Method
How a number gets printed
Every number on this site names the assistant that answered, how many answers it rests on, its range and the date. This page says what those words mean, so a client or a reader can check our work rather than take it on faith.
Five rules behind every number
- One answer is one draw. Ask an assistant the same question twice and it can name different brands, so every count comes with a range and the number of answers behind it. How the ranges work.
- We ask what buyers ask. Questions are written the way a shopper types them and name no brand, so a brand’s count is earned in the answer, not planted in the question. A study goes further and asks as simulated buyers: in a hurry, on a tight budget, buying for a child.
- We ask each assistant directly and say which one answered. Every table names the assistant, whether its search was on, and the date. Where we measure a stand-in rather than the app itself, the label says so.
- We read, not only count. A name in an answer can be a recommendation, a place to buy or one example in a list, and every cited site is checked before it counts. Counting is reading.
- We correct in public. A change to a published number goes into a dated corrections log on the page that printed it, whichever way it cuts.
Which assistants, and how each is asked
Three assistants searching the web are in the public census so far, each in its own table and never added together, plus eleven more AI systems answering from memory on the pain relief page.
Narrow screen: scroll the table sideways for more columns.
| Assistant | How it is asked | In the public census |
|---|---|---|
| Claude, with web search | Measured directly, with its web search on, rather than inside the Claude app | Pain relief (each question asked 20 times), sleep supplements and OTC hearing aids (10 times), keyboards (3 times) |
| ChatGPT, free version | Measured directly on the version closest to the free app, with web search on, and labelled a stand-in | Pain relief, sleep supplements and OTC hearing aids (each question asked 3 times, 18 answers each); also the sample study |
| Gemini, with Google Search | Measured directly, with Google Search on, one generation behind the version the free Gemini app runs; the newer one could not search on our account at the time | Keyboards (each question asked 3 times, 18 answers) |
| Eleven more, answering from memory | Asked directly with no web search, each buyer question once: what each one recalls, not what a product built on it shows a buyer | Pain relief (answering from memory) |
| Google AI Mode | Planned, through a data provider we will name | Not yet |
| Perplexity | Checked by hand so far | Not yet |
We do not publish answers gathered by imitating a person in a signed-in browser: in our first research pass, Google flagged that traffic as unusual after about 76 questions, so those answers stay out.
Where the method started: keyboards
The first census asked six questions about keyboards for programmers, three times each. Keychron, the most-named brand, was named in 14 of 18 answers on Claude, with web search.
Three answers to the same question can disagree. Asked “Best split keyboard for coding that will not take months to learn?” three times, Claude put Logitech first twice and Keychron first once. That is why every number on this site carries a range and the count behind it. Read the keyboards census.
Keychron: named in 14 of 18 answers
The detail, one rule at a time
Each opens in place.
Rangeswhy every count carries one, and which one leads
One answer is one draw from how an assistant answered that day. It says nothing durable about the brand. A brand named in half of 20 answers to the same question still carries about ±20 points of uncertainty.
Census tables print a cluster-bootstrap range as the headline figure, since it allows for the same question being asked again, not just for more answers, which widens it. A narrower Wilson range sits beside it in each full table, labelled: Wilson assumes every answer is an independent draw, which is the more optimistic of the two, so we do not lead with it.
Two 2026 papers argue for bootstrap intervals and adaptive sampling on exactly this problem. We print the interval on every table on this site.
Counting is readinghow a mention is coded, by hand where it needs judgment
A keyword count says a name appeared, not how. In the pain relief census, store brands were mentioned 39 times. We read each mention in its sentence and coded it either as a product the answer told the reader to buy or as a qualifier: a place to buy, or one example in a list. 38 were qualifiers and 1 was a product. Two mentions could be read either way; the one closer to naming a product was counted as a product, against our own finding.
The source-integrity checkthree checks before a cited site counts
Before a domain an assistant cites counts toward anything on this site, it gets three checks: its popularity rank in the public Tranco list, whether its homepage actually matches the category it is being cited for, and whether it shows signs of a recent change of owner (a domain that changed hands and now hosts something unrelated to its history). A domain that fails is marked flagged in words, never by colour. A domain that passes is marked verified. A domain we have not yet checked is marked unranked.
