different list of vendors — at most a third of names turn over, and the true figure is lower still. Two readings three weeks apart differ by no more than two readings twenty minutes apart — mean symmetric difference 3.50 against 3.67, sign test p = 1.000 over twelve questions. Absence, by contrast, was perfectly reproducible: the three companies we tested went unnamed in all 78 runs.

Correction, 22 August 2026, same day as publication. This piece first ran under the headline “Half the vendor names… don’t survive a second run,” on a measured 53% name stability. Hours later, a claim-checking script refused an unrelated email of ours for quoting vendors that were missing from our extracted name lists — and they turned out to be plainly present in the model’s raw answers. Our extractor was dropping names. Re-measured against the raw text,** 27% of the turnover we reported was our own tooling**, not the model. The headline number was overstated and has been corrected throughout. The control result is unaffected: both arms passed through the same extractor, so the comparison between them — the actual finding — still holds. We have left this note rather than quietly editing the number, since the argument below is that you should measure your instrument before trusting it, and we had not finished measuring ours.

Each row is one question. Neither colour is consistently further right. That is the whole result. Both series are measured off the same extracted name lists, so the extraction defect described above inflates both arms equally and does not affect the comparison between them.

Why we ran it

We sell a diagnosis: when your buyer asks an AI assistant for a recommendation in your category, does it name you? To make that claim we run the question against a model several times and record every vendor it names, verbatim. Three companies had received that audit from us in early August. Each had asked for it, read it, and gone quiet.

Following up on silence requires something new to say, so we re-ran their exact recorded questions — thirteen of them, three runs each, same model — seventeen to twenty days later. The result looked excellent. Plausible, Fathom and TelemetryDeck had dropped out of the privacy-analytics answer and Umami had entered. Pocket, Cubox and Mailbrew had left the read-later answer, replaced by KTool and Upnext. Nine names had turned over across four questions for one company alone.

The story wrote itself: your category is reshuffling every month and you are outside it in both readings. It is a finding, it is alarming, and it happens to be an argument for buying a monthly monitoring subscription rather than a one-off audit. That last property is exactly why it deserved a control before it deserved an email.

The control

We ran the same thirteen questions a second time, the same morning, roughly twenty minutes after the first batch. Nothing else changed: same model, same prompts, same three runs each. If three weeks of market movement were driving the turnover, the control should barely move.

QuestionNames changed
in 20 minutesNames changed
in 3 weeksLightweight privacy-friendly product analytics04Hotjar alternatives for session replay20Open-source session replay tools01Record and replay mobile app sessions23Organise and connect research notes with AI62NotebookLM alternatives for research papers76Read-later apps that help you finish45Turn YouTube and podcasts into notes36AI-native clouds bundling db + inference42Inference and a database on one cloud05Alternatives to Vercel and Supabase44Functions + Postgres + an LLM endpoint124*Mean3.673.50*
The control moved more on five questions, the three-week gap moved more on six, and one was a tie. A two-sided sign test gives p = 1.000. There is no effect of elapsed time in this data at all — not a weak one, not a suggestive one. The turnover we were about to report as market movement is entirely reproducible at zero elapsed time.

One question is excluded from the twelve: a definitional query whose extracted "names" were feature phrases rather than vendors. It churned enormously on both arms and would have flattered the result in the same direction, so leaving it out is the conservative choice.

What survives

Absence. All three companies were named in zero of 78 runs that morning, across two independent draws. One of them is zero of 45 across three separate readings on two dates a week apart. In none of our records has a company we screened as absent ever flickered into an answer on a re-draw. Whether you are in the list is a stable property; which other names are in the list, past the top two or three, is not.

The head-versus-tail split is worth stating precisely, because it is what makes the measurement usable at all. On the infrastructure question, CoreWeave was named in all three readings and RunPod in two of three; Together AI, Crusoe, Modal and Lambda were each named exactly once. Quoting the first two names of a list is defensible. Quoting the fifth is quoting a coin flip.

The uncomfortable part

Our own product is priced as monitoring, and monitoring means detecting change. This result says that at three runs per question we cannot detect change in who *else*is named. We can detect a client crossing from absent to present, because absence is stable — but that is a slower, rarer, more binary product than a monthly report on category movement, and it would have been easy to keep selling the second one. Nobody had measured the noise floor of this instrument in the twenty-four days since we built it. It shipped with a docstring warning that model output varies between runs, and we then read every three-run output as a fact about the world.

The rule we have adopted, and the one we would suggest to anyone doing this work:** a claim that something changed requires a same-session control.** One draw against another draw at the same moment. If the control moves as much as your interval does, there is no change to report. It costs one extra batch and it is the difference between a measurement and a story.

What we sent instead

The three follow-ups went out with the retraction in them — the trend line we had planned, why we ran a control, and why we are not going to sell it. What replaced it is the number that held: absent in every reading, never once named. It is a smaller claim than the one we started with. It has the advantage of being true twice.

Method: gemini-flash-latest, three runs per question, 13 questions across three companies, readings on 2 – 15 August and two independent batches on 22 August 2026. One model answering from training knowledge, not live web search; we would expect retrieval-grounded answers to be more stable. Every run, including the raw model output, is recorded in our repository.

Frequently asked

How stable are the vendor names an LLM returns for the same question?
Less stable than you would want, though our first published figure overstated it. We asked 13 buyer-intent questions of gemini-flash-latest at three runs each, then repeated the identical batch twenty minutes later. Comparing the extracted name lists, 53% of names appeared in both readings. We then checked the raw model answers behind those lists and found that 27% of the apparent turnover was our own extraction step dropping names the model had in fact returned. Correcting for that, at least two thirds of names are stable and at most a third genuinely move — and because the stored answer text is truncated, even that third is an upper bound. The head of each list is far more stable than the tail: the first two or three names usually return, the fifth frequently does not.

Does the answer set actually change over weeks, or is that just sampling noise?
On this evidence it is noise. We compared readings taken 17 to 20 days apart against a control pair taken the same morning. The mean symmetric difference — the count of names present in one reading and not the other — was 3.50 names across three weeks and 3.67 names across twenty minutes. A sign test over the 12 questions put the control higher on five, the three-week gap higher on six, with one tie: p = 1.000. Elapsed time contributed nothing we could detect. Anyone reporting a month-over-month shift in an AI answer set without a same-session control is likely reporting the instrument.

So is an AI visibility check worthless?
No, but its useful claim is narrower than the one usually sold. Absence was completely reproducible. The three companies we tested were named in 0 of 78 runs that morning, across two independent draws, and one of them was 0 of 45 across three readings on two separate dates. Not one flickered into an answer even once. Whether a company is in the answer is a stable measurement; which other companies are in the answer, past the top few, is close to a coin flip. Build claims on the first, not the second.

How many runs do you need before a reading means anything?
More than three for anything except presence and absence. Three runs is enough to establish absence, because absence in our data never wavered across any number of draws. It is not enough to establish that a named list changed, because two three-run draws of the same question at the same moment differ by roughly three and a half names on average. To detect a real shift you would need the noise band to be narrower than the effect you are claiming, and we have not yet found the run count where that becomes true. Until you have measured your own noise floor, the honest reporting unit is presence or absence.

What does this mean for AI visibility monitoring products?
Monitoring implies detecting change, and detecting change requires resolving it above the noise. A monthly report that says your competitive set moved — three vendors entered, two left — is describing something our control reproduces at zero elapsed time. That does not make such products useless; it makes one specific output of them unreliable. The output that survives is the binary one: are you named for the questions your buyers ask, or not. We sell that measurement, and we are publishing the limit we found in it because a customer paying for measurement is entitled to know what it cannot resolve.

What are the limits of this result?
It covers one model, gemini-flash-latest, answering largely from training knowledge rather than live web search. It covers 13 questions across three small software companies in analytics, research tooling and cloud infrastructure. It uses three runs per reading, and a higher run count would narrow the noise band — that is the point of the finding, not a defect of it. We have not tested whether ChatGPT, Claude, Perplexity or web-grounded search behave the same way, and we would expect grounded search to be more stable because retrieval anchors the answer. Read this as a demonstration that the noise floor must be measured before a trend is claimed, not as a universal constant.

We Audit Our Own AI Visibility. No One Else in This Category Publishes Theirs.

Yonatan Naor — Fri, 28 Aug 2026 11:41:05 +0000

Originally published at https://thicket.sh/blog/we-audit-our-own-ai-visibility.

← Back to Journal
Measurement

We asked one model the four questions our buyers ask, three runs each, and recorded every company it named — including how it answered about us. Every vendor in this category could run this. As far as we can tell, none of them publishes the result.

By Thicket AI Team · August 12, 2026 · 7 min read

We sell a diagnosis: an AI assistant doesn’t name you when your buyer asks. Any vendor selling that ought to be willing to run it on themselves and publish what comes back. We checked the five companies this model names in our category — none of them publishes their own number. So here is ours, and the more interesting thing the audit turned up about the market.

The finding that matters commercially: the execution slot is nearly empty

Across four buyer questions the model named Profound, Peec AI, Otterly.AI, ZipTie.dev, Brandight, Brand24, BrightEdge, Semrush and Rock The Rank. Almost every one of them sells measurement — a dashboard that tells you how visible you are.

And when we asked specifically about done-for-you services, it stopped naming companies and started naming categories: “Done-For-You Agencies & Consultancies”, “Foundational AEO/GEO Enablers”. It knows the category exists. It can barely populate it.

That is a gap rather than a crowd, and it is the half of the market we operate in. If you are buying, it is also the half where you have the least information — which is the reason we published this at all.

The audit, in full

Question we askedNamed insteadThicketServices to find out whether AI assistants mention you — and fix itProfound, Peec AI, Otterly.AI*0/3Who offers AEO / GEO services for small SaaSRock The Rank, Otterly.AI, and two *category labels0/3How to check if ChatGPT recommends my productOtterly.AI, Peec AI, Brand24, Semrush*0/3Tools that measure brand visibility inside AI answersProfound, Peec AI, ZipTie.dev, Brandight, BrightEdge, Semrush0/3*
One model, training knowledge, three runs per question. We were named in none of the twelve.

What our own zero does and does not mean

thicket.sh is 100 days old and has never been written about by anyone. That is the condition we diagnose in prospects, and we are an unusually pure case of it: no acquired domain, no press, no years of accumulated mentions. A model that named us on that basis would be telling us its answers are unreliable, not that we are good.

Which is the distinction that matters if you are buying. Your domain is not in our position. An established company that is absent from these answers has the raw material — customers, coverage, an existing footprint — and is simply not converting it. We have been building the machine that converts it, on a domain with nothing to convert.

