How ChatGPT, Perplexity, Gemini and Google AI Overviews choose their sources

The shared pipeline: search first, write second

None of these engines answer a “who should I use” question from memory alone. They search, then write, and the citations are the pages they used.

OpenAI describes it plainly: ChatGPT search “typically rewrites your query into one or more targeted queries” that it sends to search providers, and “may send additional, more specific queries”. Google’s documentation says AI Overviews and AI Mode “may use a ‘query fan-out’ technique — issuing multiple related searches across subtopics and data sources — to develop a response”. In AirOps’ study of 16,851 ChatGPT queries, 88.6% triggered exactly two fan-out searches.

Two consequences follow. First, you are not competing for the buyer’s question; you are competing for the narrower searches the engine invents from it. Second, being found is only half of it. In the same study, 58% of pages ChatGPT retrieved were never cited for any query they appeared in, and a page in position 1 of ChatGPT’s own retrieval results was cited 58% of the time against 14% at position 10.

Engine by engine

Engine	Where candidate pages come from	What controls access	What the vendor says you need
ChatGPT search	Rewritten queries sent to third-party search providers (Microsoft is the one OpenAI names) plus content from partners	OAI-SearchBot in robots.txt; GPTBot is training only	Let OAI-SearchBot in; blocked sites can still appear as plain navigational links
Google AI Overviews and AI Mode	Google’s own index, via fan-out searches	Normal Googlebot indexing, snippet controls, and since 31 August 2026 a Search Console setting that can exclude a site	“No additional requirements… nor other special optimizations” beyond being indexed and eligible for a snippet
Gemini app	The Google Search index, used to ground answers at prompt time	Google-Extended covers Gemini training and Gemini Apps grounding	Nothing special; but blocking Google-Extended can remove you from Gemini answers while leaving Search untouched
Perplexity	Its own crawl plus live fetches	PerplexityBot (search, not training); Perplexity-User fetches on request and “generally ignores robots.txt”	Let PerplexityBot in; changes take up to 24 hours
Microsoft Copilot	Bing: Copilot “generates a search query that it sends to the Bing search service”	Bingbot and Bing indexing	Bing Webmaster Tools now reports Copilot citations and the grounding queries behind them

Three details in that table catch people out.

OpenAI’s crawlers are independent. “A webmaster can allow OAI-SearchBot in order to appear in search results while disallowing GPTBot.” Blocking training does not have to mean disappearing from ChatGPT.

Google has added an off switch. As of 31 August 2026, every Search Console property has a generative AI control that includes or excludes the site from AI Overviews, AI Mode and Discover’s AI features. Include is the default. If you are absent from AI Overviews and nobody can explain why, check that setting before anything else.

Gemini and AI Overviews are controlled separately. Google-Extended “does not impact a site’s inclusion in Google Search”, but it does govern grounding in Gemini Apps. A site that blocked Google-Extended to opt out of training may have opted out of Gemini answers too.

What the independent studies agree on

The engines cite different things. Writesonic’s 161,286-prompt study found 3.8% of cited domains were cited by ChatGPT, Gemini, Perplexity and AI Overviews together, while 72–73% appeared on exactly one engine. Even Google’s two features diverge: Ahrefs found AI Mode and AI Overviews cited the same URLs only 13.7% of the time. In Australia, when four engines were asked for the best AEO agencies they named 24 different firms.

Ranking and citation have partly come apart, but the size is disputed. Ahrefs measured 37.9% of AI Overview citations ranking in the top ten in early 2026, down from about 76% in July 2025, though it changed its parsing method in between. BrightEdge puts the figure at about 17% and roughly flat for a year. For the assistants, Ahrefs found only 12% of cited URLs ranked in Google’s top ten for the original prompt: 28.6% for Perplexity, about 8% for ChatGPT, Gemini and Copilot. Quote the range, not one number.

Mentions beat links. In ChatGPT, AirOps found domain authority and backlinks showed no positive correlation with citation. Ahrefs, looking at AI Overviews across 75,000 brands, found brand web mentions correlated with visibility at 0.664 against 0.218 for backlinks, and warns that correlation is not causation. Both point the same way as the earned-media research: what other sites say about you matters more than what links to you.

Recently updated pages are favoured. Seer Interactive found 75% of pages cited by AI engines in 2026 had been updated in the past year (Gemini 78%, ChatGPT 73%, Perplexity 65%). The effect came from updates to existing pages, not from publishing new ones.

Source preferences shift without warning. Semrush tracked ChatGPT citing Reddit in close to 60% of responses in early August 2025 and around 10% by mid-September. The often-quoted “Wikipedia is 47.9% of ChatGPT citations” is a misreading: that was Wikipedia’s share of ChatGPT’s top ten sources in Profound’s data; its share of all ChatGPT citations was 7.8%.

Answers are close to random run to run. SparkToro’s 2,961 runs found less than a 1-in-100 chance of getting the same list of brands twice. Any tool that reports a single “AI ranking position” is reporting noise. The usable measure is how often you appear across many runs of a fixed prompt set, which is how our citation index is built.

What this means in practice

1. Let the right crawlers in. Allow OAI-SearchBot, PerplexityBot, Bingbot and Googlebot. Decide about GPTBot and Google-Extended separately, knowing the Gemini consequence.
2. Check the Search Console AI setting is on include.
3. Be properly indexed in Bing, not just Google. Copilot depends on it and Microsoft’s infrastructure sits behind part of ChatGPT search. Bing Webmaster Tools and IndexNow cost nothing.
4. Earn mentions on the pages engines retrieve for your buyers’ questions. A citation map tells you which those are.
5. Keep your cited pages current. Update them; do not spin up new ones for every variation of a question. Google’s guide says creating content for fan-out queries “primarily to manipulate rankings or generative AI responses” violates its scaled content abuse policy.
6. Measure each engine separately, on a fixed prompt set, repeatedly. The free check is the hand-run version; how to get cited by ChatGPT turns this into steps.