Measurement register

Our articles cite this register. Here it is. Every figure we publish about what AI assistants answer comes from a row below, and every row carries the question exactly as it was asked, which surface answered, the market with its location code and language, the date, and how many runs the number rests on.

Why this page exists

A number without its prompt is not a measurement, it is an assertion. Anyone can run these questions and get their own answers, and the point of publishing the prompts is that they should be able to. Where our figure and yours disagree, the disagreement is the interesting part and the prompt is how you find it.

Single-surface measurements

One question, one surface, repeated the stated number of times. Cited domains only, never the search results behind an answer: those are different things and counting one as the other is the mistake that forced four studies to be withdrawn.

Cross-surface measurements

The same question put to several surfaces at once, to see whether they cite the same pages. The last four columns are how many domains appeared on four surfaces, three, two, and exactly one.

Agency buying questions by vertical

Study: 58% of AI Agency Citations Appear Once

One row per "best marketing agency for X" question, with Google AI Overview, AI Mode and Gemini pooled. Cited is every distinct source across the three surfaces, stable is the sources present in every run, and once is the ones that appeared a single time. Three rows re-ask a question that already appears above, one of them on the same day in a fresh session; all three are excluded from the totals the study publishes, because pooling a question with itself counts one measurement twice.

Travel booking questions

Study: Same Vertical, Two Questions, 53 Points

Ten rows across two intents and four markets. Two are the same question in two sessions, kept apart rather than pooled. The surfaces column matters here: the Spanish question timed out twice on three surfaces and only completed on two, so it draws from a smaller pool and is not comparable with the other two.

One intent, four wordings

Study: Four Wordings, Four Survivors, No University

Two intents, each asked four ways in two markets: booking a hotel in the United States and Spain, and choosing a university in Spain and France. A few rows carry a note instead of a clean reading, and they are here rather than filtered out because a register that only shows the runs that worked is not a register. The wording column is a slot and not a ranking, and P1 in every market is a prompt that already appears in a table above.

Who holds the durable citation slots, by category

Study: Vendor Slots Follow the Category

One high-intent buying question per category, asked on AI Overview and kept only where a domain came back in every answered run. The vendor column counts the durable domains that sell the product the question asks about. Rows marked with another market repeat a category rather than adding one, and are not part of the headline total.

The same questions on Gemini, and a test that voided itself

Google-Extended is the control over whether content may be used for Gemini and for grounding, so Gemini is the surface where a publisher blocking it should stop appearing. Four publishers in the crawler study block it and are cited in AI Overview. Asked on Gemini the same day, all four are absent, which looks like a confirmation and is not one: only 8 of the 43 non-blocking AI Overview domains appear on Gemini either, so at that base rate the expected number of blockers present was under one, and we observed none. A test whose positive result is what chance predicts has not tested anything, and these rows carry the arithmetic rather than the headline. What they do support is smaller and real: two Google AI surfaces, the same six questions on the same day, share 8 domains out of 47.

Brand demand against citation, and which names are whole suites

The two demand studies published on 7 August 2026 compare how much a tool is searched for with whether the answer named it. Their limits paragraph deflated the suites on the uncited side and named a direction. This column is what that correction turns on: a name marked as a suite sells a whole SEO platform with an AI module inside, read from its own homepage, so its brand volume is not demand for the product being compared. Both sides carry suites, and on the cited side one of them is almost the whole total.

What the truncation does to an overlap, measured instead of reasoned

We published a caveat saying the missing part of the AI Mode set could only make the two surfaces look further apart than they are. For the two cells where we hold every marker the answer printed, we computed the overlap both ways. It moved in both directions. Which way depends on the last column: whether the domains the tool dropped are ones AI Overview also cites.