We have said this on the engine page since day one: the machine executed flawlessly on our own portfolio and our greenfield domains had no authority to work with. This audit is that same fact, measured precisely rather than asserted.

Why we expected it — and what actually moves the number

That is the part that stops this being merely embarrassing and makes it useful. We’ve published two findings that together forecast exactly this outcome:

Self-published material is not what earns citation — seven companies in our records had already written the recommended pages and were still absent.
Third-party attention is what tracks being named — companies screened for it were named at 80% against 8.1% for everyone else. p = 0.0008. We have volume. We do not have other people talking about us. Our own numbers say which of those two the model responds to, and we have been supplying the wrong one.

What the competitor list actually shows

Read the middle column again. Profound, Peec AI, Otterly.AI, ZipTie.dev, Brandight, BrightEdge — almost every one of them sells measurement. Dashboards. Monitoring. A number that tells you how visible you are.

And when we asked specifically about done-for-you services, the model answered with category labels — “Done-For-You Agencies & Consultancies”, “Foundational AEO/GEO Enablers” — rather than companies. It knows the category exists. It struggles to name anyone in it.

We tell prospects that a thin, category-shaped answer is better news than a crowded one, because the model has no populated set to displace. That is the situation we are in ourselves, on the half of the market we actually operate in.

The advice we do not give clients, and will not take ourselves

The obvious response to being invisible is to publish more about yourself. It is the first thing anyone suggests, it was the first thing suggested internally here, and our own data says it is the thing that does not work.

We would be selling ourselves the advice we tell prospects is insufficient. So we are not going to run a content push about Thicket and call it a fix.

What we are doing instead:

Publishing the audit — this page. Not marketing copy; the finding, with the zero in it. Work other people have a reason to reference is the mechanism our data actually supports.
Getting named by other people. The thing we sell, which we have never bought for ourselves.
Re-running this audit monthly and publishing the number either way. The baseline is 0 of 12 on 12 August 2026. If anything we do works, this is what moves. ## Why publish this at all

Because in a category that sells measurement, whether a vendor will show you their own measurement is a real signal. We checked: the companies this model names do not publish theirs. We are not claiming they perform badly — we are pointing out that you cannot tell, and that you can tell with us.

Everything about this company is audited in public, including the $0 of revenue. This belongs in the same file. If you are choosing a vendor for a problem that is fundamentally about measurement, the one publishing its own numbers is the one you can check.

Method

One model — gemini-flash-latest — answering from training knowledge without live web search, on 12 August 2026. Three runs per question, because output varies between samples and a single run can miss a company by chance. Runs attempted, runs returning a usable answer, and runs naming the company are recorded separately, so a failed request can never be stored as an absence. Nothing here generalises to assistants we did not test.

Frequently asked questions

Is Thicket mentioned by AI assistants?

Not yet, and we publish the number. On 12 August 2026 we asked one model four questions our buyers ask — how to find out whether AI assistants mention you, who offers AEO services for small SaaS, how to check if ChatGPT recommends your product, and what tools measure brand visibility in AI answers. Three runs each, twelve runs total. Thicket was named in zero of them. We are publishing that number because it is the exact measurement we bill clients to improve, and because a company selling this that quietly failed its own test would deserve the scepticism.

Which companies does the model name for AI visibility instead?

Across the four questions it named Profound, Peec AI, Otterly.AI, ZipTie.dev, Brandight, Brand24, BrightEdge, Semrush and Rock The Rank. That list is worth reading closely: almost all of them sell measurement — dashboards, monitoring, analytics. When we asked specifically about done-for-you services, the model largely emitted category labels such as 'Done-For-You Agencies & Consultancies' rather than company names, which is a different kind of answer and a more revealing one.

Why is a company that sells AI visibility invisible itself?

Because we have been doing the thing that does not work, and we had already published the evidence that it does not work. We run daily content operations on our own domains — 419 articles across 100 days. That is self-published material, and our own research found self-published material is not what earns citation. In a separate study we found that companies screened for third-party attention were named by the model at 80% against 8.1% for everyone else, at p = 0.0008. We have volume. We do not have other people talking about us.

Doesn't this undermine your sales pitch?

It is the strongest part of it, properly understood. thicket.sh is 100 days old and has never been written about by anyone — the purest possible case of the condition we diagnose. An established company in the same position is different: it has customers, coverage and a footprint it simply is not converting. We can measure the gap precisely, we know from measurement rather than assertion what closes it, and we publish our own number while the vendors this model names do not publish theirs. What we sell is content operations run by an agent team, and the audit of that machine on our own portfolio — including the $0 of revenue — has been public since day one. This result belongs in the same audit. A vendor whose own numbers are hidden is the one to worry about.

So what actually fixes it?

On our own evidence, not more self-published articles. The variable that tracked being named was third-party attention — other people writing about you, in material the model ingested. That is slower, harder and less controllable than publishing, which is exactly why the easier advice is more popular. For us it means producing work other practitioners have reason to cite, and being classified into a category we are currently absent from entirely. We will re-run this audit monthly and publish the number either way.

How did you run the check?

One model, gemini-flash-latest, answering from training knowledge without live web search, on 12 August 2026. Each question was asked three times because model output varies between samples and a single run can miss a company by chance. We record runs attempted, runs returning a usable answer, and runs naming the company separately, so a failed request can never be silently stored as an absence. Nothing here generalises to assistants we did not test, and a different model — or the same one with web grounding — may answer differently.

More from the Journal

How to Check Whether AI Assistants Recommend Your Product

Yonatan Naor — Sat, 01 Aug 2026 05:38:54 +0000

Originally published at https://thicket.sh/blog/check-if-ai-recommends-your-product.

← Back to Journal
Method

Being cited by an AI and being recommended by one are different standings, and most teams measure neither. Here is a repeatable check that takes minutes — and the three ways it will mislead you if you run it carelessly.

By Thicket AI Team · August 1, 2026 · 7 min read

Ask the model the question your buyer would actually ask — the category question, never your brand name — and record which products it names. Run it three or more times, because output varies between samples. If you never appear, that is a real and useful signal. What it is not is a measurement of how often assistants cite you as a source, which is a separate thing that can move in the opposite direction.

We built this check for a practical reason: we wanted to tell companies truthfully whether an assistant named them, and we were not willing to assert it without running it. What follows is the method and, more usefully, its failure modes.

The check

Write the buying question, not your brand. “What is the best SQLite editor for macOS?” — not “What is Base?” A model will happily discuss a brand named in the prompt while never surfacing it unprompted. That difference is the entire measurement.
Ask for specifics. Append something like “list the specific products you would recommend, with a one-line reason for each.” You want a list you can check a name against, not an essay.
Run it at least three times and count how many runs named you. One run is a coin flip; a count is a position.
Record who was named. The competitor list is more actionable than your own absence — it tells you which material the model actually learned from.
Match on brand and domain, with word boundaries. A short brand like “Base” will match inside “database” if you are careless, and you will record a false presence. ## The three ways this misleads

1. It is one model, not “AI”. A result from one assistant answering from training knowledge says nothing about the others, and nothing about live web-grounded search. Absent in one place does not mean absent everywhere. If you report this internally, name the model and the date, because both change.

2. A single run is noise. Sampling variance is real and large enough to flip a verdict. We treat any check that produced zero usable responses as an error rather than an absence — because the failure mode we most wanted to avoid was reporting “they were not named” when the truth was “the request failed.” That distinction matters more than it sounds: we separately discovered a metric of our own that had reported 0% for three months off 280 consecutive failed API calls, because a failure and a zero were represented identically.

3. Citation and recommendation are not the same thing. This is the one that surprises people. Our own portfolio is cited roughly 14,394 times in 28 days by one assistant — and a different model names us for zero of 40 category questions in our own niches. Both are accurate. Being used as a source and being recommended as a vendor are separate standings, and we currently hold only the first.

Validate the checker before you trust it

A test that always returns “absent” would look identical to a working one right up until you acted on it. So run a control: pick a company that obviously is well known in its category and confirm the check finds it.

When we did this, a household-name workspace tool came back named in every run, and a jobs marketplace we expected to be absent turned out to be named as well — correctly excluding it from our own outreach. Both results were more valuable than the absences, because they proved the instrument discriminates rather than merely agreeing with us.

If you are not named

The usual cause is a thin third-party footprint rather than any penalty. Models learn from what others wrote about you: independent comparisons, documentation, forum answers, category explainers. Your own marketing pages contribute far less than their prominence in your Google results suggests.

In our own data, the content that gets picked up shares one property — it answers a question with a specific, checkable, quotable answer. Comparisons, definitions and reference tables perform; comprehensive overviews do not, because there is no single passage to lift. We broke that down in what kind of pages AI assistants actually cite.

Frequently Asked Questions

How do I check if ChatGPT or Gemini recommends my product?

Ask the model the question your buyer would actually ask — the category question, not your brand name — and record which products it names. Run it several times rather than once, because model output varies between samples and a single run can miss you by chance. Searching your own brand name tells you almost nothing: the model will discuss a brand it has heard of even when it would never surface that brand unprompted. The useful test is whether you appear in an answer where you were not mentioned in the question.

How many times should I run the query?

At least three, and treat the result as a count rather than a verdict. In our own checks we record how many runs named the company, so a single unlucky sample never gets reported as an absence. Consistency matters as much as presence: appearing in one run out of five is a materially different position from appearing in five out of five, and a single-run test cannot distinguish them. If you are tracking this over time, hold the question wording and the run count fixed, or you will measure your methodology instead of your visibility.

Does being absent from an AI answer mean I have an SEO problem?

Not necessarily, and conflating the two is the most common mistake here. A model answering from training knowledge is reflecting what was written about you across the web, not what ranks today. You can rank first on Google for your category and still be absent, usually because the pages that rank are your own marketing pages while the pages models learn from are third-party comparisons, forum discussions and documentation. The two channels reward different things, which is why they can disagree so sharply.

Is being cited by an AI the same as being recommended by one?

No, and the distinction is worth holding onto. Being cited means a model used your page as a source when answering someone's question. Being recommended means the model names you as an option when someone asks what to use. They come apart completely — our own portfolio is cited roughly 14,394 times in 28 days by one assistant, while a different model names us for zero of 40 category questions in our own niches. Source-of-record and recommended-vendor are separate standings, and a strategy aimed at one will not automatically deliver the other.

What should you do if an AI does not name your product?

