How AI Chooses Which Sources to Cite: ChatGPT, Perplexity, Gemini and AI Overviews

How does AI choose which sources to cite? The short answer

AI search engines cite the sources that answer the question most directly, come from an entity they can verify, and are backed up by other independent sources. When an engine searches the web for an answer, as Perplexity does for most questions and ChatGPT, Gemini and Google AI Overviews do for many, it retrieves candidate pages, keeps the passages that best match the question, and cites the pages those passages came from.

When the engine answers from what the model already knows, it may cite nothing at all. So the practical goal has two parts: be the clearest answer on the page an engine retrieves, and be described accurately in the third-party sources the model learns from.

How each engine chooses its sources

ChatGPT answers many questions from what the model already knows and searches the web for others, especially recent or specific ones. When it searches, it cites a handful of pages per answer. In our measurements it draws on a noticeably different set of sources from the other engines: across two B2B categories it shared only 12 to 29% of its cited domains with Claude, Gemini and Perplexity (the research).

Perplexity searches the live web for almost every question and shows its sources as numbered links. It usually cites more pages per answer than the other engines and reflects new or updated pages fastest, which makes it the quickest place to see the effect of changes. For a step-by-step guide, see how to get cited in Perplexity.

Google AI Overviews are built on Google’s own index and ranking systems, so the pages they link to are mostly pages that already rank well and answer the question directly. Getting into AI Overviews is largely a search-ranking job. See how to get cited in Google AI Overviews.

Gemini and Claude also search when they need to, but both answer a share of questions from memory. In one of our payments snapshots, Claude answered 5 of 11 buyer questions without searching in every run, so no website could be cited for them (the research). For those answers, what counts is how the company is described in the sources the model learned from.

AI tools verify entities, not websites

When a user asks ChatGPT or Perplexity to recommend an accounting firm or a GEO agency, the model does not search Google. It cross-references its training data and, where applicable, live retrieval sources against a set of entity verification criteria. The businesses that get cited are the ones that have built clear, consistent authority signals across multiple trusted sources.

Understanding this process is the foundation of effective GEO. The citation decision follows a consistent pattern across platforms — entity verification, authority assessment, query matching, and recency weighting — and each step can be systematically improved.

Entity verification

The first question an AI model asks is whether this entity exists and can be verified. AI tools use Wikidata records, Wikipedia articles, and structured schema markup to resolve entity identity. A business with a clean Wikidata record, accurate schema markup, and consistent name and contact data across the web is significantly easier to verify — and therefore cite.

Businesses without a Wikidata entry exist in ambiguity. The model may know your name from web content, but it cannot verify whether the accounting firm mentioned on Reddit is the same entity as the website it found in training data. Ambiguous entities get cited less. Verified entities get cited more.

Authority assessment

Once an entity is verified, the model assesses authority by cross-referencing how many credible, independent sources mention this entity and in what context. The most-weighted sources are Reddit, LinkedIn, Wikipedia, review platforms including Trustpilot, G2, and Clutch, and editorial publications. As of March 2026, LinkedIn rose to the second most-cited domain overall for professional queries, with citation frequency doubling between November 2025 and February 2026.

Businesses that appear consistently and positively across these sources accumulate authority that AI tools recognise. The pattern is not about volume — it is about consistency and credibility. Ten genuine LinkedIn posts on your area of expertise, verified on Wikidata and mentioned in two editorial publications, outperforms 500 low-quality directory listings.

Query-to-entity matching

Citation is not just about authority — it is about relevance to the specific query. An AI model answering "best employment lawyer in London" needs to match that query to entities verified as employment lawyers operating in London. Businesses with clear, specific category positioning — reinforced through schema knowsAbout declarations, Wikidata property assignments, and consistent third-party descriptions — match queries more reliably.

Vague positioning hurts AI citation rates. A business described as "a full-service professional services firm" matches fewer queries than one consistently described as "a specialist employment law firm for technology companies in London." Specificity in your entity signals directly improves query matching accuracy.

Will AI use your PR or your blog posts as sources?

Usually the page that answers the buyer's question best, which is rarely a press release. Engines that retrieve live pages pick passages that match the question closely. A press release announcing a funding round or a partnership seldom matches a question like “which payments platform is best for marketplaces?”, so it is seldom the passage that gets used.

PR still matters, but indirectly. Its value for AI visibility is the independent coverage it produces: an article in a publication the engine already trusts that describes what you do. That coverage corroborates your entity and can be cited in its own right. A release that nobody picks up adds little.

Blog posts are cited when they work as answers. A post that opens with a plain, specific answer to one buyer question, under a heading phrased like that question, is far more extractable than a post that builds up to its point. The test is simple: if someone copied only your first paragraph into an answer, would it be correct and complete?

What makes content more likely to be cited

Across engines, the pages that get used share a handful of traits. They answer the question in the first one or two sentences. They break the topic into short, self-contained sections with question-style headings, so any one section can stand alone. They state specific facts rather than general claims, and they show who wrote them and when they were last updated.

Technically, they are readable without JavaScript, open to AI crawlers in robots.txt, and marked up with structured data such as Organization and FAQPage. Off the page, the company behind them is described the same way on review sites, directories and publications. None of these guarantees a citation; each removes a reason for an engine to pass you over. To check the technical side of your own site, run our free AI-readiness checker.

Recency signals

AI tools weight freshness. Content under three months old is three times more likely to appear in AI citations than equivalent older content. Pages with recently updated dateModified schema, fresh content, and recent third-party mentions are cited more frequently than equivalent pages with stale signals.

This means GEO is not a one-time project. A Wikidata entry established in 2024 remains valuable, but the businesses winning category positions in AI citations are the ones continuously publishing fresh content, maintaining active LinkedIn presence, and earning new editorial mentions. Authority compounds. So does recency. For what to expect at each stage, see how long GEO takes to show results.