The pages an answer cited, and whether they answer

Study: Blocking AI Crawlers Does Not Remove You

Every other set on this page records which domains an engine cites. None asks whether the citation resolves, which is one request per citation. These rows take every URL printed in four AI Overview answers across three languages, fetch each one with GET and redirects followed, and then read each host's robots.txt for what it says to named AI crawlers. One draw per question, from the cached call rather than the repeated one, because the unit here is the URL an answer printed and not how often it prints it: nothing here is a rate of citation. Blocked is its own bucket for 401, 403, 429 and 999 and is never folded into dead, and a host that refuses us its robots.txt is counted apart from a host that permits everyone.

Two shapes of a zero, and what a re-read does to each

Study: What Predicts a Nameable Source Set

Counting this register rather than measuring anything new: of the readings here with three or more answered runs, some return zero durable sources with every domain landing exactly one run short of the bar and google.com cited once, and others return zero with the best domain far below it. The first shape is what a run that returned only Google leaves behind. These rows re-read two cells of each shape at three runs to ask whether the first kind is recoverable. The dispersed rows are the control, and they are the reason the comparison means anything, because this register has already measured that re-reading alone moves the statistic a great deal. Three runs is a weaker bar than the four to nine the originals used, which favours recovery in both arms, so only the contrast between arms can carry anything.

The same question read twice, across several questions

Study: What Predicts a Nameable Source Set

The set above reads two questions many times each. This one asks the cheaper question a buyer would actually run: across several categories, do two consecutive readings of one question agree about which sources are durable? Five questions, AI Overview, three runs a reading, two comparable readings each, one day. Two readings excluded themselves by stopping early, and they are kept here with their real answered count rather than deleted, because a bar asking for unanimity is weaker at two runs than at three. The categories were chosen because earlier rows already cover them, which makes these comparable and makes them not a sample of anything.

One question read six times in a row

Study: What Predicts a Nameable Source Set

Every other set on this page reads a question once, or reads it again on another day. These rows read one question six times back to back, on one surface, three runs a reading, and then repeat the design with a second question in another market and another language. They exist to close a question this register carried open since 10 August: five readings of one prompt declined in order, and a declining sequence taken in order cannot be told apart from noise. It cannot, and it was noise. The row marked as fresh is the control that rules out a session effect: a question never asked in that session, inserted after the fifth reading of the other one. Read the answered runs column before the counts, because one reading stopped early at two runs and a source cited in every run clears a weaker bar at two than at three.

One question read deeply, instead of three times

Study: What Predicts a Nameable Source Set

Every other set on this page uses a strict bar: a source counts if it was cited in every answered run, with three runs. These rows ask what sits underneath that bar by requesting up to ten runs of one question and recording how many answered runs cited each domain. The denominator of every count is the runs that happened rather than the runs asked for, and the two differ often enough to matter: 26 of these 59 readings stopped early and 33 completed the request. The Set column says which question a row belongs to. The pair is the original deep read that found the shape; the sweep tested whether it generalises across categories; the market rows ask whether a fixed source set belongs to a category or to the country it was measured in; the bar rows take cells that looked fixed at three runs and read them deeply; and the wording row asks the same intent as a sweep row in different words. Each set is scoped separately on purpose, because pooling them would move figures the articles quote.

The same string put to the ranked page and to the answer

Study: What Predicts a Nameable Source Set

The rows above ask how many domains an answer cites. These ask where those domains come from, by putting the identical string to Google's organic results on the same day and counting how much of the cited set was already on page one. The count that also ranks sits between 6 and 10 in all six categories while the cited set runs from 9 to 104, so what separates a narrow category from a wide one is what gets cited from off the ranked page. The distinct-domain count of the organic page itself is not compared across categories anywhere, because a top-twenty page is bounded at twenty by construction and its flatness belongs to the instrument rather than to the market. The last row is a control and not a buying question: it asks how to rank in AI Overviews, a different intent in a different vertical, and it lands at 53 per cent of its ranked page.

Who ships an llms.txt, asked of two populations on one day

The catalogue vendors and the domains this register has seen cited three or more times, both asked for /llms.txt on 14 August 2026 with the same instrument. Reachable is asked minus bot challenges and fetch errors, because those are facts about an edge and not about whether a file exists. Vendors ship the convention far more often than the sites that actually get cited. That is a comparison of two populations and not a mechanism, and nothing here shows the file helps or hurts: what it does show is that winning citations does not require one.

