ChatGPT citations, counted: the source mix behind 306 readings on 15 October 2026
All 306 of the ChatGPT readings came back with a source, and six of the eight hosts at the top of the list are a platform describing its own product.
Published by AI Knows Us · Updated 15 October 2026
ChatGPT returned at least one source on every one of 306 readings taken on 15 October 2026, from 102 buying questions put to it in each of three markets, the United States, the United Kingdom and India, with every source URL written down. The hosts it cited most often were documentation: developers.google.com appeared in 176 of the 306 readings, support.google.com in 138, blog.google in 125, help.openai.com in 111 and developers.openai.com in 86. Across both engines in that run, 85.6% of the 10,652 citations recorded went to the website of the company or agency selling the thing being recommended, and review directories such as G2 and Capterra accounted for 0.8%.
Every ChatGPT reading had a source slot in it
Two engines were read, ChatGPT and Gemini, and the two behaved differently enough that a combined figure hides the useful part. Gemini, asked the same 102 questions in the same three markets on the same day, returned a source list on 181 of its 291 usable readings. ChatGPT returned one 306 times out of 306.
That difference is the arithmetic of the exercise. A reading with no source list cannot be won with a page, because there is no page in it to be. A reading with a source list can be. Every ChatGPT reading was in the second group, which makes it the engine where a page you publish this month has somewhere to land.
The hosts ChatGPT cited most often, counted in readings
The unit here is readings, not citations. A host counts once per reading that cited it, however many of its URLs appeared, so the middle column reads as appeared in this many of 306.
| Host | Readings citing it, of 306 | What it is |
|---|---|---|
| developers.google.com | 176 | Google Search Central and developer documentation |
| support.google.com | 138 | Google product help |
| blog.google | 125 | Google's own announcements |
| developers.google.cn | 122 | Google developer documentation, China domain |
| help.openai.com | 111 | OpenAI help centre |
| github.com | 110 | Code and specification repositories |
| developers.openai.com | 86 | OpenAI developer documentation |
| semrush.com | 84 | A vendor in the category, citing its own research |
Read the top of that table as a closed door. Six of the eight hosts are a platform describing its own product. On a question such as how an engine picks its sources, or which crawler token controls what, the platform documentation is the primary source and no page any of us writes will displace it. That is not a problem to solve. It is a question type to stop spending money on.
Two of the hosts near the top are not the platforms' own
Worth recording because it was not expected. Alongside the official documentation, ChatGPT also cited developers-google-com.hckseu1-xqad.com in 75 readings and developer-openai-com.sitemirror.store in 48. Neither hostname belongs to Google or to OpenAI, and both sit inside the top twenty of the combined domain list from this run.
The reading is narrow and useful: a host does not have to be the authority to be returned as one. What comes back is what the retrieval step found, and on this date it found copies.
The two kinds of question, and only one of them is winnable
The 102 questions sat in 17 batches, and the source mix splits them in two.
- Platform questions. What is llms.txt. Should I allow GPTBot. How does an engine decide what to cite. These went to documentation, and the counts above are what that looks like. Answer them on your own site for your own readers, and expect the citation to go to the platform.
- Buying questions. Which tool tracks whether assistants mention my brand. Which agency does this work in my city. What does it cost. This is where the 85.6% self published share lives, because there is no official documentation for who the best supplier is. A company's own page is the only thing in the field, and on this evidence it is being cited.
The standing advice in this category is to go and collect directory reviews. On this run the directories were 0.8% of what the two engines cited. If you want the category map before the engine detail, what AI SEO is, and the whole landscape sets out the terrain these questions sit in, and the mechanics of how ChatGPT reaches a page at all are in ChatGPT SEO, how ChatGPT picks sources.