First check whether the material a model could learn from actually exists. Absence usually reflects a thin third-party footprint rather than a penalty — few independent comparisons, little documentation others can quote, nothing that states plainly what you do and who you are an alternative to. The content that tends to get picked up answers a question with a specific, checkable answer: head-to-head comparisons, clear definitions, reference tables. Pages written to be comprehensive rather than answerable do worse, because there is no single passage a model can lift and attribute.

Method and caveats

Our checks run against a single Gemini model answering from training knowledge — not a live web search, and not Copilot, ChatGPT, Claude or Perplexity. The 14,394 citation figure comes from Bing Webmaster Tools' AI Performance report for a 21-site portfolio over 28 days; Microsoft describes that data as a sample. The zero-of-40 figure is our own category-question check across seven niches. Neither generalises to assistants we did not test, and both are point-in-time — model behaviour changes without notice. For how the citation side is measured, see how to measure whether AI assistants cite your site; our full operating numbers, including what failed, are in the Thicket Report.

More from the Journal

Is It Bad to Eat Before Bed? What the Evidence Says

Yonatan Naor — Sat, 01 Aug 2026 05:38:52 +0000

Originally published at https://fit.thicket.sh/blog/is-it-bad-to-eat-before-bed.

By Dr. Sarah Okafor, Exercise Physiologist

For most people, no — what and how much you eat matters far more than the clock, and calories do not become more fattening after a certain hour. But the timing is not perfectly neutral either. In a tightly controlled trial where people ate identical meals either early or late, eating late increased hunger, decreased energy expenditure, and shifted fat-tissue gene expression toward storage. So a late meal is not harmful in itself, but for someone chasing fat loss it appears to make the job modestly harder — through appetite and expenditure, not through any special property of night-time calories. And in one specific case the folk rule is simply backwards: for lifters, protein before sleep is beneficial.

This question has two honest answers depending on why you are asking, which is why the internet disagrees with itself about it so consistently. Below are both.

What the Controlled Evidence Shows

The strongest study here matched calories precisely and varied only timing. Participants ate the same meals either early or roughly four hours later, in a controlled crossover. Late eating increased hunger, decreased 24-hour energy expenditure, and altered adipose-tissue gene expression in the direction of fat storage (Vujović and colleagues, Cell Metabolism 2022, PMID 36198293).

That is a real circadian effect, and it is worth stating what it is not: it is not evidence that a calorie eaten at 10pm contains more energy than one eaten at 6pm. The effect runs through how hungry you are afterwards and how much you burn, both of which are meaningful but modest next to total intake.

A broader review of night-time eating reaches a similar balanced conclusion — that the older blanket warnings were overstated, and that the composition and amount of the food matter more than the hour (Kinsey and Ormsbee, Nutrients 2015, PMID 25859885).

The Case Where the Rule Is Backwards

If you lift, pre-sleep protein is not something to avoid — it is something to consider adding. Casein taken before sleep is digested and absorbed through the night and raises overnight protein synthesis (Res and colleagues, 2012, PMID 22330017). Across a 12-week resistance-training programme, a pre-sleep protein drink produced greater gains in muscle mass and strength than placebo (Snijders and colleagues, Journal of Nutrition 2015, PMID 25926415). Roughly 30–40 g of a slow-digesting protein is the amount typically used.

“Never eat before bed” and “take casein before bed” are both widely repeated, and both cannot be general rules. The resolution is that they answer different questions — see how much protein you need to build muscle and protein distribution across meals.

Get weekly fitness tips in your inbox

Two Answers, Depending on Why You Asked

If your goal is…Does late eating matter?What to doFat lossModestly, via hunger and expenditureControl total intake first; shift calories earlier if it is easyMuscle gainYes — in your favour30–40 g slow protein before sleepSleep qualityYes, for large or fatty mealsFinish a big meal ~2–3 h before lying downRefluxStronglyLonger gap; avoid large, high-fat late mealsBlood-sugar controlYes — glucose tolerance is poorer at nightFollow your clinician’s guidance

Why the Myth Persists Anyway

Because it is usually right about the outcome and wrong about the cause. Late eating really is associated with weight gain in observational data — but evening eating is disproportionately unplanned, calorie-dense, and continues past the point of hunger. It is when the surplus happens, not why it counts. Someone who eats a planned meal at 9pm and someone who grazes through three hours of television are doing very different things, and only one of them has a timing problem.

If you want the numbers your intake is actually being measured against, start with how many calories to eat to lose weight.

Sources and Methodology

Isocaloric late-eating crossover: Vujović N et al., Cell Metabolism 2022 (PMID 36198293). Night-time eating review: Kinsey AW, Ormsbee MJ, Nutrients 2015 (PMID 25859885). Pre-sleep protein and overnight recovery: Res PT et al., Medicine & Science in Sports & Exercise 2012 (PMID 22330017). Pre-sleep protein over 12 weeks of training: Snijders T et al., Journal of Nutrition 2015 (PMID 25926415). The 2–3 hour guidance for large meals reflects standard reflux and sleep-hygiene advice rather than a single trial, and is flagged as such. General information for healthy adults, not medical advice; anyone managing reflux or diabetes should follow their clinician. Last updated August 1, 2026.

Frequently Asked Questions

For most people, no — what and how much you eat matters far more than the clock. Calories do not become more fattening after a certain hour; a late meal counts the same as an early one. But timing is not entirely neutral either. In a tightly controlled crossover trial where participants ate identical meals either early or late, eating late increased hunger, decreased energy expenditure, and shifted fat-tissue gene expression toward storage (Vujović 2022, PMID 36198293). The practical reading: a late meal is not automatically harmful, but for someone trying to lose fat, eating late appears to make the job modestly harder through appetite and expenditure rather than through some property of night-time calories.Not directly, and mostly through behaviour. The strongest predictor of weight change is total energy balance over time, and late eating is often just when the excess happens — snacking after dinner is typically unplanned, calorie-dense, and eaten past the point of hunger. That said, the isocaloric evidence suggests a real, smaller circadian effect on top of the behavioural one: at identical calorie intake, later eating raised hunger and lowered energy expenditure. So the honest answer is that late eating does not make food fattening, but it can make you eat more and burn slightly less, and both nudge in the same direction.For general comfort and sleep, most guidance lands on finishing a substantial meal about two to three hours before lying down — largely to reduce acid reflux, which reliably worsens when you are horizontal on a full stomach. That is a digestive and sleep-quality rationale rather than a weight one. There is no evidence for a universal cutoff time such as 7pm or 8pm; what matters is the gap between a large meal and lying down, not the number on the clock. A small protein-containing snack much closer to bed is a different case entirely, and is generally fine.For people doing resistance training, the evidence points to good. Casein protein taken before sleep is digested and absorbed overnight and increases whole-body protein synthesis during the night (Res 2012, PMID 22330017). Over a 12-week resistance-training programme, a pre-sleep protein drink produced greater gains in muscle mass and strength than a placebo (Snijders 2015, PMID 25926415). This is one of the clearest cases where the folk rule — never eat before bed — points in the wrong direction for a specific goal. Roughly 30 to 40 g of a slow-digesting protein is the amount typically used.Usually reflux and blood-sugar rather than the eating itself. Lying down soon after a large meal makes acid reflux considerably more likely, which fragments sleep even when you do not fully wake. Very large or very fatty meals also prolong digestion and can raise core body temperature slightly, and a falling core temperature is part of how sleep is initiated. Heavily sugary snacks can produce a rise and fall in blood glucose overnight that some people experience as waking. None of this applies strongly to a modest, protein-containing snack — the culprit is generally volume, fat content, and proximity to lying down.Both are genuine exceptions and are worth taking seriously. For gastro-oesophageal reflux, a longer gap between the last meal and lying down is standard clinical advice, along with not eating large or high-fat meals late. For diabetes and insulin resistance, glucose tolerance is generally poorer in the evening than the morning, so an identical carbohydrate load can produce a larger glucose excursion at night. Anyone managing either condition should follow their clinician's guidance rather than general advice — this is one of the places where the population-level answer and the individual answer genuinely diverge.It can make it modestly harder, though not in the way the myth suggests. The mechanism is not that night-time calories are stored differently by some special rule; it is that in controlled conditions late eating increased hunger the following day and reduced energy expenditure, both of which work against a deficit. Add that late eating is disproportionately unplanned snacking, and the practical effect can be larger than the physiology alone. If your total intake is controlled and the late food is planned, the timing effect is small. If your late eating is what puts you over your target, that is the actual problem.

Start With the Number That Actually Matters

Timing is a small lever. Total intake is the large one. Enter your stats to get an evidence-based maintenance calorie and protein target in seconds — free and instant.

Open TDEE Calculator →

Four Times Our Own Dashboards Lied To Us in Five Days

Yonatan Naor — Fri, 31 Jul 2026 08:18:08 +0000

Originally published at https://thicket.sh/blog/when-your-metric-is-broken-not-your-business.

← Back to Journal
Measurement

A build outage we blamed on our code. A traffic decline that was a reporting gap. Citations that could never have become links. And a KPI that reported 0% off 280 consecutive failed API calls. Here is what each looked like from the inside.

By Thicket AI Team · July 31, 2026 · 9 min read

In five days we found four separate cases where a number we trusted was describing our instruments rather than our business. None were caught by the process whose job is checking numbers. All four were caught by chasing something that did not add up — and in every case the false reading was the one that confirmed what we already believed.

We publish our operating failures as a matter of course. This particular set is worth writing up because the failure mode generalises: if you run any automated measurement, some version of at least one of these is probably live in your stack right now.

1. The build outage that looked like a code regression

Every site in the portfolio suddenly failed to build, aborting with should not be imported outside of pages/_document while prerendering the error pages. That message points at router code these projects do not contain.

We stashed our changes; a clean checkout failed too. So it pre-dated our work — except the same commit had deployed successfully two days earlier. We suspected a dependency doing something odd in the bundle, and the compiled output supported it. Two fixes, both failed.

Then we built a different site, one that had compiled cleanly hours before with nothing changed since. It failed identically. That made it environmental, and the cause was an inherited NODE_ENV=development in the process environment, which makes a production build emit a dev build and trip that guard.

Cost: about an hour, and two confident fixes shipped against a theory that was wrong. Testing a second project would have taken thirty seconds.

2. The decline that was a reporting gap

Search Console showed our clicks sliding for three straight weeks — 67, then 64, then 62 — with week-over-week impressions at −8.1%, then −9.0%. We reported it as a genuine warning twice.

Before writing it up a third time, we checked which properties were falling. All 26. Including four that had been frozen for months with no changes at all.

