What did not work: our own domain was cited 3 times in 597 readings, and three other results that went against us

Our own domain was cited 3 times in 597 readings of ChatGPT and Gemini, taken on 15 October 2026. Zero of those came from the United States, on either engine. Zero came from Gemini, in any market. We built the measurement and we sell the fix, and on our own measurement we are close to invisible. This page publishes that result, the three other findings from the same run that went against us, and the baseline the next run gets measured against.

68 of the 102 build spec conversations we ran that day asked for failures and negative results to be shown rather than only wins. 64 asked for the commercial interest to be disclosed on the page. Both are easy to agree with in principle and expensive to do once the number is in front of you, so here is ours, alongside the rest of the run in what AI assistants actually cite across 597 readings in three markets.

Result one: 3 citations in 597 readings, market by market and engine by engine

Market	Engine	Readings	Readings citing us
United States	ChatGPT	102	0
United States	Gemini	102	0
United Kingdom	ChatGPT	102	1
United Kingdom	Gemini	102	0
India	ChatGPT	102	2
India	Gemini	87	0
All three	Both	597	3

The United States row is the one that matters commercially and it is a pair of zeros. Gemini is the other: across three markets and 291 readings on that engine, our domain was not cited once. The one British citation and the two Indian ones all came from ChatGPT.

Two engines produced this grid. Perplexity, Claude and Google AI Overviews were not in the run, so nothing here describes them. Our product measures five surfaces; this result covers two of them.

Why a number this bad is published rather than held back

The same run found that 85.6% of all citations recorded, 10,652 of them across 2,708 domains, went to a page on the vendor's or the agency's own website. In a category where roughly six in seven are self published, the only thing separating a measurement from a brochure is whether the measurement is allowed to come out badly. Ours did. Publishing the wins and keeping this one in a spreadsheet would make every other number we publish worth less, including the ones we are proud of.

It also sets a floor that cannot be moved later. 3 of 597, on 15 October 2026, on two named engines, against a frozen question set. Nobody can quietly restate that as progress.

Result two: a market parameter that reported success and did nothing

The first design for the run set the market with an API parameter. The engine surfaces we read accept gl, hl, location and country, they return success for every value, and they ignore it. Set to the United States, one query answered for Waterloo, Iowa. Set to India, that same query answered for Corinth, Mississippi.

Nothing failed. There was no error to find and no log line to read, and a three market dataset built that morning would have contained one market that was a town in Mississippi. The market was moved into the prompt text instead, and that was then measured: the United States prompt answered for the United States, the United Kingdom prompt answered for London in pounds, the India prompt answered for Mumbai. The method is written up in full on how we measure, and what we cannot see.

Result three: a finding we had to withdraw

ChatGPT volunteered, without being asked, that our pricing page showed rupee prices and no dollar prices. A verification fetch of the live page agreed with it. That looked like a clean third party catch of a real defect.

It was not a defect. The site prices by location, the fetch came from an Indian IP, and so it returned Indian prices correctly. Both the engine and the check were reading from one place and neither of them said so. The finding was withdrawn and the rule it produced now governs the whole programme: a claim about what an engine sees on a website has to state where it was read from, or it is not a finding.

Result four: the Google side of the same domain

Citations in an assistant and clicks from a search engine are different measurements on different surfaces, and the second one is also published. The Search Console export on our proof page, taken 2026-10-14, covers aiknowsus.com from 2026-10-04 to 2026-10-12: 9 days with data, 5,965 impressions, 3 clicks, and a best day of 2,685 impressions at an average position of 41.6. The domain was registered days before that window opened.

The same export page carries a second domain, clawlaw.in, over a 92 day window to 2026-10-12, with 18,537 clicks and a daily average that moved from 59 to 349. That is a different site with its own history, and the proof page states plainly that neither case is evidence for the other. We keep them in separate paragraphs for that reason.

What these 597 readings do not entitle us to say

• They do not say the engines never cite us. They say 3 readings in 597 did, on one day, on two engines.
• They do not say anything about Perplexity, Claude or Google AI Overviews, which were not in the run.
• They do not produce a citation rate with a range around it. One reading per question measures a tendency, and the rate needs repeats.
• They do not transfer to another company's domain. The honest use of this page is as a worked example of the measurement, not as a claim about your site.
• They do not settle what works in AI assisted search in general, a wider question handled in artificial intelligence in SEO, what works and what is oversold.

Questions about our own audit result

Is 3 citations out of 597 readings really the whole figure?

It is. 3 readings out of 597 cited a page on aiknowsus.com: one in the United Kingdom and two in India, all three on ChatGPT, taken 15 October 2026.

Does this mean generative engine optimisation does not work?

It means our own domain had not been read yet on 15 October 2026. The same run shows what the engines were citing instead: 85.6% of citations went to vendor and agency owned pages, and small specialist sites were cited dozens of times each. The channel was carrying other people's pages, which is a measured fact about the channel rather than about the idea.

Why publish before the number improves?

Because a baseline published afterwards is not a baseline. The date, the N and the zeros are fixed now, by us, in public, so the next run can be checked against them by somebody who does not trust us.

Which engines and markets does this cover?

ChatGPT and Gemini, in the United States, the United Kingdom and London, and India. 102 questions per market, 72 of them identical across all three.

When is it measured again?

The benchmark page commits to a re-run of the same question set every 60 days, reported with its own date.

The baseline, stated so it can be used against us

3 citations in 597 readings. 0 in the United States on ChatGPT and 0 on Gemini. 0 on Gemini in the United Kingdom and 0 on Gemini in India. 1 on ChatGPT in the United Kingdom, 2 on ChatGPT in India. Question set frozen, both engines named, 15 October 2026.

AI Knows Us sells AI visibility measurement and the work that follows it, so a page like this one is against our short term interest and we are publishing it anyway. The next run uses the same 102 questions, the same two engines and the same three markets. If the number has not moved, that will be on this site too, with the same table.