What AI assistants actually cite: 597 readings across ChatGPT and Gemini in three markets
A reproducible measurement of where ChatGPT and Gemini get their recommendations, run on 15 October 2026, with the question set published.
Published by AI Knows Us · Updated 15 October 2026
We asked ChatGPT and Gemini 102 buying questions each, in three markets, and recorded every source they cited. 597 readings, 10,652 citations, 2,708 distinct domains, run on 15 October 2026. The headline finding: 85.6% of everything these two engines cited was published by the company or agency selling the thing being recommended. Review sites such as G2 and Capterra accounted for 0.8%.
The prompt set is published below so anyone can re-run this and check the result.
Correction, 15 October 2026. The first version of this page published the vendor share as 91.3%. That was wrong. The classifier that produced it had a faulty pattern which required the literal string "developers.." and therefore never matched, so Google's and OpenAI's own documentation, 178 readings citing developers.google.com and 138 citing support.google.com, was counted as vendor marketing rather than as platform documentation. The corrected figure is 85.6%. The classifier is now a single tested file that refuses to run if any of 24 named examples is sorted wrongly. The finding did not change; the number did, and a page that argues for checkable numbers has to show its own corrections.
What was measured
| Item | Value |
|---|---|
| Date of run | 15 October 2026 |
| Engines | ChatGPT and Gemini. Two engines, not five. Perplexity, Claude and Google AI Overviews were not part of this run. |
| Questions | 102 per market, 72 identical across all three markets plus 30 market specific |
| Markets | United States, United Kingdom and London, India |
| Readings | 597 usable, from 612 attempted |
| Citations recorded | 10,652 across 2,708 distinct domains |
| Readings that returned any source | 487 of 597 |
Who gets cited: the source mix
This is the number that should change how a marketing team spends its budget in this category.
| Source type | Citations | Share |
|---|---|---|
| The vendor's or agency's own website | 9,114 | 85.6% |
| Official platform documentation | 1,031 | 9.7% |
| News or a large publisher | 265 | 2.5% |
| Social or community | 158 | 1.5% |
| A review directory (G2, Capterra and similar) | 84 | 0.8% |
The common advice in this category is to get listed and reviewed on the software directories. On this measurement those directories were 0.8% of what the two engines actually cited. Roughly six citations in seven went to somebody's own website.
Second correction, the domain table, 15 October 2026. An earlier version of the table below listed ccn.web.tr and developers-google-com.hckseu1-xqad.com among the most cited domains. Neither is a publisher. Opening the cited URLs shows the first is an iframe proxy carrying Google own documentation and the second a reverse proxy mirror of developers.google.com with the same path structure, so both are almost certainly the data provider fetch infrastructure rather than a source any engine chose. They were an artefact of how this was measured. Both are now resolved back to the page they point at, which is why developers.google.com reads 180 here rather than 178. Doing this across the whole dataset moves the vendor own site share from 85.6 per cent to 85.3 per cent, inside what a single date measurement can support, so the headline is left as measured and recorded here instead.
The most cited domains
| Domain | Type | Readings citing it | ChatGPT | Gemini |
|---|---|---|---|---|
| developers.google.com | official or platform | 180 | 178 | 2 |
| support.google.com | official or platform | 138 | 138 | 0 |
| blog.google | official or platform | 125 | 125 | 0 |
| developers.google.cn | official or platform | 122 | 122 | 0 |
| help.openai.com | official or platform | 112 | 111 | 1 |
| github.com | social or community | 110 | 110 | 0 |
| semrush.com | vendor or agency own site | 95 | 84 | 11 |
| developers.openai.com | official or platform | 86 | 86 | 0 |
| tryprofound.com | vendor or agency own site | 61 | 54 | 7 |
| openai.com | official or platform | 59 | 59 | 0 |
| otterly.ai | vendor or agency own site | 58 | 56 | 2 |
| gov.uk | vendor or agency own site | 57 | 56 | 1 |
| bing.com | official or platform | 55 | 55 | 0 |
| ahrefs.com | vendor or agency own site | 54 | 52 | 2 |
| blogs.bing.com | official or platform | 53 | 53 | 0 |
Platform documentation dominates the technical questions, which is unsurprising and probably permanent: on "how does this engine choose its sources", the engine's own documentation is the primary source and nothing is going to outrank it. Note also how one sided that column is. The documentation gravity in this dataset is almost entirely ChatGPT's, toward both OpenAI's and Google's docs. Gemini barely appears at the top of this list at all.