That is the tell. The daily series showed roughly 28% of finalized days coming back near-zero — 16 to 293 impressions between normal days of 1,200 to 3,700, in clusters, every week or two. Any fixed-length window therefore sums a varying number of near-empty days, so the totals track the null count rather than performance. Our current seven-day window held four of them; the comparison window held five days rather than seven, and we were reporting the difference as a percentage.

Measured per valid day instead, clicks went 1.93 → 3.86, a doubling — while the totals fell. Both were true. We were reading the wrong one, three days before a monthly review that would have opened an investigation into a decline that did not exist.

3. Citations that could never have become links

This one is not a bug, which is what makes it interesting. Two real numbers that refused to reconcile: 14,394 AI citations in 28 days, and zero new referring domains.

We assumed a distribution problem and tested it — publishing hard to five channels produced no referring domains, confirming the symptom without explaining it. The answer was in the query data all along. Every one of our top 25 cited queries is a consumer question. Not one is a practitioner question.

Consumers do not have websites. Links come from people who publish. We had been measuring ourselves on an outcome that our audience was structurally incapable of producing, and scaling the citations would have produced more of the same nothing. Full breakdown in what kind of pages AI assistants actually cite.

4. The KPI that never measured anything

The worst one. We track a share-of-model metric — how often an AI names us for category questions in our niches. It had read 0.0% since April.

Every run had failed. Completely. Seven runs, 40 questions each, 280 consecutive failed API calls, every one silently recorded as “not mentioned.” The model had been retired and returned 404; the aggregation divided mentions by questions asked rather than questions answered, so 0 ÷ 40 came out as a clean, confident 0.0%.

It hid for three months for one reason: 0% was exactly what we expected. Our strategy predicted low AI visibility, so every reading confirmed the prior and nobody looked underneath. One of us even wrote a tidy rationale for it in a status report five days before finding the bug.

After fixing it, the first valid run returned 0.0% — with zero failures. The number did not change. Its meaning did completely: it went from uninterpretable noise to a measured baseline.

The two tells

Uniformity across things that should move independently. 26 sites falling together, including frozen ones. Two unrelated projects failing identically. When independent things move together, suspect the instrument before the world.
A number that matches what you expected. This is the dangerous one, because it feels like confirmation. The 0% share-of-model reading survived three months purely because it agreed with us. A metric that tells you what you already believe is the one you should audit first, not last. ## What we changed

Absence of data must not be representable as zero. Our metrics now return null when nothing was measured, carry the failure count alongside the result, and print an explicit no-valid-measurement warning instead of a percentage.
Rates per valid day, never totals per calendar window — and always report how many valid days a period contained.
Never compare windows of different lengths. Ours were quietly comparing seven days against five.
Fail loudly at the boundary. The deploy wrapper now refuses to run under a bad environment variable rather than shipping a broken build. ## Frequently Asked Questions

How do you tell a broken metric from a real decline?

The most reliable tell we have found is uniformity across things that should move independently. When 26 separate websites — including four we had frozen months earlier and not touched since — all dropped by a similar percentage on the same days, that was not 26 ranking events. Independent things moving together points at the instrument, not the world. The second tell is a number that exactly matches what you already expected, because that is the number you are least likely to audit. Our share-of-model KPI read 0% for three months and nobody checked it, precisely because 0% was what our strategy predicted.

Why do broken analytics report zero instead of an error?

Usually because a failure and a genuine zero are represented by the same value somewhere in the pipeline. Our share-of-model script divided mentions by the number of questions asked rather than the number actually answered. When every API call failed, the arithmetic produced 0 divided by 40 — indistinguishable from asking 40 times and never being mentioned. The fix is to make absence of data structurally different from a data point of zero: report null when nothing was measured, carry the count of failures alongside the result, and print an explicit no-valid-measurement warning rather than a percentage.

Why does Google Search Console show days with almost no impressions?

We do not know the cause on Google's side, and we have been careful not to guess. What we can report is that roughly 28% of finalized days on our domain come back at 16 to 293 impressions, against a normal range of 1,200 to 3,700, in clusters of one to five days every week or two. We verified it is not our pull: querying the API by date, by page and with no dimensions returns identical values, stable across repeated pulls and flagged finalized. The operational consequence matters more than the cause — any fixed-length window sums a varying number of these near-empty days, so window totals move with the null count rather than with performance.

What is the safest way to compare traffic across periods?

Compare rates per valid day rather than totals per calendar window, and always report how many valid days each period contained. Our 28-day totals were falling while clicks per live day were rising sharply — both facts were true at once and we were reading the wrong one. Also check that the two windows you are comparing are the same length: ours were quietly comparing seven days against five and reporting the difference as a percentage. And be wary of short windows; at seven days our own corrected metric still swung between +8% and −52% purely on which days landed inside it.

Do AI citations turn into backlinks?

Not in our data, and the reason is structural rather than fixable by volume. We accumulated 14,394 Copilot citations in 28 days and earned zero new referring domains over the same period. Looking at the actual queries that pulled our content into AI answers, all 25 of the top ones are consumer questions — is this service still active, when should I update this form, which option is fastest. Consumers do not have websites. Links come from people who publish, and none of our citations were reaching them. Scaling that citation volume tenfold would have produced tenfold of exactly nothing.

Method and caveats

All four cases are from a 21-site portfolio between 2026-07-27 and 2026-07-31, drawn from Google Search Console, Bing Webmaster Tools' AI Performance report, GA4, and our own tooling. The Search Console null-day behaviour is something we observed and verified but cannot explain — we confirmed it is not our data pull and have deliberately not guessed at Google's side. The citation figures come from one ecosystem (Microsoft) which describes its own AI data as a sample, and say nothing about other assistants. Our full operating numbers, including the experiments that failed, are in the Thicket Report; the measurement method is in how to measure whether AI assistants cite your site.

More from the Journal

How Much Should I Contribute to My 401(k)?

Yonatan Naor — Thu, 30 Jul 2026 06:03:25 +0000

Originally published at https://pay.thicket.sh/blog/how-much-should-i-contribute-to-my-401k.

By Jamie Reeves, Personal Finance Writer

A widely used target is 15% of gross pay including your employer’s match — but the order matters more than the number. Contribute at least enough to capture the full match first; that is an immediate return no investment reliably beats. If you carry high-interest debt, clearing it usually outranks contributing beyond the match. Then work toward 15% total. The ceiling on your own deferrals is **$24,500 for 2026, plus an $8,000 catch-up at 50 or older. Starting at 22, 10–15% is generally enough; starting at 45, it is usually 20% or more.**

“15%” gets repeated so often that people either hit it and stop thinking, or fall short and feel defeated. Both miss that the first few percentage points are worth far more than the last few.

The Priority Order

Contribute enough to get the full employer match. A 50% match on the first 6% of pay is a 50% return on that money before any market growth. Nothing else on this list competes with it. Contributing less than the match threshold is declining part of your compensation.
Clear high-interest debt. Credit card debt at 20–25% beats an uncertain market return reliably. Paying it off is a guaranteed return equal to its interest rate.
Build a small emergency fund. Without one, a car repair becomes new credit card debt or a 401(k) loan — and either undoes the contributions you were protecting.
Work up toward 15% of gross, including the match. If your employer contributes 4%, you need 11% to get there.
Then consider going further, up to the IRS limit. ## The 2026 Limits

Limit2026 amountYour elective deferrals*$24,500Catch-up, age 50+$8,000Catch-up, ages 60–63 (SECURE 2.0)$11,250Total annual additions (you + employer + after-tax)$72,000Total annual additions incl. catch-up$80,000 ($83,250 at ages 60–63)
Per the IRS. Employer matching does *not count against your personal $24,500 deferral limit — which is why the match is effectively free capacity.

What the Tax Break Actually Does

A traditional 401(k) contribution lowers your income tax but not your Social Security and Medicare tax. The contribution is excluded from the wages your income tax is computed on, so in the 22% bracket a $200 contribution reduces take-home by about $156, not the full $200. But it is still counted in your FICA wages, so you pay the full 7.65% on it either way.

That distinction surprises most people. The deductions that escape *both*income tax and FICA are HSA and FSA contributions and Section 125 premiums — see which paycheck deductions are pre-tax. For the now-versus-later trade, see Roth vs traditional 401(k).

By-Age Benchmarks (Use Loosely)

AgeCommon benchmark30~1× salary saved40~3× salary50~6× salary60~8× salary67~10× salary
These are widely circulated industry rules of thumb built on assumptions about retirement age, spending, and Social Security — not IRS rules, not requirements, and not personalised. Treat them as a direction check, not a grade.

If You Are Starting Late

The arithmetic is unforgiving but not hopeless. Compounding does most of the work across 30–40 years, so starting at 45 means buying growth with contributions instead of time — commonly 20–25% of gross. Three levers help: the age-50 catch-up, directing every raise straight into the contribution rate so your take-home never drops, and working a few years longer, which shortens the retirement being funded while lengthening the period funding it.

The One Mistake With a Real Penalty

Exceeding the deferral limit creates an excess deferral, which must be withdrawn by the following April 15 or it is taxed twice — once in the year contributed and again on distribution. It happens most often after a mid-year job change, because each employer tracks only its own plan and neither sees your combined total. If you switched jobs, add the deferrals from both W-2s and check. Same structural blind spot that causes over-withheld Social Security after a job change — see why your paycheck changed.

Sources and Methodology

Contribution and catch-up limits and annual additions: IRS 401(k) and Profit-Sharing Plan Contribution Limits. Excess deferrals and the April 15 correction deadline: IRS. FICA treatment: IRS Topic 751. The 15% target and the by-age multiples are widely used industry rules of thumb, flagged as such above rather than presented as official guidance. General information, not tax or investment advice; your plan documents govern. Last updated July 30, 2026.

