How AI Search Engines Pick Their Sources
ChatGPT, Perplexity, and Google AI Overviews don't pick sources the way you'd expect. Here's how retrieval, training data, and citation binding actually work.
AI search engines pick sources through two separate systems working together: retrieval, which fetches live pages from the web at query time, and training data, which is the model's baked-in knowledge from before its cutoff. For any question about current recommendations, prices, or comparisons, retrieval does almost all the work. And within retrieval, there's a distinction most guides skip past: being fetched is not the same as being cited, and being cited is not the same as being mentioned.
This post covers the mechanics underneath the tactics. If you want the step-by-step playbook, read how to do GEO; this is the "why it works" that makes the playbook make sense.
Two systems, doing different jobs
Training data is what the model absorbed before its knowledge cutoff, compressed into its weights. If your brand was well documented across the web for months before that cutoff, the model has some baseline sense of you, the way it "knows" any well-established entity. This layer is slow to change and rewards long-term, consistent presence.
Retrieval is what happens when the engine searches the live web to answer a question. This is where nearly all commercial and recommendation queries get answered, and it's the layer GEO work actually targets. The model issues one or more searches, fetches a set of pages, and builds an answer from what it finds.
One nuance retrieval studies keep surfacing: engines often break a complex question into several narrower sub-questions and search each independently before combining the results. "Best CRM for a 20-person remote team" doesn't run as one search; it fans out into something like CRM feature comparisons, pricing pages, and team-size-specific reviews, each retrieved separately. Content that only answers the whole question, and ignores the sub-questions, misses citations it should have won.
Fetched, cited, and mentioned are three different outcomes
This is the distinction most advice collapses into one thing, and it changes how you should read your own analytics.
Fetched means the engine's crawler pulled your page during retrieval. It doesn't mean you'll appear anywhere in the answer.
Cited means a specific sentence in the answer is attributed to your page, usually as a clickable footnote tied to one claim. Citation binds to a precise piece of text, not to general topical relevance. Being broadly relevant to the question isn't enough; you have to be the best support for one specific claim the model is making.
Mentioned means your brand name appears somewhere in the answer, sometimes as a linked chip, without being the cited source of any particular claim. You can be mentioned without being cited (the model knows of you from training data or a different source) and cited without being prominently mentioned (a footnote nobody reads).
One independent analysis of ChatGPT's own network traffic found a stark gap between fetch rates and citation rates for different content types: Reddit and YouTube pages were fetched at similar volumes, but Reddit converted to citations far more often. The likely mechanical reason is that when the engine fetches a Reddit thread, it gets the actual text of the discussion; when it fetches a YouTube page, it typically gets metadata, not a transcript. There's nothing to bind a citation to. This is a big part of why video content underperforms text for citations even when it's heavily viewed and topically relevant, and it's a reason to publish transcripts alongside any video you make.
What decides the citation, once a page is fetched
Four things consistently correlate with a fetched page converting into a cited one.
Extractability. A citation has to bind to a clean, self-contained claim. Pages with a direct-answer passage near the top, ideally under 80 to 120 words and free of hedging, convert to citations at meaningfully higher rates than pages that bury the answer in a long introduction. One structural detail worth knowing: the majority of cited answer passages in one large audit contained no outbound links at all. A link-free, self-contained block reads as the authoritative answer itself; a block full of links reads as a signpost to somewhere else, which seems to make models treat it as less quotable on its own.
Specificity. Vague claims don't bind to anything worth citing. "We're the best marketing tool" attaches to nothing. "AI referral traffic converted 4.4x better than organic across 12,500 tracked queries" is a specific, attributable claim a model can lift and defend.
Format. Comparative content, listicles and "X vs Y" structures, accounts for a large share of citations across multiple independent audits, more than any other single content type. Tables get lifted almost verbatim because the structure itself does the work of organizing the claim.