The crawler and convention audit of 11 August 2026

Study: All 66 Passed the Crawler Audit

Three checks every tool in this market sells, run against all sixty-six vendors in the catalogue that day, plus what each one publishes by convention. The counts are frozen at what 11 August measured: the catalogue has grown since, and this is a fact about a date rather than a denominator that follows it. EchoWi is one of the sixty-six.

Crawler names looked for, so a zero is a zero of something: GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, anthropic-ai, PerplexityBot, Perplexity-User, Google-Extended, Bytespider, CCBot, Applebot-Extended

The page behind the domain

Study: The Page Behind the Domain

Every other citation study here counts domains. This one records the exact pages a single AI Overview cited for each question, because a domain does not tell a buyer whether one page holds the answer or several do. One draw per cell on purpose: repeating the call returns a cached answer, so this describes a set of questions and is not a rate.

Citations the answers printed, and the pages they resolve to: 92 / 91 (ai_overview), 120 / 120 (ai_mode)

Homepages among the cited pages: 0 / 91 (ai_overview), 3 / 120 (ai_mode)

Who declares Content Signals in robots.txt

Study: Who Declares Content Signals

Content Signals are directives inside robots.txt that state what a crawler may do with a page after reading it: search, ai-input and ai-train, plus a fourth field, content-use, still being tested. Both rows below are a GET of /robots.txt and a read of every Content-Signal line, on 14 August 2026. A challenge is counted apart from a missing file, because it is evidence about the edge and not about the file. EchoWi is inside the vendor row and inside its seven, so the method is stated in full rather than the conflict being left implicit.

What the border-crossing domains turned out to be

The two-layer rows above cite 451 distinct domains across two markets and 13 of them appear in both. Those 13 were published as a list of names first, with the classification declared as not done, because a domain name is a hypothesis about what a site sells and its home page is the answer. This is that pass: each one opened on 14 August 2026 and recorded by what its own page says. Two came back behind a bot challenge and are recorded as unclassified rather than guessed, and the study gives its figure with and without them. Seven of the thirteen are the vendors themselves.

The same five questions on a second and third surface

Study: One Surface Is Not the Web

Every other set on this page was measured on AI Overview, so a finding about categories could be a finding about one product. These rows ask the five United States questions above on AI Mode and on Gemini, same market, same run count, same vendor rule, and only the surface changes. Three Gemini rows stopped before three runs and are marked, so their bar is lower and they are not compared. No vendor share is computed from Gemini: this route is known to return no sources at all for one assistant, so a short citation list cannot be told apart from a short capture.

Whether the surface answers at all, by question type

Study: No AI Overview for Local Best-Of

Two surfaces asked in the same call. The AI Overview column counts runs that produced an answer, not citations: a zero there means the surface did not render, which is a different fact from a surface that ran and cited nobody. The last row is the control, local in the same way and informational rather than commercial.

A buyer's keyword against a tool's prompt, same surface and day

Every product in this category asks questions, because that is what someone types into an assistant. An AI Overview also fires on the keyword phrase a buyer types into Google. 13 categories asked both ways, three runs requested each, in one market. The first arm is AI Overview; a second arm repeats four of the categories on AI Mode, and the surface column says which. `answered` is read from the result, and a pair whose two sides answered a different number of times is excluded from the median rather than compared across two different bars. The rows marked as a re-read are the control: the categories whose forms disagreed are also categories this register knows are unstable, and only re-reading the same form separates those two explanations.

Which questions still reach Reddit

13 cells across four categories, United States, English, 20 August 2026, on Google's two AI surfaces. Four runs requested per cell and the answered count read from each result rather than assumed. Reddit is cited in every small-business and consumer cell and in none of the enterprise ones, and the question's phrasing does not predict it: holding the buyer fixed, the plain commercial question pulls Reddit as reliably as the experiential one. The domains column is what makes an absence readable, because a zero in an answer citing 53 sources is a different fact from a zero in a thin one. This study cannot speak to ChatGPT: our route returns no citations from that surface at all.

