What ChatGPT and Gemini Said They Need Before Citing a Vendor Page: 102 Build Spec Conversations, Counted

On 15 October 2026 we ran 102 build spec conversations on ChatGPT and Gemini across the United States, the United Kingdom and India. Each one worked the same way: show the engine the answer it had just given, point out that it had recommended somebody else, and ask what a page would have to carry before it would cite that page instead. Ten requirements came back often enough to count. All 102 asked for a date, a last verified line and a refresh frequency, and all 102 asked for pricing in the reader's own currency. 101 asked for a before and after with the window stated, 99 for a disclosed methodology, 98 for the engines named one by one instead of the phrase AI, 83 for a confidence interval or a repeat count, 80 for the sample size, 78 for the prompt set published, 68 for failures and negative results, and 64 for the commercial interest declared on the page itself. We sell AI visibility measurement, so this list is a specification we have to meet before we quote it at anybody.

The ten requirements, with the count for each

The unit is one build spec conversation. A requirement counts once per conversation that asked for it, and the denominator is 102 in every row.

Asked by	What the engines asked a page to carry
102 of 102	A date, a last verified line and a refresh frequency. Three fields, not one
102 of 102	The price in the currency the reader is buying in
101 of 102	A before and after, with the window between them stated
99 of 102	A disclosed methodology, written down and attached to the result
98 of 102	The engines named individually rather than called AI
83 of 102	A confidence interval, a variance or a repeat count
80 of 102	The sample size N
78 of 102	The prompt set published, so the reading can be taken again
68 of 102	Failures and negative results, not only the wins
64 of 102	The commercial interest disclosed on the page

Read the top row twice, because it is the one that gets collapsed. A publication date, a line saying when the facts were last checked, and a statement of how often they will be checked again are three separate things, and a page with only the first has answered one third of the most requested item in the set. Our own restated facts have gone stale inside two days: on 16 October 2026 we read OpenAI's bot documentation and found four crawler identities where two of our live pages had named three, so both now carry the fourth name and the date it was read.

The contradiction in the same dataset, and what it actually means

Two findings from the same programme sit against each other, and the gap between them is where this whole list earns its keep.

All 102 said they discount a page published by the company selling the thing. 41 said they would cite such a page anyway, given measurement a reader could check. And 82.0% of the 20,143 citations recorded on 16 October 2026 went to a vendor or an agency own site. The engines describe a discount and then take four fifths of their sources from the kind of page they say they discount.

So the discount is a higher bar rather than a locked door. A self published page competes against other self published pages, not against an independent authority, because in this category there usually is not one. What separates the vendor page that gets used from the one that does not is the ten rows above. Review directories took 0.7% of those 20,143 citations, so a year of directory profile building is a year spent on under one hundredth of the field.

The four rows almost nobody publishes

Six of the ten are cheap: a date, a refresh line, a currency, a methodology paragraph, the engine names and a disclosure sentence are an afternoon of editing. Four are expensive, and those four decide whether the page reads as evidence.

• The sample size, 80 of 102. A percentage without an N cannot be compared to any other percentage. Our own best cell is 3 readings of 210, and printed that way it is a fact somebody can check rather than a figure somebody has to believe.
• A confidence interval or a repeat count, 83 of 102. This one costs real money, because it means asking the same question more than once. Our product reads repeatedly and prints a 95% Wilson range with every rate for that reason. The two audits behind this page asked each question once, which is why they report a source mix rather than a rate with a range.
• The prompt set published, 78 of 102. Publishing the questions lets a sceptical reader reproduce the result and get a different answer, which is the point. A benchmark nobody can re run is a screenshot.
• Failures and negative results, 68 of 102. On 16 October 2026 two of our own four access findings did not survive a second reading hours later, because what looked like a block was a rate limiter and one host returned an explicit HTTP 429. Both are off the table with the reason on the page.

What these 102 conversations are, and what they are not

They are a count of what two engines asked for, on one date, in three markets. They are not a ranking factor list. ChatGPT and Gemini were the two engines here; Perplexity, Claude and Google AI Overviews were not asked, and our product reads five surfaces while these conversations reach two. A wider pass on 16 October 2026 ran 126 build spec conversations, and the counts above come from the 102 of 15 October, so the two sets stay apart in every sentence.

There is also no quarterly index behind this, and no per industry rate exists in either run. The questions aimed at verticals were worded generically and no trade was named, so a software figure could only reach a page by being invented. These conversations were attached to a run whose own failure modes, excluded readings and domain counts are published as what AI assistants actually cite across 597 readings of ChatGPT and Gemini in three markets.

How to use the list when somebody is selling you a number

The ten rows make a serviceable intake test for any report you are handed, in the order of the counts. Ask for the date and the refresh interval first, then the engine names, then the N, then the prompt set. A supplier who can produce all four in a minute has done the work. One who offers a single blended percentage across unnamed engines has produced a presentation. Reading that evidence is a skill of its own, and choosing a generative engine optimisation agency and how to read the evidence it shows you takes a pitch apart line by line.

Questions about the 102 build spec conversations

Did the engines actually say they would cite a vendor page?

41 of the 102 said they would, given measurement the reader can check. All 102 said they apply a discount to a page published by the seller. Both statements are in the same set, and the 82.0% vendor own site share in the readings is the behaviour that goes with them.

Is this list the same as a ranking factor list?

No. It is a count of what two engines asked for when asked directly, on 15 October 2026. Treat it as a specification for a page rather than as a mechanism inside a model.

Which requirement should a small team do first?

The date, the last verified line and the refresh frequency, because 102 of 102 asked for it and it takes an hour. Then name the engines you tested instead of writing AI, which 98 of 102 asked for and costs nothing.

Why would publishing our failures help us get cited?

68 of 102 asked to see them, and in a field where 82.0% of citations go to pages written by sellers, a page that reports a result against itself is doing the one thing its competitors are not. On 16 October 2026 six readings of 1,260 used a page of ours as a source, and that fraction travels with its denominator wherever we quote it.

Were all five surfaces involved in these conversations?

No. The build specs were run on ChatGPT and on Gemini, and nothing in the counts above reads on Perplexity, Claude or Google AI Overviews.

The one row we find hardest, and why it is on this page

Of the ten, the expensive one for us is the repeat count. 83 of 102 asked for a variance or a confidence interval, and both audits asked each question once, which is enough to report which sources an engine reaches for and not enough to report a rate with a range. Saying so costs us the cleanest headline in the dataset and keeps the rest of it usable. AI Knows Us is paid to take this measurement and write the pages that follow it, which is the disclosure row met in plain sight: if the list above is the standard, we are judged by it too.