ExplainerChatGPT, Gemini & co.
How AI Search Engines Choose Their Sources
by the AEO GEO Labs team5 min read
How AI search engines choose sources comes down to a two-stage process: a search step retrieves candidate pages from an index, then the model reads passages from those pages and cites the ones it actually uses to write the answer. You can't buy your way into either stage. You can make sure you are crawlable, indexed, and holding the clearest passage for the question.
This is a supporting piece for our LLM SEO guide. Below we stick to what each engine has documented, as of October 2026, and label the parts that are inference.
The shared pattern: retrieve, read, cite
Every major AI search engine uses retrieval-augmented generation (RAG): it searches first, then writes from what it found. Google's generative AI optimization guide describes it directly: its systems use core Search ranking to "retrieve relevant, up-to-date web pages from our Search index," then review specific information from those pages to generate the response with links.
The common steps across engines:
- Rewrite the question into searches. Engines often run several related queries, which Google calls query fan-out.
- Retrieve candidates from a web index, either the engine's own or a partner's.
- Rank passages, not just pages. The model needs a specific span that answers the sub-question.
- Write the answer and attach citations to the sources it drew on.
Because step 3 works at passage level, a page ranking eighth can be cited over the page ranking first if its paragraph answers more directly.
How AI search engines choose sources, engine by engine
Each engine documents a different part of the pipeline. This table summarizes what is public as of October 2026.
| Engine | Where candidates come from | What's documented | How to opt out |
|---|---|---|---|
| Google AI Overviews and AI Mode | Google's Search index | Core ranking systems plus query fan-out | nosnippet, or Search Console setting |
| ChatGPT search | OpenAI's crawler index plus third-party search providers | OAI-SearchBot must be allowed for full citations | Block OAI-SearchBot |
| Perplexity | Perplexity's own index | Hybrid retrieval, passage-level reranking | Block PerplexityBot |
| Microsoft Copilot | Bing's index | Grounding queries reported in Bing Webmaster Tools | NOARCHIVE meta tag |
| Claude | Anthropic's web search tool | Claude decides when to search; citations are always on in the API | Block Claude-SearchBot |
Google says AI Overviews and AI Mode are "rooted in our core Search ranking and quality systems." A page must be indexed and eligible for a snippet. Fan-out widens the pool, so pages ranking for related sub-questions can be cited even if they don't rank for the query typed.
ChatGPT
OpenAI's crawler documentation says OAI-SearchBot is used "to surface websites in search results in ChatGPT's search features," and blocked sites won't appear in ChatGPT search answers beyond navigational links. OpenAI's publisher FAQ adds that ChatGPT may rewrite a prompt into several queries and also use third-party search providers.
Perplexity
Perplexity runs its own index. In its write-up on the Perplexity Search API, it describes tracking over 200 billion URLs, combining lexical and semantic retrieval, and reranking at the level of sections and spans rather than whole documents.
Microsoft Copilot
Copilot grounds answers in Bing. Microsoft's AI Performance report in Bing Webmaster Tools, launched in February 2026, shows publishers their citations and the "grounding queries" the AI used to retrieve their content.
Do AI engines cite the same pages that rank in Google?
Often not. Ahrefs' 2025 study of 15,000 queries found only 12% of links cited by ChatGPT, Gemini and Copilot appeared in Google's top 10 for the same prompt. Perplexity overlapped most, at 28.6%.
Our reading, labeled as inference: engines with their own index or a different search partner see a different web than Google's top 10, and fan-out queries pull in pages ranking for sub-questions. Ranking well in Google helps with Google's AI features, but it doesn't guarantee citations elsewhere. That gap is the core of how GEO differs from SEO.
What makes a passage more likely to be chosen
Engines don't publish source weights, so the factors below combine documented guidance with published research.
- Access. A page blocked to the engine's search crawler can't be retrieved. Check yours with our AI Crawler Checker.
- A direct, self-contained answer. Microsoft's October 2025 guidance recommends concise one-to-two sentence answers and warns against long walls of text and answers hidden in tabs.
- Specific, verifiable content. The GEO paper (KDD 2024) found that adding citations, quotations and statistics could raise visibility in generative engine responses by up to 40% on its benchmark.
- Non-commodity content. Google says its AI systems look across many sources, so a unique viewpoint helps a page stand out.
- Freshness for time-sensitive topics. Microsoft lists keeping content current among its recommendations for AI answers.
The practical playbook is in our guide on how to earn AI citations.
Frequently asked questions
Can I pay an AI search engine to cite my website?
No engine sells organic citations as of October 2026. Some run publisher revenue-share programs, such as Perplexity's Publishers' Program, and ads are a separate product. Citations in the answer itself come from retrieval and ranking. The levers you control are crawler access, indexing, and having the clearest passage for the question.
Why does ChatGPT cite different sources than Google?
ChatGPT retrieves from OpenAI's own crawl and from third-party search providers, and may rewrite the prompt into several searches. That produces a different candidate pool than Google's top 10. Ahrefs found only about 8% of ChatGPT's cited links appeared in Google's top 10 for the same prompt in its 2025 study.
Do AI search engines prefer big brands?
Partly. Industry studies find that brand mentions across the web correlate with AI visibility, and widely referenced sites appear often. But engines rank passages, so a precise answer on a smaller site can beat a vague one on a large site. As of October 2026, no engine has published a rule that favors brand size directly.
Does schema markup decide which sources AI engines choose?
No public evidence says schema decides citations. Google states that structured data isn't required for its AI features. Microsoft recommends JSON-LD to help its systems interpret content. Our view: accurate schema helps machines read your page, but the visible text is what gets quoted.
If you want this done for your site, AEO GEO Labs runs GEO, AEO and SEO programs for B2B and SaaS teams. See our services.