What ChatGPT cited of ours, which was almost nothing
Across all 597 usable readings in the run, both engines together, our own domain was cited 3 times. On ChatGPT that was zero in the United States, one in the United Kingdom and two in India. On Gemini it was zero in every market. We publish measurement of this kind and we sell the work that improves the number, so the baseline belongs on the page rather than in a drawer: the full method, the published question set and that result are in the 597 reading benchmark, which is where these 306 ChatGPT readings come from.
How the 306 readings were taken
- 102 questions per market, 72 identical across all three plus 30 written for the market.
- No question named our company or any vendor. A prompt with a brand in it pulls the answer towards that brand, and the reading then measures the prompt rather than the engine.
- One question per reading, asked 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. The country parameter on the provider used here is accepted and then ignored, which was measured first.
- One reading per question, so these counts describe a tendency on 15 October 2026 rather than a fixed ranking.
- Two engines were in the run, ChatGPT and Gemini. The other three surfaces our product covers were not read, so this page says nothing about Perplexity, Claude or Google AI Overviews.
What to publish if you want ChatGPT to cite you
- Pick the buying question, not the platform question. One page, one question, with the answer in the first two sentences.
- Put a date on every figure, and a last verified line on the page. When we showed both engines their own answers and asked what would make them cite a vendor page, all 102 of those conversations asked for a date.
- Price in the reader's currency. All 102 asked for that too, and a page quoting rupees to a buyer in Ohio has answered somebody else's question.
- Name the engines you tested rather than writing AI. 98 of the 102 asked for the named list instead of the category word.
- Give the sample size and show what failed. 80 of 102 asked for the N, and 68 asked for negative results.
- Keep the facts in server rendered HTML. A price that appears only after scripts run is a price the retrieval step did not read.
Questions about ChatGPT citations in this run
Did ChatGPT return sources on all 306 readings?
Yes, on this question set and this date. Each of the 306 readings recorded a live search and at least one source URL. That is one of the cleanest results in the run and also one of the easiest to re-check, because the question set is published with the benchmark.
Does 85.6% self published mean my own page can be the source?
On the buying questions, yes, and that is the finding worth acting on. At 85.6% the field in this category is mostly sellers describing themselves, which means your page is competing with other vendor pages and not with an independent authority. What separates a self published page that gets cited from one that does not is whether it carries a date, a stated method, a sample size and a price.
Why is Google's documentation at the top of a ChatGPT list?
Because a large share of these 102 questions are questions about how search and assistants work, and the documentation is the primary source for those. It is also the clearest signal in the table about where not to compete. Note which engine was doing the citing: Gemini contributed 2 of the 178 readings that cited developers.google.com and none of the 125 that cited blog.google.
Should I chase G2 and Capterra listings to get cited by ChatGPT?
Review directories were 0.8% of everything the two engines cited on these 102 questions on 15 October 2026. Listings do other useful work for a business, and they belong in a plan for other reasons. As a route into a ChatGPT citation, a quarter spent on directory profiles is a quarter spent on 0.8% of the field.
Does any of this describe Perplexity, Claude or Google AI Overviews?
No. The run covered ChatGPT and Gemini, read through one provider on one method so the comparison between the two stays clean. Adding a third engine through a different method would mix two measurements. Those three surfaces need their own readings, and the numbers here do not stand in for them.
Re-reading your own source mix
The number here worth copying is not 176 or 86. It is the habit: ask your buying question blind, write down every host that came back, and count how many you could realistically be. The list splits into documentation nobody will beat, vendors you compete with, and three or four pages you could join this week. AI Knows Us takes that reading across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews and lists what each one cited per surface. The 306 of 306 on this page is ChatGPT alone, and a count like it is what turns into a decision about what to write next.
Correction to the domain table, 15 October 2026. An earlier version of this table 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 is 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 the engine chose. They are an artefact of how this was measured, not a finding, and both rows have been removed. Resolving every proxy back to the real page it points at moves the vendor own site share from 85.6 per cent to 85.3 per cent, which is inside the variation a single date measurement can support, so the headline figure is left as measured and this note records the difference instead.