Gemini returned sources on 181 of 291 readings, against 306 of 306 for ChatGPT

Gemini returned a source list on 181 of the 291 usable readings we took from it on 15 October 2026. The other 110 came back with no source at all. Given the same 102 buying questions, in the same three markets, on the same day, ChatGPT returned a source list every time: 306 of 306. For anybody publishing pages that gap matters more than which sites either engine picked, because it decides how many answers there was anything to be cited in.

What a rate of 181 in 291 does to the arithmetic

If the reason you are reading this is to get your business named, the first question about any engine is not which sites it prefers. It is whether a given answer has a source list at all. An answer written without one is drawing on what the model already holds, and a page you publish this month is not in that. An answer with a source list is a slot, and slots can be filled.

On this run ChatGPT offered a slot every time. Gemini offered one in roughly two readings out of three. Three things follow, and they are all practical.

• Judge your Gemini hit rate against the sourced readings. If you count your citations against every Gemini answer you tested, you are dividing by a denominator that includes readings nothing could have won. Report both numbers: answers tested, and answers that carried a source.
• Expect a page to buy you fewer chances on Gemini than on ChatGPT. Same page, same week, smaller surface to land on. That is an argument about where to measure first, not an argument against publishing.
• Treat the unsourced third as a separate problem. Being absent from sourced answers is a question of reach and evidence. Being absent from unsourced ones is about how widely your name was already written about, and it moves slowly whatever you ship this quarter.

The per market numbers, because they are not the same

The three markets did not behave alike, and anyone re-running this should expect the spread rather than one global figure.

Market	Gemini readings that returned a source	ChatGPT, same questions
United States	61 of 102	102 of 102
United Kingdom and London	64 of 102	102 of 102
India	56 of 87 usable	102 of 102
All three	181 of 291 usable	306 of 306

Across all three, Gemini cited our own domain zero times. Not once in the United States, not once in the United Kingdom, not once in India. We sell the work that is supposed to change that number, which is a reason to read the method rather than to take our word for the conclusion.

Gemini is largely missing from the domains that top the combined list

This is the part most likely to be read backwards. The combined most cited list from the run is dominated by Google properties, and it would be natural to assume Google's own assistant is the one leaning on them. On this measurement it was the other way round.

Host	Readings citing it, both engines	ChatGPT	Gemini
developers.google.com	178	176	2
support.google.com	138	138	0
blog.google	125	125	0
help.openai.com	112	111	1
semrush.com	95	84	11
tryprofound.com	61	54	7
otterly.ai	58	56	2

So the documentation gravity in this dataset is ChatGPT's, not Gemini's. Gemini's sourced answers spread themselves much more thinly across marketing and vendor sites, and no single host ran away with them. For a publisher that is the encouraging half of the page: on the readings where Gemini did return sources, there was no Google property sitting permanently at the top of the list to be displaced.

What that means for the pages you write

Across both engines, 85.6% of the 10,652 citations recorded went to the website of the company or agency selling the thing being recommended. Review directories such as G2 and Capterra were 0.8%. That split holds for the Gemini half too, and it points the work in one direction: your own page, answering one buying question, is the thing being cited in this category, and the directory profile is a minor channel.

• Write the page that answers the question as asked, including the market. The question set here was written per market for a reason, and the answers differed by market.
• Carry a date and a last verified line. A date was asked for in all 102 of the conversations where both engines were shown their own answers and asked what would earn a vendor page a citation.
• Quote prices in the reader's currency. All 102 asked for that as well, so a page that quotes one currency to three markets is answering one of them.
• Name the engines you tested. 98 of the 102 asked for the named list rather than the words AI search.
• Keep the facts in server rendered HTML, because a fact that appears only after scripts run is a fact the retrieval step did not read.

Our full method, the published question set and the result we did worst on are in the 597 reading benchmark, which is where these 291 Gemini readings come from. The gap between what gets claimed about assistants in search and what a count actually supports is the subject of artificial intelligence in SEO, what works and what is oversold.

How to take this reading on your own questions

The design for a Gemini source audit was already written up in where a Gemini answer came from, which published the protocol and said plainly that it carried no number yet. This is that number, taken on 15 October 2026 on a larger set, and the method behind it was short.

• The set was 102 questions per market: 72 asked identically everywhere, 30 written for the market itself.
• Nothing in any question named us or a competitor. A prompt carrying a brand returns an answer about that brand, which measures the prompt and not the engine.
• Each reading was a fresh single turn with a live search requested, and every URL that came back was written down.
• Market was set in the wording of the prompt rather than by a location parameter, because the parameter was tested first and the provider ignores it.
• One pass per question, so 181 of 291 is the shape of one day, 15 October 2026, and further passes would move it at the margin.
• Fifteen of the 306 Gemini attempts were unusable, which is why the denominator is 291. The run reached two engines only, so nothing here covers Perplexity, Claude or Google AI Overviews.

Questions about Gemini source rates in this run

Why would Gemini return no source on a third of the readings?

Because it answered those from what the model holds rather than from pages fetched at the time. Google does not publish the rule that decides which way a question goes, so the figure to work from is a measured one, and on these 102 buying questions on 15 October 2026 it was 181 sourced readings out of 291.

Does a missing source list mean Gemini named no companies?

No, and that is why the number matters. An unsourced answer can still name three companies confidently. It simply gives you nothing to point at and no page to compete for, which is a different kind of visibility problem from losing a sourced answer to a competitor.

Is 181 of 291 the rate I should expect on my own questions?

Treat it as the reading for one question set in one category on one date. Questions that turn on current prices, recent changes or a named place pull a search more often than broad category questions do, so a prompt set weighted towards the first kind will come back with a higher sourced share.

Does Google-Extended explain a low Gemini source rate?

That token is a separate control, and it governs whether your content is used by Gemini apps and grounding rather than whether Gemini runs a search for a given question. If your own pages are missing from Gemini answers, check access first, and the dated robots.txt and access control test sets out how to confirm it from your own logs.

Does this tell me anything about Google AI Overviews?

No. AI Overviews is a different surface, reached a different way, and it was not in this run. Gemini readings describe the Gemini product surface only, and a result from one Google surface is not evidence about another.

Where to look next on Gemini

If you take one thing from 181 of 291, make it the habit of recording whether a source list appeared before recording whether you were in it. That single column turns a vague sense of being invisible into two separate, fixable facts. AI Knows Us reads ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews with the brand left out of every question, and reports what each surface cited, so the sourced readings and the unsourced ones stay apart instead of averaging into one misleading score.