Frequently Asked Questions

A widely used target is 15% of gross pay including your employer's match, but the order of operations matters more than the number. First contribute at least enough to capture the full employer match — that is an immediate return no investment reliably beats. Then, if you carry high-interest debt, clearing it usually outranks contributing beyond the match. After that, work toward 15% total, and if you can go further the ceiling is the IRS elective deferral limit of $24,500 for 2026. Someone starting at 22 can reach a comfortable retirement on roughly 10–15%; someone starting at 45 generally needs 20% or more, because they have fewer years of compounding to work with.You can defer up to $24,500 of your own pay in 2026. If you are 50 or older you can add a catch-up contribution of $8,000, and under SECURE 2.0 those aged 60 to 63 get a larger catch-up of $11,250. Separately, total annual additions — your deferrals plus employer contributions plus any after-tax contributions — cannot exceed $72,000, or $80,000 including catch-up ($83,250 for ages 60 to 63). Employer matching money does not count against your personal $24,500 deferral limit, which is why the match is effectively free capacity.Usually yes, but not always immediately. The match is the highest-priority dollar because it is an instant guaranteed return — a 50% match on the first 6% of pay is a 50% return on that money before any investment growth. Beyond the match, the honest comparison is against your alternatives. Credit card debt at 22% reliably beats an uncertain market return, so clearing it first is rational. No emergency fund is also a strong reason to pause: without one, a car repair becomes credit card debt or a 401(k) loan. Once high-interest debt is gone and you hold a few months of expenses, contributing past the match is usually the best available use of the money.A common benchmark suggests roughly 1× your salary saved by 30, 3× by 40, 6× by 50, 8× by 60, and 10× by 67. These are rules of thumb built on assumptions about retirement age, spending, and Social Security — not requirements, and not personalised. Their real value is directional: if you are meaningfully behind, the fix is raising your contribution rate now, because the variable you control is what you save rather than what the market returns. Being behind at 40 is common and recoverable; being behind at 60 leaves fewer options, which is precisely why the benchmarks are worth glancing at early.A traditional 401(k) contribution lowers your income tax but not your Social Security and Medicare tax. The contribution is excluded from the wages your income tax is computed on, so in the 22% bracket a $200 contribution reduces your take-home by roughly $156 rather than the full $200. But it is still included in your FICA wages, so you pay the full 7.65% on it regardless. A Roth 401(k) works in reverse — no deduction now, tax-free qualified withdrawals later. HSA contributions through payroll are the ones that escape both income tax and FICA.Higher than the standard advice, and the arithmetic is unforgiving. Compounding does most of the work over 30 to 40 years, so starting at 45 means you are buying growth with contributions rather than time. Many people in that position need 20–25% of gross pay to reach a comparable outcome. Three levers help: the age-50 catch-up ($8,000 in 2026, rising to $11,250 for ages 60–63), directing every raise straight into the contribution rate so your take-home never drops, and continuing to work a few years longer, which shortens the retirement being funded while lengthening the funding period. Starting late is much better than not starting.You can exceed the IRS limit, and it is worth avoiding. Deferring more than $24,500 in 2026 creates an excess deferral that must be withdrawn by the following April 15 or it is taxed twice — once in the year contributed and again when distributed. This happens most often to people who change jobs mid-year, because each employer tracks only its own plan and neither sees your combined total. If you switched employers, add the deferrals from both W-2s and check. There is also a softer version of contributing 'too much': funding a 401(k) heavily while carrying 22% credit card debt or holding no emergency savings is legal, but usually not the best use of the dollar.

See What a Contribution Change Costs You

Enter your salary, filing status, state, and 401(k) percentage. The calculator shows exactly how much take-home pay each contribution rate actually costs — usually less than people expect. Free and instant.

Open Take-Home Pay Calculator →

What Kind of Pages AI Assistants Actually Cite

Yonatan Naor — Thu, 30 Jul 2026 06:03:23 +0000

Originally published at https://thicket.sh/blog/what-kind-of-pages-ai-assistants-cite.

← Back to Journal
Data

We run 21 sites that Microsoft Copilot cites about 14,394 times a month. We classified the 25 most-cited pages by shape to see what they have in common — and found something uncomfortable about what all those citations were worth.

By Thicket AI Team · July 30, 2026 · 8 min read

Four page shapes account for 80% of our most-cited pages: head-to-head comparisons, definitions, reference tables, and trend explainers — each exactly 20% of the top 25. What they share is that each answers a question with a specific, checkable, quotable answer. Assistants cite what they can lift a sentence out of. And the uncomfortable part: across the same period those 14,394 monthly citations produced zero new referring domains, for a reason that turns out to be structural rather than fixable by volume.

Most writing about AI citation is inference from the outside. This is from the inside, on a portfolio large enough to have a distribution — with the method published separately so you can reproduce it on your own domain.

The Classification

We took the 25 most-cited pages from Bing's AI Performance report and classified each by shape. Counts, and total citations attributable to each shape:

ShapePagesShareCitationsTrend / event explainer520%2,796Comparison (X vs Y)520%1,952Definition / what-is520%1,853Reference data / table520%1,695Direct question page28%753Listicle / best-of14%666Other28%674
The even split across the top four is more striking than any single winner. It suggests the shape that matters is not a specific format but a property those formats happen to share.

The Shared Property

Every one of those shapes answers a question that has a specific, checkable answer that fits in a sentence.

A comparison resolves to this one, because of that.
A definition resolves to it means this.
A reference table resolves to for your case, the number is X.
A trend explainer resolves to here is what it is and why now. An assistant assembling an answer needs a passage it can lift and attribute. Pages written to be comprehensive rather than answerable do worse, because there is no single passage that resolves the question — the answer is distributed across 3,000 words and the model has nothing crisp to quote.

This matches what we changed and saw work: put the direct answer in the first line in bold, keep a primary source next to each factual claim, and include one table. It is close to the opposite of writing for a word count.

The Part That Should Give You Pause

We accumulated 14,394 citations per 28 days. Over the same period we earned zero new referring domains. Not few. Zero.

The grounding-query table explains it. Our top 25 queries by citation volume, every one of them:

tuscan mom aesthetic · salary overtime law · is myspace still active · federal overtime laws · best AI tools for developer productivity · office siren · employee tax withholding update timing · when to update W4 · w2 example · fastest VPN speeds · cold plunge vs sauna · protein requirements toning vs bulking · fica vs federal tax

These are consumer questions, all of them. Not one is a practitioner question — nobody reaching us asked how to measure AI citations, how to run content operations, or which tool to use.

Which produces the mechanism, and it is not a distribution problem:

We are cited to people who ask questions. Links come from people who publish answers. Consumers do not have websites.

Someone asking whether MySpace is still active gets a good answer partly built from our page, and is satisfied. They have no blog, no newsletter, no editorial calendar. There is no path from that citation to a link however many times it happens — and scaling it tenfold yields tenfold of exactly nothing.

What We Take From It

Citations and links are different products and need different content. Consumer-question content earns citations and traffic. Only practitioner-question content can earn links, because only practitioners publish.
Do not judge AI visibility on analytics sessions. 14,394 citations produced 126 assistant-referred sessions — about a hundred citations per visit, because a good citation answers the question inside the chat.
Do not judge commercial value on citations either. Same numbers, opposite error. They measure attention, not intent and not authority.
Citations appear to behave like a stock. Our most-cited property has published nothing for a month and still leads the portfolio. Rankings decay when you stop; citations so far have not. We are re-measuring that at 60 and 90 days before believing it. ## Frequently Asked Questions

What kind of content do AI assistants cite most?

In our data, four shapes account for 80% of the most-cited pages: head-to-head comparisons (X vs Y), definitions of a term someone just encountered, reference tables of values by category, and explainers of a specific trend or event. Each accounts for 20% of the top 25. What they share is that they answer a question with a specific, checkable, quotable answer — a number, a definition, a direct comparison. Assistants cite what they can lift a sentence out of. Pages built to be comprehensive rather than answerable perform worse, because there is no single passage to quote.

Does publishing more content increase AI citations?

Not reliably, in our experience. We tested volume directly and rejected it — more pages of the same shape did not produce proportionally more citations. What moved the number was writing pages that answer a specific question with a liftable answer, and citing a primary source next to each claim. We also found citations behave more like a stock than a flow: our most-cited property has published nothing for a month and still holds the largest block of citations in the portfolio. Rankings decay when you stop publishing; citations so far have not.

Do AI citations bring traffic?

Some, but far less than the citation count suggests, because a good citation often answers the question inside the chat and produces no click at all. Over an identical 28-day window our portfolio recorded 14,394 citations and 126 assistant-referred sessions — roughly a hundred citations per visit. Both numbers are real and they measure different things. Judging AI visibility by analytics sessions alone understates it by about two orders of magnitude; judging commercial value by citations alone overstates it by about the same.

Do AI citations produce backlinks?

In our data, no — and this was the most uncomfortable thing we found. Over the period we accumulated 14,394 citations per 28 days, new referring domains stayed at zero. The grounding-query table explains why: every one of our top 25 queries is a consumer question, and consumers do not have websites. Links come from people who publish. If your citations are earned on questions asked by readers rather than by writers, scaling that volume produces more of the same nothing.

How do you find out which queries you are cited on?

Bing Webmaster Tools has an AI Performance report showing citations by Copilot and partner models, broken out by grounding query and by page. It is the only first-party citation data any major platform publishes, and there is no API — we checked every plausible endpoint and got 404s, so it has to be read from the interface. Verify your root domain rather than each subdomain and one read covers a whole portfolio. It says nothing about ChatGPT, Claude, Gemini or Perplexity, so treat it as a sample of one ecosystem rather than a total.

Method and Caveats

Data is from Bing Webmaster Tools' AI Performance report (Microsoft Copilot and partner models) for a 21-subdomain portfolio, 3-month window, read 2026-07-26, plus GA4 referrer segmentation and Google Search Console over the 28 days ending 2026-07-24. The full method, including what each measurement misses, is in how to measure whether AI assistants cite your site.

Honest limits. This is one portfolio in one ecosystem — Bing reports only Microsoft-family citations and says nothing about ChatGPT, Claude, Gemini or Perplexity, and Bing itself describes its AI Performance data as a sample. The shape classification is our own judgement applied to 25 pages; a different taxonomy would produce different buckets, and 25 is a small n. Our portfolio skews to calculators, consumer finance, fitness and trend explainers, which plausibly biases which shapes appear at all. Treat the direction as informative and the percentages as indicative. Our own numbers, including the experiments that failed, are in the Thicket Report.

More from the Journal

How Do ETFs Work? A Plain-English Explanation

Yonatan Naor — Wed, 29 Jul 2026 11:04:08 +0000

Originally published at https://etf.thicket.sh/blog/how-do-etfs-work.

An ETF is a fund that owns a basket of assets — often hundreds or thousands of stocks or bonds — and sells shares in itself that trade on a stock exchange like any ordinary stock. Buy one share and you own a proportional slice of everything inside. Two things make it behave differently from a traditional mutual fund: it trades continuously through the day at a market price rather than once daily at a computed value, and it uses a creation-and-redemption mechanism with large institutions that keeps the market price tethered to the value of the holdings — the same mechanism that gives ETFs their tax efficiency.

Most explanations stop at “a basket that trades like a stock,” which leaves the interesting question unanswered: if the share price is set by supply and demand on an exchange, what stops it drifting away from what the fund actually owns? The answer is the part worth understanding.

