AI brand visibility in the US, measured: what counts as one citation, and why the same 1,260 readings produce 20,143 or 36,576

Two AI visibility numbers can describe exactly the same data and differ by 16,433, with neither one wrong. On 16 October 2026 we put 210 buying questions per market to ChatGPT and Gemini, 1,260 readings in total including 420 American ones, none discarded. Counting one domain in one reading, deduped, gives 20,143 citations across 4,947 distinct domains. Counting raw URL occurrences on the identical data gives 36,576. That is not a correction, it is a different measurement wearing the same word, and the two cannot be compared. Any American business comparing AI visibility reports is almost certainly comparing numbers counted under different rules, because the rule is the thing that never makes it onto the page. Every figure here uses the deduped definition, and the American cells are 1 citation of our own domain in 210 ChatGPT readings and 0 in 210 Gemini readings.

The counting unit, which is the rule that produced both totals

A domain cited four times inside one answer is one citation here, not four. The same domain cited once in each of four answers is four. That single choice is the whole difference between 20,143 and 36,576, and it is almost never printed. Without it, a citation count is a number with no unit attached, like a distance reported without saying miles or feet.

The other half of the confusion is the word being counted. Appearing in the text of an answer, being one of the sources the answer was assembled from, and being put forward to the buyer as a choice are three different events, and a business can have any one without the others. Every figure on this page is the middle one, a source, counted by the rule above.

The American readings, and the whole run they sit in

What was measured	16 October 2026
Questions per market	210, in 21 batches, 630 in total
Readings recorded	1,260, all 1,260 usable
US readings	420, half on each engine
Citations, deduped by domain and reading	20,143 across 4,947 distinct domains
The same data counted as raw URL occurrences	36,576
Readings citing our own domain, US ChatGPT	1 of 210
Readings citing our own domain, US Gemini	0 of 210

Sorted by type: 82.0% a vendor or agency own site, 12.3% official or platform documentation, 3.2% news or a large publisher, 1.9% social or community, 0.7% a review directory. The run of 15 October 2026, from a different question set, is in what AI assistants actually cite across 597 readings of ChatGPT and Gemini in three markets, and its American cut is in what the two engines cited for United States buyers.

The second rule that moves a headline figure: proxies

712 of the cited URLs in this run were not publishers at all. They were the data provider own fetch infrastructure, an iframe proxy and a reverse proxy mirror of Google documentation, serving the real documentation page under a hostname of their own. Left in, they sit in a most cited domains table as though a proxy service were a publisher of search guidance. Resolved back to the page they actually serve, they move the vendor own site share from 82.4% to 82.0%, and they moved the earlier run from 85.6% to 85.3%.

Two of our own live pages listed those hostnames before this was caught, and both are corrected with the explanation on the page. The same artefacts will be in anybody else's data, because they come from how data is fetched rather than from how any engine behaves. When a report shows you a domain you have never heard of, open the cited URL. If it resolves to somebody else's page, the row should not be there.

The ten domains that won, and what it costs an American publisher to compete

Cited in N readings	Domain	Type
446	developers.google.com	official or platform
328	support.google.com	official or platform
325	developers.google.cn	official or platform
309	blog.google	official or platform
269	help.openai.com	official or platform
243	github.com	social or community
230	developers.openai.com	official or platform
199	semrush.com	vendor or agency own site
197	theverge.com	news or publisher
147	searchengineinfo.org	vendor or agency own site

Six of the ten are a platform answering questions about itself, which is why a page written to beat developers.google.com on Google's own mechanics is budget spent for sport. What makes the rest of the list interesting is the widening of the question set. The earlier run's top ten, built from 102 questions per market, had four specialist AI visibility tools in it: tryprofound.com cited in 61 readings, otterly.ai in 58, depra.ai in 53 and peec.ai in 49. This top ten, built from 210 per market, has no specialist visibility tool in it at all.

The encouraging half is the 82.0%. Outside the platform documentation layer the slots are held by self published pages rather than by directories, which took 0.7% of citations. An American firm weighing a year of directory profile building against a quarter spent publishing evidence now has a number for that argument instead of two opinions.

The controls behind the American readings

American phrasing, the country written into the question rather than switched on as a parameter, and each answer checked for dollars and American context before it entered the count. Our own business name appeared in none of the questions. Source classification runs through one tested file that refuses to load if any of 24 named examples sorts wrongly, which is how the proxy problem above was caught rather than argued about. The full method of the earlier run, including its failure modes, is in how the question set and the 597 readings were built.

The two runs agree, which is worth more than either on its own. 15 October: 3 citations of our domain in 597 readings, 85.6% vendor own site, 0.8% review directories. 16 October: 6 in 1,260, 82.0% and 0.7%. Different questions, twice the readings, same answer.

What this American dataset does not contain

No rate by trade, because no trade was named in the 210 questions and inventing one is the thing this page argues against. No American city figure in this run; city readings came from a separate scan that read ten named cities one at a time. And the count stops at two engines. Google AI Overviews, Perplexity and Claude were never asked these questions, we read five surfaces for customers, and two is not five in any sentence here.

Questions about counting AI citations in the US readings

How often do these engines recommend American businesses?

Putting a business forward as a choice was not scored here. What was scored is which domains the answers were assembled from, under the counting rule at the top of this page. The recommendation score is the next thing to add, and it will arrive with its own numerator, denominator and date.

Which engine cites American businesses more often?

Under the rule above, one American ChatGPT reading of 210 used our domain as a source and no American Gemini reading did. Add both runs together and Gemini gave 921 readings with our domain in none of them. Report the two engines separately rather than averaging them, because they are not two samples of the same behaviour.

Why not publish the larger citation total, since it is from the same data?

Because every figure published on this site has used the deduped rule, including the earlier run, and switching the rule would make a new number look like growth. Where the raw count is useful it is published beside the deduped one with the rule stated, which is what this page does.

Does a bigger sample make a benchmark more reliable?

Only alongside the attempted count and the usable count. 1,260 of 1,260 usable is a different claim from a benchmark that quietly discarded a share of its attempts, and a page reporting one total without the other has left out the part most likely to matter. Repeated answers to the same question are also not independent observations, so running one question three times does not make three samples.

What belongs in an American AI visibility report?

The counting rule, the attempted and usable reading counts, the engines by name, the numerator and denominator per engine and market, the date, and the negative cells. What a report needs to survive a sceptical reader is worked through in building a report that survives a sceptical reader.

The sentence to look for in the next report you are shown

One line decides whether an American AI visibility number can be used: the one that says what a citation was counted as. If it is missing, the figure cannot be compared to any other figure, and the gap between 20,143 and 36,576 on identical data is the size of the error that line prevents. We sell this measurement and the publishing that follows it, so the demand applies to us first. Our own American result is 1 reading of 210 on ChatGPT and 0 of 210 on Gemini, on 16 October 2026, counted one domain per reading, deduped, with the proxies resolved before counting.