4,947 domains were cited, and ten of them are the ones everybody argues about
A run of 1,260 readings produced 20,143 citations spread across 4,947 distinct domains, which puts almost all of the field outside the table anybody publishes.
Published by AI Knows Us · Updated 19 October 2026
The 16 October 2026 run of 630 buying questions on ChatGPT and Gemini produced 20,143 citations across 4,947 distinct domains. Ten of those domains make up the top ten that gets quoted, reprinted and argued over. The remaining 4,937 are the part of the field a publisher can actually enter, and they are the part nobody writes about because a long list is harder to turn into a headline than a short one.
That is not a complaint about the top ten. It is an argument about where a plan should look, and the arithmetic is on the side of the tail.
Why a top ten is the least useful part of a source table
- The rows at the top are occupied by the subject of the question. Platform documentation holds the questions about the platform, and the publisher of a thing is the primary source for that thing.
- A short list cannot show you a gap. Gaps are in the middle of a distribution, where several domains appear a handful of times each and none of them has settled the question.
- The top of a table is the least likely part to change. The tail is where a reading taken next month will differ from this one, which is the only place a plan can have an effect.
- Your buyers' questions are not the top ten's questions. Price, suitability, who to hire and anything tied to a place all sit in the tail by construction.
What the tail is made of
Our own data gives named examples of tail domains, from the earlier reading of 15 October 2026 and from the separate city scan. They are small, specific sites, and their counts are small too.
| Domain | Citations or readings | Which reading |
|---|---|---|
| whito.co.uk | 46 | 15 October 2026, 597 readings |
| rank4ai.co.uk | 39 | Same reading |
| aeoscore.co.uk | 32 | Same reading |
| upgrowth.in | 52 | Same reading |
| themediaverse.in | 48 | Same reading |
| nicodigital.com | 44 | Same reading |
| hawkorbit.com | 33 | Same reading |
| localseo.nyc | 3 of 12 | City scan, New York, six questions on two engines |
| chicagoseoagency.co | 3 of 12 | City scan, Chicago |
| houdigitalmarketing.com | 3 of 12 | City scan, Houston |
None of those is a large company. Several of them are a domain name plus a handful of pages about one place or one market. In a top ten they are invisible; in the reading that a buyer in their market actually gets, they are the answer.
The counting rule that decides how big the tail looks
Before any of this is comparable to anybody else's figure, one line has to be on the page. Each citation in our totals is one domain inside one reading, so four appearances of a host within a single answer still come to one. Run the identical data by raw address occurrence and the total is 36,576 rather than 20,143. Neither number is wrong and the two cannot be set beside each other in a sentence.
A second rule moved the shares before they were published. 712 cited addresses belonged to our data provider's own fetching setup, not to any publisher, and each was pointed back at the page it hands over before counting, which took the vendor own site share from 82.4% to 82.0%.
How to find your own tail, which is the part that matters
You are not looking for the most cited domain in your category. You are looking for the questions where two or three small domains appear once each and nothing dominates, because those are the readings a new page can change this quarter.
- Ask between thirty and a hundred buying questions in your buyers' own words, with no company name anywhere in them.
- Record every cited domain and which engine returned it, in its own column.
- Sort the questions by how concentrated their sources are, not by how many sources they had.
- Start with the flattest questions. Those are the ones with room in them.
Keeping that list as a reading list rather than a scoreboard is the habit that makes it useful, and the same distinction between a map of sources and a ranking of them runs through a map of the subjects and the sources worth your time.
Where this reading stops
It covers 630 questions in one category on a single date, put to two named apps and no others, which were ChatGPT and Gemini. Our product reads five surfaces and three of them were asked nothing here. There is no per industry figure inside it, because no industry was named. City numbers come from the separate scan of ten named cities and nowhere else. And every count is a citation, meaning a domain in a source list, which is a different event from a business being named in the answer text.
We sell this measurement and the publishing work that follows, which is why our own row is printed in the same run: 6 readings of 1,260 had a page of ours in the source list. The search console side of our own record, both windows and the export date, is on our proof page.
Questions about the long tail of AI citations
How many different domains were cited in the run?
4,947 distinct domains, holding 20,143 citations, from 1,260 readings on 16 October 2026.
How much of that goes to the top ten?
The top ten is reported in readings citing each domain rather than in citations, so the two units cannot be combined into a share. What the table does show is ten named rows and 4,937 other domains.
Is a small domain really worth targeting?
The targets are questions, not domains. A question whose sources are three small sites at one citation each is a question where a page you publish has somewhere to land, and the named examples above are what those sites look like.
Why do published studies always show the same top ten?
Because the top of a distribution is stable and short, which makes it quotable. The useful part of a table moves, which is also what makes it hard to headline.
Which engines and date is this?
ChatGPT and Gemini, 210 questions per market in the United States, the United Kingdom and India, read on 16 October 2026. The earlier reading in the same programme, with its full method, is 597 readings across ChatGPT and Gemini in three markets.
The number to put in front of a board, and the one to work from
Put 4,947 in front of the board, because it is the figure that answers whether there is room. Work from your own flattest questions, because that is where the room actually is. A plan written against the top ten is a plan written against the subject of the question, and the subject always wins that one.