Narrow screen: scroll the table sideways for more columns.
| Domain | Cites | By assistant | Created | Tranco rank | Status |
|---|---|---|---|---|---|
| Wedding-venue domain, name withheld | 47 | Claude, with web search, 46 (three versions) · Claude app 1 · Google AI Mode 0 | 1997 | Not ranked | Flagged |
Roughly one in seven of the citations Claude gave in our research got flagged for a human look this way; on a human read, about a third of flags turned out to be on-topic after all. That is why a flag is a reason for a human look, never a finding about the site. The precision figure stays out of client reports until flags are hand-labelled at scale, which is also why this page, not a client report, is where the caveat lives.
What we will not promiseno number without its assistant and its count
Every mention rate on this site names the assistant, the count behind it and a date, which is the discipline this page exists to enforce. That discipline rules out the shortcuts that make a number easy to sell before it is earned: a ranking promised in advance, a score with no count behind it, or a consensus that was manufactured (bought reviews, hidden crawler text, incentivised posts) rather than measured. A number that cannot carry its assistant and its count does not get printed, on this page or anywhere else on the site.
Which AI crawlers can read this siteall the major ones, and the trade-off
robots.txt here allows the major search and AI crawlers: Googlebot, Bingbot, OAI-SearchBot, GPTBot, ClaudeBot, Claude-SearchBot, PerplexityBot and Google-Extended. Google-Extended specifically controls whether Google’s Gemini app can train on and ground its answers in this content. It does not control whether this site shows up in Google Search or in AI Overviews, which Googlebot governs on its own. We leave it allowed anyway: a census that cannot be read is not a census, and the trade-off is that anything readable here can also end up reused by AI companies we have no agreement with.
Appendix for analyststhe counting rules in short, without the question library
Mention rule. A brand counts as named when one of its listed names appears in the body of the answer, matched as a whole word and regardless of case. The answer’s list of sources and every link are removed first, so a link to a brand’s own site is not a mention. A brand that is not on the tracked list cannot be counted.
Denominator. Each census table prints its denominator: the answers it counts, the questions and how many times each was asked. In a study, every brand shares one denominator, the answers to buyer scenarios that name no tracked brand, so a scenario that names a brand never lifts that brand’s rate.
Ranges. Census tables lead with a two-stage cluster bootstrap (questions drawn with replacement, then answers within each question) and show a Wilson score interval on the pooled count beside it. A study prints the span of the two, which is never narrower than either.
Position. The order in which tracked names first appear in the answer body. “In the first three” means among the first three tracked names.
Failed answers. Timeouts, errors, empty answers and block pages are counted in the sampling record and left out of every rate.
Sources. Taken from the assistant’s own citation record where it returns one, and otherwise from the sources the answer lists, which is the assistant’s own account of where it looked.
Hand coding. Where a count needs judgment, such as whether a store-brand mention names a product to buy, each mention is read in its sentence, and a case that could go either way is coded against our own finding.
What stays private. The full question library and its phrasings, how simulated buyers are written, how questions are assigned across buyers and assistants, the instructions each assistant is given, flag thresholds and the fix playbook. A client’s own analyst may inspect the setup of their study under a non-disclosure agreement.
Corrections
Changes to this page. Each census page keeps its own log.
- 2026-10-18: the collection table adds ChatGPT’s free-version stand-in on three health censuses and eleven more AI systems answering from memory on the pain relief census. The detail sections now open in place. No number changed.
- 2026-10-17: the page now names assistants in plain words, opens with the five rules behind every number and ends with an appendix for analysts. No number changed.
- 2026-10-17: the collection table now covers both published censuses, and the keyboards census grid replaces an illustrative grid as the worked example. Each census page keeps its own corrections log.
- 2026-10-16: first publication of the method page.
We measure and source. We do not write, edit or suggest any statement about what a product does; that stays as your label states it. Brand names in our censuses are measurement results, not recommendations.
Read the census