How ChatGPT, Perplexity and Gemini each pick their sources

The same question, asked of three AI assistants, will often produce three different sets of sources. Not because one is right and the others wrong, but because they retrieve differently. Knowing how each behaves tells you where your effort actually lands.

One caveat before the detail: none of these companies publish their retrieval logic, and all of them change it. What follows is drawn from observed behaviour, not documentation. Treat it as a working model, not a specification.

Perplexity — search-first, citation-heavy

Perplexity behaves most like a search engine with a writer attached. It retrieves live results for nearly every query and cites visibly, with numbered sources next to the claims.

What it favours: pages that answer the specific question directly, recency, and clean extractable structure. Because citations are shown, there is a visible payoff for being quotable rather than merely relevant.

What this means for you: this is the platform where content shape matters most. A self-contained paragraph that resolves a question outperforms a longer page that circles it. It is also the fastest platform to reflect changes — useful when you want to measure whether a fix worked.

ChatGPT — mixed retrieval, brand-weighted

ChatGPT answers some questions from training and others by searching live, and the boundary is not announced to the user. For anything time-sensitive or specific it retrieves; for general knowledge it often does not.

What it favours: when it retrieves, established sources and clear entity signals. When it does not retrieve, whatever was consolidated during training — which is where brand mentions across many independent sites pay off, because that is what made it into the model in the first place.

What this means for you: two separate jobs. Crawler access and page structure serve the retrieval path. Entity building across external sites serves the other one. Neither substitutes for the other.

Google AI Overviews — the classic index, distilled

AI Overviews sit on top of Google's existing index and infrastructure. If you are invisible in Google's index, you are invisible here.

What it favours: everything classic SEO already favours — crawlability, structured data, topical depth, site quality signals — plus a strong preference for content that maps cleanly onto a question.

What this means for you: this is the platform where GEO and SEO overlap most. Structured data does real work here. So does having your facts stated as facts rather than implied by marketing copy.

Gemini — knowledge-graph leaning

Gemini draws heavily on Google's entity understanding. It tends to be conservative about recommending things it cannot confirm exist as distinct entities.

What it favours: confirmed entities with consistent attributes across sources. This is the platform where the consistency problem hurts most — contradictory data about your business is worse here than almost anywhere.

Bing Copilot — index freshness matters

Copilot runs on Bing's index, which most sites neglect entirely. That neglect is an opportunity: the competitive field is thinner.

What this means for you: submit your site to Bing's webmaster tools and use the instant-indexing protocol so new pages register in hours rather than weeks. It is a small task most competitors skip.

The work that pays across all five

Chasing each platform separately is a poor use of time. Four things pay off everywhere:

1. Let the crawlers in. Nothing else matters if this is wrong, and it is wrong more often than people expect.
2. State facts in machine-readable form. Structured data is read by every system here.
3. Shape content as answers. Helps retrieval-based systems directly and everything else indirectly.
4. Be consistently described in several places. Helps entity-based systems directly and retrieval-based systems as a quality signal.

Platform-specific tuning is worth doing after those four are solid — and not before.