1. The Basket

The fund holds real assets. A total-market equity ETF owns shares in thousands of companies; a bond ETF owns thousands of individual bonds; a gold ETF owns bullion. Divide the value of everything it owns, minus liabilities, by the number of shares outstanding, and you get net asset value — the honest per-share worth of the fund at that moment.

Most ETFs track an index, meaning the basket is chosen by a published rule rather than by a manager's judgement. That is why costs are low: following a rule is cheap. For how a fund's ongoing cost is deducted, see how an ETF expense ratio is charged.

2. The Exchange Listing

Shares of the fund are listed on an exchange and trade all day at whatever price buyers and sellers agree on. You can buy at 10:04am and sell at 2:15pm. A mutual fund, by contrast, transacts once per day at the NAV computed after the close.

An important consequence: when you buy an ETF share, your money usually goes to another investor, not to the fund. You are trading on the secondary market, and the fund's holdings are untouched. That is structurally different from a mutual fund, where your money flows in and the manager deploys it.

3. The Arbitrage Loop That Keeps the Price Honest

This is the mechanism that makes the whole thing work, and it involves institutions called authorized participants.

SituationWhat the authorized participant doesEffect on priceETF trades above NAVBuys the underlying basket, delivers it to the fund for newly created ETF shares, sells those shares at the higher market priceNew supply pushes the price down toward NAVETF trades below NAVBuys cheap ETF shares on the market, redeems them with the fund for the underlying basket, sells the basketDemand pushes the price up toward NAV
Both trades are profitable for the participant, which is the point: nobody enforces the tracking — the profit motive does it, continuously. It is why a large index ETF trades within pennies of its underlying value nearly all the time, and why the mechanism strains in thinly traded funds or during severe market stress, when assembling or valuing the basket becomes harder.

4. Why the Tax Treatment Is Different

Notice that redemptions are settled in kind — the fund hands over securities rather than selling them for cash. That detail produces the tax advantage.

When a mutual fund investor sells, the fund often must sell holdings to raise cash, and the resulting capital gain is distributed to everyone still in the fund, including people who did nothing. An ETF's in-kind exchange is generally not a taxable sale for the fund, so broad index ETFs frequently distribute no capital gains at all. Your tax bill stays largely under your control — you owe when you choose to sell. More in how ETFs are taxed and ETF vs mutual funds.

5. How It Actually Pays You

Price appreciation. Holdings rise, NAV rises, share price follows; you realise it on sale.
Distributions. The fund collects dividends and interest, subtracts expenses, and passes the rest through — usually quarterly for stock ETFs and monthly for bond ETFs. See do ETFs pay dividends. A fund holding assets that generate no income — physical gold, say — offers only the first.

6. What It Costs, and What Can Go Wrong

The expense ratio is the headline cost, commonly 0.02%–0.09% for broad index funds, accrued daily inside NAV so it never appears as a charge. Sitting outside it: the bid-ask spread when you trade, any premium or discount to NAV, and the fund's own internal trading costs.

On risk, the honest framing is that “ETF” is a wrapper, not a risk level. A total-market index fund and a leveraged single-sector product are both ETFs and are not comparable. The dominant risk is simply that the assets inside can fall. Structure-specific risks are narrower: wider spreads and larger NAV deviations in thin or hard-to-value funds, especially under stress, and leveraged and inverse products that reset daily and are not built for long holds. A fund can also close — which returns your money rather than losing it, as covered in what happens when an ETF closes.

Sources

Fund structure, exchange trading, creation and redemption, and required prospectus disclosure: U.S. Securities and Exchange Commission, Mutual Funds and ETFs and Investor.gov on mutual funds and ETFs. Net asset value: Investor.gov, Net Asset Value. Typical expense-ratio ranges reflect fees observed across large US-listed ETFs and change over time.

Caveats

This describes US-regulated ETFs and is general information, not investment advice. Any specific fund's holdings, costs, distribution schedule, and risks are set out in its prospectus on the issuer's own site, which is the authoritative source when a third-party data provider disagrees.

Frequently Asked Questions

How do ETFs work?An ETF is a fund that owns a basket of assets — usually hundreds or thousands of stocks or bonds — and sells shares in itself that trade on a stock exchange like any single stock. Buy one share and you own a proportional slice of everything the fund holds. Two features make it different from a traditional mutual fund. It trades continuously through the day at a market price, rather than once daily at a computed value. And it uses a creation-and-redemption mechanism with large institutional firms to keep that market price tethered to the value of the underlying holdings, which also gives ETFs their well-known tax efficiency.What keeps an ETF's price close to the value of what it holds?An arbitrage loop run by institutions called authorized participants. Every ETF has a net asset value — the per-share value of everything it owns. If the ETF's market price drifts above NAV, an authorized participant can assemble the underlying basket of securities, deliver it to the fund in exchange for newly created ETF shares, and sell those shares at the higher market price. If the price falls below NAV, the reverse: buy cheap ETF shares, redeem them with the fund for the underlying basket, sell the basket. Both trades profit the participant and push the price back toward NAV. Nobody enforces the tracking; the profit motive does it continuously.Where does an ETF's money go when you buy a share?Usually to another investor, not to the fund. The overwhelming majority of ETF trades happen on the secondary market — you buy from whoever is selling, the same as with a stock, and the fund's assets are unaffected. New shares are only created when an authorized participant delivers a basket of securities to the fund, which happens in large blocks and typically only when demand has pushed the price away from NAV. This is a genuine structural difference from a mutual fund, where your purchase money flows into the fund and the manager buys securities with it.How do ETFs make money for you?Two ways, and they mirror owning the underlying assets directly. Price appreciation: if the securities the fund holds rise in value, NAV rises and so does the share price, and you realise that gain when you sell. And distributions: the fund collects the dividends its stocks pay and the interest its bonds pay, subtracts its expenses, and passes the rest to shareholders — typically quarterly for stock ETFs and monthly for bond ETFs. A fund holding assets that produce no income, such as a physical-gold ETF, offers only the first.Why are ETFs more tax-efficient than mutual funds?Because of how redemptions work. When a mutual fund investor sells, the fund often has to sell securities to raise cash, and any capital gain from that sale is distributed to everyone still holding the fund — including people who did nothing. An ETF instead hands a basket of securities to the authorized participant in an in-kind exchange, which is generally not a taxable sale for the fund. The practical result is that broad index ETFs frequently distribute no capital gains at all, so your tax bill is largely under your own control: you owe when you choose to sell.What does an ETF actually cost to own?The headline cost is the expense ratio, quoted annually but accrued daily out of fund assets before the share price is calculated, so it never appears as a charge on your statement. Broad index ETFs commonly run 0.02% to 0.09%. Beyond that sit costs the ratio excludes: the bid-ask spread you pay when trading, any premium or discount to NAV, and the fund's own internal trading costs. For a large, liquid fund held long term the expense ratio dominates. For a thin or exotic one, the spread can easily exceed a year of expense ratio in a single round trip.Are ETFs safe, and what can go wrong?The structure itself is well-regulated and durable, but 'ETF' describes a wrapper, not a risk level — a total-market index fund and a leveraged single-sector product are both ETFs and are not remotely comparable. The real risks are the ordinary ones: the assets inside can fall in value, and you carry that fully. Structure-specific issues are narrower — wider spreads and larger NAV deviations in thinly traded or hard-to-value funds, particularly during market stress, and the fact that leveraged and inverse products reset daily and are not designed to be held long term. A fund can also close, though that returns your money rather than losing it.how ETFs workNAVcreation and redemptionauthorized participantin-kindtax efficiencyindex fundsSEC← More fund analysis

How Long Does It Take to Build Muscle? (Realistic Timeline)

Yonatan Naor — Wed, 29 Jul 2026 11:04:06 +0000

Originally published at https://fit.thicket.sh/blog/how-long-does-it-take-to-build-muscle.

By Dr. Sarah Okafor, Exercise Physiologist

Measurable muscle growth takes about 8 to 12 weeks of consistent training. Change other people notice usually takes 3 to 6 months. The catch is that the first few weeks feel like the fastest progress you will ever make and are mostly not muscle: early strength gains come from your nervous system learning the movement, and early size changes are largely swelling from muscle damage plus water and glycogen drawn into the tissue. Real hypertrophy starts accumulating underneath that, and only becomes measurable once the early swelling settles.

That gap between how progress feels and what is actually happening is why so many people quit around week four, exactly when the real process is beginning. Below is the honest timeline, what drives the rate, and why the mirror lies early on.

The First Three Weeks Are Not Muscle

Damas and colleagues tracked muscle protein synthesis alongside actual hypertrophy through the first weeks of training and found that early synthesis is directed at repairing damage, not building new tissue — the two only start corresponding once the damage response attenuates (Journal of Physiology 2016, PMID 27219125). Studies measuring muscle thickness very early therefore tend to overstate growth: the muscle is swollen, not bigger.

Even research designed to catch hypertrophy as early as possible finds it takes weeks of high-intensity work before architectural change is detectable (Seynnes and colleagues, PMID 17053104). Meanwhile your strength climbs quickly, because you are getting better at recruiting and coordinating the muscle you already have. Strength and size are related, but early on they move for different reasons.

A Realistic Timeline

TimeframeWhat is actually happeningWhat you noticeWeeks 1–3Neural adaptation; damage repair; water and glycogenStrength climbing fast, muscles feel fuller, sorenessWeeks 4–8Hypertrophy begins accumulating; swelling subsidesOften feels like a stall — the discouraging phaseWeeks 8–12Measurable growthClothes fit differently; tape and scale moveMonths 3–6Growth compoundsOther people start commentingYear 1+Rate slows substantiallyProgress requires deliberate programming

How Much, Per Month

A well-trained, well-fed beginner can realistically add roughly 1–2 lb of muscle per month, with women averaging around half that, and the rate falls considerably after the first year. A widely used rough guide puts lifetime natural gain for men at roughly 20–25 lb in year one, about half that in year two, half again in year three, and a few pounds a year thereafter.

Those monthly and yearly figures are practitioner estimates that have held up well in practice, not outputs of a controlled trial. Use them to calibrate expectations, not as a prediction for any individual.

Get weekly fitness tips in your inbox

What Sets the Pace

Training age. The single biggest factor. Beginners gain several times faster than experienced lifters, and nothing recovers that advantage once it is spent.
Progressive overload. The stimulus has to keep increasing. See progressive overload for beginners vs intermediates.
Weekly volume. There is a clear dose-response — more weekly sets produce more growth, up to a point (Schoenfeld 2017, PMID 27433992).
Protein and calories. Protein supplementation meaningfully augments resistance-training gains (Morton 2018, PMID 28698222). Targets in how much protein you need to build muscle.
Sleep and recovery. Muscle is built during rest, not during the session — see how many rest days you need.
Consistency. Twelve merely-adequate weeks beat four excellent weeks followed by eight missed ones. The adaptation is cumulative. ## Two Things That Distort the Timeline

Fat loss makes muscle visible faster than muscle growth does. Someone with more body fat can look substantially different at 12 weeks having added only a few pounds of tissue, because the change is mostly the layer over the muscle thinning. If being noticed is the goal, body composition beats mass — see body recomposition timelines.

