82.0% of the citations landed on a page somebody owned, and that is the whole asymmetry
Of 20,143 citations recorded on 16 October 2026, 82.0% went to a vendor's or an agency's own website, which is the only part of the field anybody can change by deciding to.
Published by AI Knows Us · Updated 19 October 2026
Sorted by the kind of page they landed on, the 20,143 citations recorded in our 16 October 2026 run of ChatGPT and Gemini break down like this: 82.0% on a vendor's or an agency's own website, 12.3% on official or platform documentation, 3.2% on news or a large publisher, 1.9% on social or community, and 0.7% on a review directory. The first row is the one you control. Everything underneath it belongs to somebody else, and no amount of effort transfers a single citation from one of those rows into your account.
That asymmetry is the finding. Four fifths of the field is made of pages that somebody decided to publish, and the decision is available to you on the same terms it was available to them.
The shares, and what each row is
| Kind of page | Share of 20,143 citations, 16 October 2026 | Who decides what goes on it |
|---|---|---|
| A company's or agency's own website | 82.0% | The company. You, for your own |
| Official or platform documentation | 12.3% | The operator whose product it documents |
| News or a large publisher | 3.2% | An editor |
| Social or community | 1.9% | A community, and its moderators |
| A review directory | 0.7% | A platform, and its ranking rules |
Read the right hand column down and the strategic question answers itself. The row where you hold the pen is also the largest row by a distance of nearly seventy points over the next one.
The figure held when we tried to break it
A single share from a single day is a number, not a finding. This one has been measured twice on two deliberately different question banks. On 15 October 2026, 102 questions per market and 597 usable readings gave 85.6% for the same row out of 10,652 citations. On 16 October 2026, 210 questions per market and 1,260 readings gave 82.0% out of 20,143. The second bank was built to go into eight directions the first never went, and the row moved by under four points.
One rule had to be applied before either figure was published. 712 cited addresses in the second run belonged to our data provider's own fetching setup rather than to a publisher, and pointing each one back at the page it hands over moved the row from 82.4% to 82.0%. The first run shifted the same distance, 85.6% to 85.3%, and the headline stays as measured with the difference recorded.
What the 82.0% does not license
It does not say any page on your own site will be cited. It says the citations in this category sit mostly on pages of that kind, so in those readings you are up against other people's marketing rather than against an independent authority. That contest turns on what is actually printed on the page, and the fullest list we have of what the engines themselves asked a vendor page to carry, counted across 102 conversations on 15 October 2026, is published as what ChatGPT and Gemini said they need before citing a vendor page. Two items from it are worth repeating here because they are the cheapest to act on: a date with a last verified line, asked for in every one of the 102, and a price in the reader's own currency, also asked for in every one.
A build spec is an engine's opinion about what would earn a citation rather than evidence about what is true, and we treat it as such. As a list of things to put on a page you were already writing, it costs nothing to follow.
The budget arithmetic, stated carefully
The temptation is to split a content budget in proportion to the shares above. That is the wrong move for one reason: a citation share is not a value share. A directory profile at 0.7% of citations still gets read by buyers, and a press placement at 3.2% still carries weight in a sales conversation. Both have value that has nothing to do with a source list.
So the honest instruction is narrower. Spend against the shares only for the part of the budget whose stated purpose is being cited, and keep the rest of the budget judged on its own outcomes. Written down that way, most plans turn out to have labelled something a citation strategy when it was a sales asset, which is a naming problem rather than a spending one and it is fixed in an afternoon.
What stays true either way is which side of the table you can move, which is the pages on your own domain and the questions you choose for them. The same ownership argument shows up when the instrument changes rather than the channel, as in what each of those two instruments can actually measure.
The edges of this reading
Two apps, ChatGPT and Gemini, on one date, with 630 questions in three markets. Our product reads five surfaces and three of them are absent from every figure here. There is no per industry rate in the run, because no industry was named in a question. Anything city level in this programme was measured by a different scan, of ten named places, and is not in these shares. And a citation is a domain in a source list, which is a different event from a business being named in the answer text.
We sell the measurement and the publishing work, and our own row in the same run is 6 readings of 1,260. The search side of our own record, with both windows and the export date on the page, is our proof page.
Questions about where AI citations actually land
What share went to companies' own websites?
82.0% of the 20,143 citations recorded on 16 October 2026. The earlier run of 15 October 2026 put the same row at 85.6% of 10,652 citations.
Does that mean I should stop using directories?
It means a directory profile is a poor instrument for being cited, at 0.7% of citations. Keep it for the buyers who read it, and stop counting it against a citation target.
Why do the two runs disagree slightly?
Different question banks, eleven days apart in design, with the second one twice the size and aimed in eight new directions. A row that moves by under four points across that is a row that replicated.
Which pages on my own site are the ones that get cited?
Ours is not a page type benchmark, so the honest answer from this run is the kind of page rather than the template: one question per page, the answer at the top, with a date, a stated method, a sample size and a price in the reader's currency.
Which engines produced these shares?
ChatGPT and Gemini, 210 questions per market, read on 16 October 2026. The earlier reading, with the method and the question design published alongside it, is 597 readings across ChatGPT and Gemini in three markets.
The one row you hold the pen on
Every other row in that table is somebody else's editorial decision, and each one has a gatekeeper who did not ask for your plan. The largest row has no gatekeeper at all. Pick your questions, publish the answers with the fields above attached, and date them, because that is the part of the field that moves when you decide it should.