The result we did worst on, published because leaving it out would make the rest worthless
Across 597 readings, our own domain was cited 3 times. Zero in the United States on either engine. Zero on Gemini in any market. One in the United Kingdom and two in India, both on ChatGPT.
We run this measurement and we sell the work that improves it, and at the time of this run we were close to invisible on our own measurement. Any benchmark published by a vendor that happens to show the vendor winning should be read with that in mind, including ours. This one is the baseline we will be judged against when it is re-run.
How it was run, so you can check it
Each question was asked once per engine per market, as a single turn with a live web search, and every source URL the engine returned was recorded. The market was fixed in the prompt text, not by an API location parameter, and that choice was made after measuring the alternative.
The parameter route does not work, and this matters if you try to reproduce this. The API used here accepts a country parameter on these engines and returns a success response for every value, and then ignores it. Asked the same question three times with the country set to the United States, the United Kingdom and India, the answers came back localised to Iowa, to nowhere in particular, and to Mississippi. Setting the market in the prompt text was then tested and does work: the United States prompt answered for the United States, the United Kingdom prompt answered for London in pounds, and the India prompt answered for Mumbai.
What this measurement does not show
- Two engines, not five. Perplexity, Claude and Google AI Overviews were not in this run. Any claim about "AI search" from this data is really a claim about ChatGPT and Gemini.
- One reading per question. These engines do not return the same answer every time. A single reading per question measures a tendency, not a fixed ranking, and a repeat run will differ at the margin. Repeat sampling is the first thing that should be added.
- One date. 15 October 2026. This category changes month to month.
- The market was set in the prompt, not by location. That is weaker than reading from a real connection in each country, and a buyer in that country may see something different.
- Classification is a judgement. Sorting a domain into "platform documentation" rather than "vendor own site" is a decision, and the correction above shows what happens when it is made by an untested pattern. The rules are now one tested file with named examples.
- We sell the service this measures. That is a reason to check the method rather than take the numbers on trust, which is why the questions are published below.
The question set
The 102 questions are grouped in 17 batches: the category by its own names, tools and software, agencies and services, diagnosis, measurement and reporting, technical surfaces, platform specific, the adjacent search buyer, content strategy, reputation and risk, buying and procurement, verticals, and five market specific batches per market. A representative sample:
- What is generative engine optimisation and does a business actually need it?
- What is the best tool to track whether AI assistants mention my brand?
- What software monitors a brand across ChatGPT, Claude, Perplexity and Gemini?
- Why does my company not appear when I ask ChatGPT for recommendations in my industry?
- What sources do AI assistants actually use when recommending a business?
- What is llms.txt and does a website need one?
- Should I allow or block AI crawlers such as GPTBot and ClaudeBot?
- How do I get a page into Google AI Overviews?
- What kind of content do AI assistants cite most often?
- What do I do if ChatGPT says something wrong about my company?
No question named our own company. Naming a brand in the question changes the answer, so the only honest way to measure whether an engine names you is to ask without naming yourself.
Questions about this benchmark
Why only ChatGPT and Gemini?
Those two were reachable through one provider on one method on the day of the run, which keeps the comparison clean. Adding a third engine through a different method would mix two measurements and make the comparison between engines meaningless.
Can I re-run this myself?
Yes, and that is the point of publishing the question set and the method. You will not get identical numbers, because these engines vary between runs. If your result is wildly different from ours we would like to know.
Does 85.6% self published mean reviews are worthless?
It means review directories were 0.8% of what these two engines cited on these 102 questions on this date. Reviews do other useful work. As a route to being cited by ChatGPT or Gemini, on this evidence, they are a minor channel and the advice to prioritise them is not supported.
How often will this be re-run?
The intention is every 60 days, using the same question set, so the numbers are comparable. The run date is on every figure for that reason.
What would change your mind about these findings?
A repeat run with several samples per question showing a materially different source mix, or the same run from real connections in each country showing that the prompt based market lock distorted the result. Both are planned.
Why we published the part that makes us look bad
We asked both engines what would make them cite a vendor's own page despite the obvious commercial interest. Every one of the 102 answers said the same thing in different words: a stated method, a sample size, a date, and the failures shown alongside the wins. Two thirds specifically asked for negative results.
So the three citations are in this page, and so is the correction to our own headline figure. If this measurement is worth anything to you, it is worth something because the numbers that embarrass us are in it. We measure this for a living at AI Knows Us, which is a reason to check the method above rather than to trust the conclusion.