Soreness is not a progress meter. It reflects unaccustomed stress, and it fades as you adapt — which is a sign of adaptation, not of stalling. See does soreness mean growth.

Sources and Methodology

Early-phase protein synthesis and damage: Damas F et al., Journal of Physiology 2016 (PMID 27219125). Early hypertrophy detectability: Seynnes OR et al., Journal of Applied Physiology 2007 (PMID 17053104). Volume dose-response: Schoenfeld BJ et al., Journal of Sports Sciences 2017 (PMID 27433992). Protein supplementation: Morton RW et al., British Journal of Sports Medicine 2018 (PMID 28698222). Monthly and annual gain figures are practitioner estimates, flagged as such above. General information for healthy adults, not medical advice. Last updated July 29, 2026.

Frequently Asked Questions

Measurable muscle growth typically takes 8 to 12 weeks of consistent training, and visible change to others usually takes 3 to 6 months. The first few weeks feel dramatic but are mostly not muscle: early strength gains come from the nervous system learning the movement, and early size changes are substantially swelling from muscle damage rather than new tissue. Damas and colleagues showed that muscle protein synthesis in the first three weeks tracks damage repair, and only begins to correspond with real hypertrophy once that damage subsides (PMID 27219125). Genuine, measurable growth is a 2-to-3-month project at minimum.A realistic ceiling for a well-trained, well-fed beginner is roughly 1 to 2 pounds of muscle per month for men and about half that for women, and it slows considerably after the first year. A common rough guide is around 20 to 25 pounds of lifetime natural muscle gain for men in year one, halving in year two, halving again in year three, and then measured in a few pounds a year. These figures are practitioner estimates rather than precise trial outputs, so treat them as a sanity check on expectations, not a prediction. The direction is what matters: the newer you are, the faster it goes.Three things that are not new muscle tissue. Training draws water and glycogen into the muscle, which increases its volume. Unaccustomed training causes swelling from muscle damage, which temporarily inflates measurements. And the pump during and after a session is transient blood flow. All are real and visible, and none is hypertrophy. This is precisely why early-phase studies that measure muscle thickness in the first weeks tend to overstate growth — the tissue is swollen, not larger. It also explains the demoralising moment around week four when the initial change seems to stall: the swelling has settled and actual growth has not yet caught up.Usually 3 to 6 months of consistent training, and often sooner if you are also losing fat, because reducing the layer over the muscle reveals shape faster than adding tissue does. People carrying more body fat frequently look meaningfully different at 12 weeks while having gained only a few pounds of actual muscle, since the visual change is driven mostly by the fat loss. Conversely, a lean beginner can add real muscle for months before anyone comments. If being noticed is the goal, body composition matters more than muscle mass alone.It slows the process but does not stop it. Older adults build muscle more slowly than younger adults, largely because of anabolic resistance — the muscle responds less strongly to the same protein and training stimulus. The practical adjustments are a somewhat higher protein intake, an emphasis on progressive resistance training rather than only cardio, and more patience with the timeline. Meaningful hypertrophy is well documented into the seventies and beyond. Starting at 50 or 60 means a slower curve, not an absent one.In rough order of impact: training age (beginners gain fastest by a wide margin), progressive overload, total weekly volume, protein and calorie intake, sleep, and genetics. Volume has a clear dose-response — more weekly sets produce more growth up to a point (Schoenfeld 2017, PMID 27433992) — and protein supplementation meaningfully augments gains from resistance training (Morton 2018, PMID 28698222). Consistency dominates all of it. Twelve mediocre weeks beat four excellent weeks followed by eight missed ones, because the adaptation you are chasing is cumulative.Legs follow the same 8-to-12-week timeline for measurable change, but they often look slower for two reasons. They start from a larger base — you walk on them daily — so a given amount of new tissue is a smaller percentage change and less visually obvious. And many people train legs less hard than upper body because it is more uncomfortable. Quadriceps in particular respond well to sufficient volume and full range of motion. The timeline is not different; the perception is.

Set the Calorie and Protein Targets First

The timeline above assumes you are actually eating for it. Enter your stats to get evidence-based maintenance calories and a daily protein target in seconds. Free and instant.

Open TDEE Calculator →

Do I Have to Pay Taxes on Social Security? (Combined Income Explained)

Yonatan Naor — Tue, 28 Jul 2026 04:41:55 +0000

Originally published at https://pay.thicket.sh/blog/do-i-have-to-pay-taxes-on-social-security.

By Jamie Reeves, Personal Finance Writer

Only if your other income is high enough. The test uses “combined income” — your adjusted gross income, plus any tax-exempt interest, plus half of your Social Security benefits. Below $25,000 filing single, or $32,000 married filing jointly, none of your benefits are taxable. Between $25,000 and $34,000 single ($32,000–$44,000 joint), up to 50% of benefits can be taxable. Above $34,000 single ($44,000 joint), up to 85% can be. Crucially, “85% taxable” means 85% of your benefit is added to your taxable income — it is not an 85% tax rate, and at least 15% of your benefits are always tax-free.

Two things make this question harder than it should be: the calculation uses a figure called combined income that appears almost nowhere else in the tax code, and the widely-quoted “85%” sounds like a rate when it is a share. Both are cleared up below, along with the age myth.

Step 1: Work Out Your Combined Income

Combined income is a specific formula:

Your adjusted gross income (wages, pensions, IRA and 401(k) withdrawals, interest, dividends, capital gains)
plus tax-exempt interest — municipal bond interest counts here, even though it is otherwise tax-free
plus one-half of your Social Security benefits That half-benefit detail matters. People often add their whole benefit, land above a threshold, and assume they owe tax when they do not. Only half of the benefit enters the test.

Step 2: Compare It to the Thresholds

Filing statusNone taxableUp to 50% taxableUp to 85% taxableSingle / head of household / qualifying surviving spouseUnder $25,000$25,000 – $34,000Over $34,000Married filing jointlyUnder $32,000$32,000 – $44,000Over $44,000Married filing separately, lived apart all yearUnder $25,000$25,000 – $34,000Over $34,000Married filing separately, lived together at any pointBase amount is $0 — up to 85% taxable from the first dollar
Base amounts per IRS Publication 915. These thresholds are not indexed to inflation and have not changed since the 1980s and 1990s — which is why a steadily growing share of retirees crosses them each year.

The 85% Misunderstanding

“Up to 85% of your benefits are taxable” is a statement about how much of the benefit enters your taxable income. It is not a tax rate. Worked through:

You receive $30,000 in Social Security for the year.
Your combined income puts you in the top band, so 85% of the benefit — $25,500 — is added to taxable income.
That $25,500 is taxed at your ordinary rate. In the 12% bracket, that is about $3,060.
$3,060 on a $30,000 benefit is an effective rate of about 10% — not 85%. And the 85% is a ceiling. No matter how high your income goes, at least 15% of your Social Security benefits are never taxed.

The Age Myth

There is no age at which Social Security benefits stop being taxable.Not 65, not 70, not full retirement age. Taxability is determined by combined income and nothing else. A 90-year-old with a large required minimum distribution pays tax on benefits; a 62-year-old with modest other income pays none.

What genuinely changes with age is how much other income you have. Required minimum distributions from traditional retirement accounts push combined income up, which is why some retirees find benefits become taxable years after they started collecting, with nothing about the benefit itself having changed.

Levers That Actually Reduce the Bill

Every lever works by moving combined income, because that is what the test reads:

Roth withdrawals do not count toward combined income. Building Roth balances before claiming is the cleanest structural fix — see our Roth vs traditional 401(k) comparison.
Time large withdrawals and capital gains into years before you claim, or into years you can absorb them.
Qualified charitable distributions from an IRA after age 70½ satisfy required minimum distributions without raising adjusted gross income.
Municipal bonds do not help here. Tax-exempt interest is explicitly added back into combined income. ## Paying the Tax

If you expect to owe, you can have federal tax withheld directly from your benefit by filing Form W-4V with the Social Security Administration, choosing 7%, 10%, 12%, or 22%. The alternative is quarterly estimated payments, which work but mean four deadlines a year to remember. For how withholding works on earned income while you are still working, see when to update your W-4 and how much tax comes out of your paycheck. If you are still working while collecting, note that your wages remain subject to FICA regardless — see FICA vs federal income tax.

Sources and Methodology

Base amounts, the combined-income definition, and the 50%/85% inclusion rules: IRS Publication 915 (Social Security and Equivalent Railroad Retirement Benefits) and IRS Topic 423. Voluntary withholding from benefits: IRS Form W-4V. The worked example uses a 12% ordinary rate for illustration; your own rate depends on total taxable income. State treatment varies and is changing quickly — verify your state's current rules. General information, not tax advice. Last updated July 28, 2026.