Consensus. For recommendation queries specifically, the model is looking for agreement across the sources it retrieved. If your brand shows up in four of the six pages retrieved for a query, you get named with confidence. If you show up in one, the model hedges or skips you. This is the mechanic behind why Reddit threads and third-party listicles often matter more than your own site: your own domain is one vote, and the model is counting votes across a set. We go deeper on building that consensus deliberately in how to get your brand recommended by ChatGPT.
Why engines differ from each other
The four engines you'd typically track don't retrieve the same way, and treating them as one channel is a common mistake.
ChatGPT leans on a hybrid retrieval setup, part live web search, part licensed data partnerships, that shifts over time and isn't fully disclosed. What's stable is the outcome: a small set of domains, especially Reddit and major review or comparison hubs, dominate citations for commercial queries, and citation is heavily concentrated in extractable, specific text.
Perplexity retrieves on essentially every query and shows its full source list by default, which makes it the most transparent engine to study directly. It leans even harder on community sources like Reddit for recommendation-style questions.
Google AI Overviews draws disproportionately from pages outside its own organic top ten, which is the most counterintuitive fact in this whole space. A page that never cracked page one of Google can still be the quoted source in an AI Overview above it, because the citation model and the ranking model aren't optimizing for exactly the same thing.
Claude tends toward a more conservative citation pattern, favoring structured, well-sourced content and skewing toward fewer, higher-confidence citations rather than a wide spread.
The practical implication: run your prompt audit across all of them, not just ChatGPT, because a brand that's well cited in one engine can be nearly invisible in another for the same query.
What this means for what you publish
Three concrete decisions fall directly out of the mechanics above. Publish transcripts for any video or audio content, since untranscribed media gets fetched but rarely cited. Write comparison and listicle content deliberately, since format alone shifts citation odds. And treat your own domain as one input among several rather than the whole strategy, since consensus across sources matters more than optimizing any single page in isolation. The checklist for turning this into a repeatable weekly and monthly practice is in the GEO checklist.
Running the audit yourself
You don't need special tooling to see this in action. Pick five buying-intent questions in your category, run each in a fresh ChatGPT and Perplexity chat with search on, and expand the sources. Note which pages got cited, whether the citation is bound to a specific sentence or feels like a general mention, and what format that page uses, direct answer, table, listicle, and where it came from, owned domain, review platform, forum. Do this for your own brand and for the competitor who's winning the query. The gap is usually visible in the first hundred words of whichever page won.
Okara's GEO agent runs this exact comparison continuously across ChatGPT, Perplexity, Gemini, and Claude, and the writer and Reddit agents produce the extractable, consensus-building content this mechanic rewards. Add your site at okara.ai to see where you currently stand.
FAQ
Does ChatGPT crawl the web? Yes, through a retrieval layer that combines live web search with licensed data sources, though the exact mix isn't fully disclosed and shifts over time. This is separate from the model's training data, which is fixed as of its knowledge cutoff.
Why does Perplexity cite different sources than ChatGPT for the same question? Each engine runs its own retrieval and ranking logic. Perplexity retrieves on nearly every query and leans hard on community sources like Reddit; ChatGPT's mix varies by query type and includes licensed partnerships. Track both separately rather than assuming one predicts the other.
Can AI engines read JavaScript-heavy sites? Most AI crawlers, including the major ones from OpenAI, Anthropic, and Perplexity, read raw HTML and don't execute JavaScript. A page that renders fully in a browser can return as an empty shell to these crawlers. Confirm your key pages serve real content in their raw HTML response.
Does schema markup affect AI answers? It helps engines parse structure and entity identity reliably, but controlled testing has found schema alone doesn't directly lift citation rates. Treat it as the reliability layer under good content, not a citation tactic by itself.
Why would a page get fetched but never cited? Usually because the content couldn't bind to a specific claim, no self-contained answer passage, no specific statistic, or because a competing source answered the exact same claim more precisely. Fetching is necessary but not sufficient; citation requires the page to win a specific sentence, not just be relevant to the topic.