Asking the identical question twice, and what comes back

Every other stability figure here is taken inside one reading: ask N times, keep what survives all N. This asks a different question. Take that surviving set, ask the identical string again minutes later, take it again, and compare the two. Each pair is read at a matched bar, two runs in the first pass and three, four or five in the second, because a set that survived five runs is tougher than one that survived two and comparing across bars measures the bar rather than the market. The two passes are kept apart for the same reason, and the row marked later the same day is not a within-session reading at all: it is the same string at the same five-run bar, asked again hours afterwards. Core is the domains in both readings, rim is those in only one. The row marked off-bar is corroboration and is excluded from every comparison.

What a page an AI cited declares about itself

The population here was chosen by the engines and not by us: every URL is one that appeared as a source in a reading elsewhere in this register. That makes it selected on the outcome, so it cannot show that structured data causes citation. What it can do is settle whether the pages getting cited in this category are heavily marked up. Declares is the subset of JSON-LD types that say what the page is rather than who published it. A page that answered with almost no body text is marked shell, because that is a fact about where it keeps its content, and a page we could not read is kept out of every count.

Who writes the GEO-agency rankings, in three markets

Page one of the buying query for a GEO agency in Germany, the United States and France, read on 19 August 2026. Elsewhere this register counts this from what Google prints in its snippet and says so; here every page was fetched and its list read in document order, which is what produces the rows a snippet cannot see: a publisher inside its own list at position 3 or 5 rather than first. Kind comes from each publisher's own homepage title, never from its domain name, and it splits the table into those who sell in the category being ranked and those who do not. Self position is where the publisher appears in its own list, and zero means it is not in it. Entry lists for the American and French rows are truncated where a page runs long, so their length is a floor and no figure counts it. A page behind a bot challenge is kept out of every count.

Do the never-cited vendors publish category lists too

33 vendor domains, every vendor the named-versus-cited study ever saw named in an answer, read on 20 August 2026. The count is URLs in the vendor’s own sitemap whose slug matches a category-list shape, a set declared before any site was opened: best-, top-, -vs-, alternatives, comparison, comparativa, mejores-, mejor-, beste-. It is a loose proxy and the looseness runs one way: it inflates both groups and pushes toward finding no difference, which is what was found. Six domains served no readable sitemap and are marked as unread rather than counted as an absence.

Does the answer name the vendors whose pages it reads

72 vendor observations in 2 markets, on AI Overview and Gemini: 48 in the United States in English and 24 in Spain in Spanish, all on 20 August 2026. The vendor list was frozen per category before each call, so the count describes those vendors and not the category. Naming is read from the answer prose with URLs stripped first, because the body carries markdown links and a raw search would credit a name the answer never said. AI Mode and ChatGPT are absent by design: neither returns a usable citation domain by this route, so every observation would fall into “named, not cited” by construction. The linked column marks the separate case where the body links a vendor domain that the citation list omits.

Which questions still reach Reddit

13 cells across four categories, United States, English, 20 August 2026, on Google's two AI surfaces. Four runs requested per cell and the answered count read from each result rather than assumed. Reddit is cited in every small-business and consumer cell and in none of the enterprise ones, and the question's phrasing does not predict it: holding the buyer fixed, the plain commercial question pulls Reddit as reliably as the experiential one. The domains column is what makes an absence readable, because a zero in an answer citing 53 sources is a different fact from a zero in a thin one. This study cannot speak to ChatGPT: our route returns no citations from that surface at all.

Which vendors serve markdown to an agent

Every active vendor in our verified catalogue, asked for markdown two ways on 19 August 2026: the homepage with an Accept header of text/markdown, and the sibling path /index.md. A body counts as markdown only if it is not HTML and carries a markdown structure, so a plain-text error page cannot pass. A host behind a bot challenge is its own bucket and never counted as absent. The detector was validated first against three cases whose answer was already known.