What sources get cited by AI search engines, measured on 602 controlled prompts.

The count varies by more than a factor of two between engines, 6.88 citations per response for ChatGPT, 12.06 for Google's AI surfaces and 16.35 for Perplexity, and the count is the less important half of the finding.¹ The more important half is mean influence per fetched page: 0.2713 for ChatGPT against 0.0584 for Google and 0.0646 for Perplexity, roughly a four-fold difference in how much of the finished answer a single source accounts for.

A citation is therefore not a fixed-value unit across engines, and any report that adds citations from different engines into one total is adding quantities with different meanings. Being one of seven sources on an engine that leans hard on each one is a materially different outcome from being one of sixteen on an engine that samples widely and shallowly. Set the target per engine, and say which outcome you are aiming for.

This also reframes what a low citation count means. An engine citing fewer sources is not being stingy, it is concentrating, which makes it higher variance for a brand: you are either substantially inside the answer or absent from it. The wide-sampling engine offers thinner, more frequent appearances. Neither is better in general.

Not the ones cited most often. The same dataset also scores mean influence by domain type, and that ranking does not match the share table above: encyclopedia sources average 0.2144, commercial sources 0.1028, news media 0.0726 across 1,546 citations, last of the seven types reported.¹ News is 31.17% of ChatGPT's citations and near the bottom for how much of the answer it accounts for, which the authors read as engines reaching for explanatory pages after timely ones.

A second cut scores the role a passage plays rather than the kind of site it sits on. Definition passages average 0.1531 and comparison passages 0.1524, against 0.0801 for background and 0.0529 for reference-only citations, roughly a third of a definition. What a page is for inside the answer matters more than what hosts it: a commercial page that supplies a clean definition is doing the job an encyclopedia usually does.

A minority of it on most commercial questions, even though official sources are the largest single category overall. The 34–46% official share is spread across every brand in the answer, competitors included, and the rest is news, vertical publications and discussion threads you do not control. A 4,706-query academic audit of Google AI Overviews, collected September 2025 in the US and Germany, found 53% of the domains it consults were absent from the organic top 10 and 27% from the top 100.⁵ A 55,393-query preprint collected between 13 March and 21 April 2026 measured the comparable first-page gap at 29.8%, so the defensible statement is a range of roughly 30% to 53% with both collection windows named, not a point from either study.⁸ An 11,500-query SIGIR 2026 study makes the point from another angle: for one query, the sources returned by Google organic search, AI Overviews and Gemini overlap at under 0.2 average Jaccard similarity.¹⁰ Kirsten counts domains consulted, not domains cited.

Two consequences follow. The first is that the pages worth writing are the ones that can be selected on their own merits: documentation that answers a real sub-question, a pricing page that states prices in text, a comparison page honest enough to be quotable. Google is blunt about what qualifies a page at all: it must be indexed and eligible to be shown in Search with a snippet, and there is no special structured data to add.⁶ Its optimization guide adds that these features are “rooted in our core Search ranking and quality systems”.⁹

The second is that being fetchable is a precondition, not a detail. A 2026 preprint, reported via the critical survey, found 27.1% of URLs already cited in AI answers could not be scraped, being inaccessible, removed or non-textual, and roughly 16% of the 19,154 pages retrieved were labelled AI-generated by its detector.⁷ A page behind a login, an interstitial or a bot-blocking gate is out of the running however well it is written, the subject of Technical GEO.