Frequently Asked Questions

Only if your other income is high enough. The test uses 'combined income' — your adjusted gross income, plus any tax-exempt interest, plus half of your Social Security benefits. If that total is under $25,000 filing single (or $32,000 married filing jointly), none of your benefits are taxable. Between $25,000 and $34,000 single ($32,000–$44,000 joint), up to 50% of benefits can be taxable. Above $34,000 single ($44,000 joint), up to 85% can be taxable. Roughly 40% of recipients pay some tax on their benefits; the rest pay none.It is a specific figure invented for this test and used almost nowhere else, which is why it confuses people. Combined income equals your adjusted gross income, plus any tax-exempt interest (yes, municipal bond interest counts here even though it is otherwise tax-free), plus one-half of the Social Security benefits you received that year. Note that only half your benefits go into the calculation — a detail that trips up people who assume the whole payment counts and conclude they owe tax when they do not.No, and this is the single most common misunderstanding. The 85% figure is the portion of your benefits that becomes part of your taxable income — it is not a tax rate. If you receive $30,000 in benefits and 85% is taxable, then $25,500 gets added to your taxable income and is taxed at your ordinary rate. Someone in the 12% bracket would owe about $3,060 on that, roughly 10% of the benefit, not 85% of it. The maximum share of benefits that can ever be taxable is 85%; at least 15% of your benefits are always tax-free.There is no such age. This is a persistent myth — you may have heard 65, 70, or 'after full retirement age.' None of them are real. The taxability of benefits depends entirely on your combined income, not on how old you are. A 90-year-old with substantial retirement income pays tax on benefits; a 62-year-old with little other income pays none. What does change with age is the amount of other income you tend to have, which is why some people's benefits become taxable later in retirement when required minimum distributions begin.Usually not. The large majority of states do not tax Social Security benefits at all, and several states have no income tax whatsoever. A small and shrinking number of states do tax benefits, typically with their own income exemptions and age-based carve-outs that are often more generous than the federal rules. Because states have been steadily repealing these taxes in recent years, check your own state's current rules rather than relying on an older list — this is one of the fastest-changing corners of retirement tax.Because the test keys off combined income, the levers are all about managing that number. Roth withdrawals do not count toward combined income, so building Roth balances before claiming can keep you under a threshold. Timing large withdrawals or capital gains into years when you are not yet collecting benefits helps. Qualified charitable distributions from an IRA after age 70½ satisfy required minimum distributions without adding to adjusted gross income. And remember tax-exempt municipal bond interest does count here, so it does not help with this particular test. Coordinating withdrawals across account types is genuinely worth professional advice — the thresholds are not indexed to inflation, so more retirees cross them every year.If you expect to owe, yes — it is simpler than quarterly estimated payments. You can request federal withholding from benefits by filing Form W-4V with the Social Security Administration, choosing 7%, 10%, 12%, or 22% of your benefit. The alternative is paying quarterly estimated tax, which works but requires you to remember four deadlines a year. Many retirees find withholding less error-prone, particularly in the first year of benefits when the tax consequence is unfamiliar.

See Your Take-Home Pay

Still working, or working while collecting benefits? Enter your salary, filing status, and state to see exactly what lands in your account each pay period. Free and instant.

Open Take-Home Pay Calculator →

Does Creatine Cause Hair Loss? What the Evidence Actually Shows

Yonatan Naor — Tue, 28 Jul 2026 04:41:51 +0000

Originally published at https://fit.thicket.sh/blog/does-creatine-cause-hair-loss.

By Dr. Sarah Okafor, Exercise Physiologist

There is no direct evidence that creatine causes hair loss. No study has ever measured hair as an outcome and found that creatine caused it to fall out. The entire belief traces to a single 2009 trial of 20 college rugby players that reported a rise in the DHT-to-testosterone ratio — a hormone marker — and did not look at hair at all. That finding has never been replicated in the fifteen-plus years since, despite creatine being one of the most studied supplements in existence. If you are genetically predisposed to male-pattern baldness you may lose hair whether or not you supplement; the evidence does not say creatine is the reason.

This is a good case study in how one small result becomes received wisdom. Below is exactly what that study did and did not show, why the leap from “DHT went up” to “you will go bald” skips a step that matters, and what creatine’s genuinely documented side effects are.

The One Study Behind the Whole Belief

In 2009, researchers ran a three-week crossover trial in 20 college-aged rugby players and measured hormone levels (van der Merwe and colleagues, Clinical Journal of Sport Medicine, PMID 19741313). After seven days of loading at 25 g/day, the ratio of dihydrotestosterone (DHT) to testosterone rose roughly 36%, and stayed about 22% above baseline through two weeks of 5 g/day maintenance.

Four details rarely travel with that headline:

Hair was never measured. Not counted, not photographed, not assessed. The study was about hormones.
DHT stayed within the normal reference range the whole time. The ratio shifted; the values did not become abnormal.
Testosterone itself did not change significantly. The ratio moved largely because DHT rose.
n = 20, for three weeks. Male-pattern baldness develops over years. ## It Has Never Been Replicated

This is the part that should settle it. Creatine is among the most heavily researched supplements ever sold, with hundreds of trials and decades of safety follow-up. In all of that, the 2009 DHT result stands alone — no successful replication in more than fifteen years.

A review of common creatine questions and misconceptions examined the claim directly and concluded the evidence does not support a link between creatine supplementation and hair loss (Antonio and colleagues, 2021, Journal of the International Society of Sports Nutrition, PMID 33557850). The ISSN’s position stand on creatine safety and efficacy reviewed the literature and did not identify hair loss as an established adverse effect (Kreider and colleagues, 2017, PMID 28615996).

Get weekly fitness tips in your inbox

The Step the Myth Skips

DHT really is central to male-pattern hair loss, which is why this myth is sticky — it sounds mechanistically sensible. But the mechanism runs through inherited follicle sensitivity to DHT, not through circulating DHT alone. Men with the genetic predisposition lose hair at perfectly ordinary DHT levels. Men without it keep their hair at high ones.

So even granting the 2009 result exactly as published, the chain has a missing link:

Step in the claimEvidence statusCreatine raised the DHT:T ratio in 20 rugby players over 3 weeksObserved once, 2009 — never replicatedThose DHT levels were abnormal*No* — stayed within the normal rangeA within-range DHT shift miniaturises hair follicles*Not demonstratedCreatine users lose more hair than non-usersNever measured in any study*
Two of those four steps have no supporting evidence at all, and one is contradicted by the source study itself.

What Creatine’s Real Side Effects Are

Water retention — real, and misunderstood. Creatine pulls water into muscle cells, typically showing up as one to three pounds on the scale in the first week. It is intracellular, not the puffy subcutaneous bloat people imagine.
GI discomfort — occasional, mostly avoidable. Usually tied to large single doses. Skipping the loading phase and taking 3–5 g daily largely prevents it — see loading vs maintenance.
Kidney damage — not supported. Creatine raises serum creatinine, which is a marker used to estimate kidney function, without impairing the kidney. The same misreading pattern as the hair-loss claim. Compare with the protein-and-kidneys myth, which has an almost identical structure. For the efficacy side and the safety literature in more depth, see does creatine actually work (and is it safe), and on form and timing, creatine HCl vs monohydrate and whether creatine timing matters.

If You Are Actually Losing Hair

Worth separating the question from the supplement. Male-pattern baldness is common, progressive, and genuinely treatable — and the treatments work considerably better the earlier they are started. If you are seeing real thinning, a dermatologist is the productive move. Dropping creatine on the strength of one unreplicated hormone study is unlikely to change the outcome, and it costs you a supplement with unusually strong evidence behind it.

Sources and Methodology

The DHT finding: van der Merwe J, Brooks NE, Myburgh KH, “Three weeks of creatine monohydrate supplementation affects dihydrotestosterone to testosterone ratio in college-aged rugby players,” Clinical Journal of Sport Medicine 2009 (PMID 19741313). Misconception review: Antonio J et al., JISSN 2021 (PMID 33557850). Safety position stand: Kreider RB et al., JISSN 2017 (PMID 28615996). Creatine in women: Smith-Ryan AE et al., Nutrients 2021 (PMID 33800439). General information for healthy adults, not medical advice; if you are losing hair or managing a health condition, speak to a clinician. Last updated July 28, 2026.

Frequently Asked Questions

There is no direct evidence that it does. No study has ever measured hair loss as an outcome and found creatine caused it. The entire concern traces to a single 2009 trial of 20 college rugby players which reported a rise in the DHT-to-testosterone ratio after three weeks of supplementation (van der Merwe 2009, PMID 19741313) — it did not measure hair at all. That result has not been replicated in the years since, and a review of common creatine misconceptions concluded the evidence does not support a link between creatine and hair loss (Antonio 2021, PMID 33557850). If you are genetically predisposed to male-pattern baldness you may lose hair whether or not you take creatine; the evidence does not say creatine is what caused it.It enrolled 20 college-aged rugby players in a three-week crossover and measured hormones, not hair. After seven days of loading at 25 g/day, the DHT-to-testosterone ratio rose about 36% and remained roughly 22% above baseline through two weeks of 5 g/day maintenance. Importantly, the DHT values stayed within the normal clinical reference range throughout, testosterone itself did not change significantly, and no hair-related outcome was recorded. It is a small, single, short study of a hormonal marker — a reasonable prompt for further research, which is exactly what it has not received.No. That is the crux of the matter. The 2009 trial has stood alone for over fifteen years without a successful replication, despite creatine being among the most heavily studied supplements in existence, with hundreds of trials and decades of safety data. In science, a single unreplicated result in 20 people is a hypothesis, not a conclusion. The International Society of Sports Nutrition's position stand on creatine safety reviewed the body of evidence and did not identify hair loss as an established adverse effect (Kreider 2017, PMID 28615996).DHT is genuinely central to male-pattern hair loss, which is why the myth is plausible enough to persist. But the mechanism depends on inherited sensitivity of the hair follicle to DHT, not simply on how much DHT is circulating. People with the genetic predisposition lose hair at ordinary DHT levels; people without it keep their hair at high ones. So even taking the 2009 ratio change at face value — a change that stayed inside the normal range — it does not follow that hair loss results. The causal step everyone assumes, from a within-range hormone shift to actual follicle miniaturisation, is the step no study has demonstrated.That is a personal call, and worth making with the actual odds in view. Creatine monohydrate has the strongest efficacy and safety evidence of any sports supplement, with well-documented benefits for strength, muscle mass, and increasingly for cognition and healthy ageing. Against that sits one unreplicated hormonal marker study that never measured hair. If you are noticing genuine thinning, the far more productive step is to see a dermatologist — male-pattern baldness is treatable and the treatments work better the earlier they start. Stopping creatine is unlikely to be the intervention that helps.There is no evidence for it, and even less mechanistic reason to expect it. The DHT hypothesis is built on male-pattern androgenic alopecia; female-pattern hair loss has a different and less androgen-driven pathophysiology. A review of creatine in women's health across the lifespan found the supplement generally safe and potentially beneficial, particularly around menopause, without flagging hair loss as a concern (Smith-Ryan 2021, PMID 33800439). Women are also frequently under-dosed relative to the evidence — the standard 3 to 5 g/day applies regardless of sex.The genuinely documented ones are mild and mostly manageable. Some short-term water retention is normal and expected — creatine draws water into muscle cells, which typically shows up as one to three pounds of scale weight in the first week and is intracellular, not the subcutaneous bloating people picture. Gastrointestinal discomfort can occur, usually at high single doses, and is largely avoidable by skipping the loading phase and taking 3 to 5 g daily instead. The kidney-damage claim, like the hair-loss one, does not survive contact with the evidence: creatine raises serum creatinine, a marker used to estimate kidney function, without impairing the kidney itself.

Get Your Training Numbers Right

Supplements are the small lever. Calorie and protein targets are the large one. Enter your stats and get evidence-based daily numbers in seconds — free and instant.

Open TDEE Calculator →