How AI search engines choose what to cite

Stage 1: Retrieval — getting into the candidate set

Before anything is cited, a search happens. Perplexity retrieves on essentially every answer; ChatGPT, Gemini, Claude, and Copilot retrieve when their search tools engage; Google AI Overviews synthesize from the Google index. Retrieval looks like search because it mostly is search: query terms, freshness, authority, and crawlability decide the candidate set.

Two implications get missed. First, engines rewrite queries — a buyer's vague question fans out into several sharper sub-queries, so pages that answer the sharp version ("X pricing 2026", "X vs Y for teams") get retrieved for the vague one. Second, crawl failures are silent: a page AI crawlers can't fetch or render simply doesn't exist at this stage, no matter how good it is. That's the case for checking AI readability before anything else.

Stage 2: Extraction — containing the liftable passage

From each candidate page, the engine pulls passages relevant to the question. This is where structure pays:

• Answer-first sections. An H2 that states the question, followed immediately by the direct answer, is a machine-liftable unit. The same answer buried in paragraph six of a narrative isn't.
• Specific claims. "Costs $49/month for 5 seats" beats "affordably priced for teams". Engines quote precision; they paraphrase vagueness — usually without attribution.
• Tables and lists. Comparisons and steps survive extraction intact. Dense prose gets summarized, and summaries lose the citation.
• Self-contained atoms. A passage that makes sense without the surrounding page ("X is a Y that does Z") can be cited alone. Passages full of "as mentioned above" can't.

Stage 3: Attribution — being the source worth naming

When several candidates support the same claim, engines pick which to attribute. Patterns visible across platforms: original sources beat aggregators of the same fact; pages with visible freshness beat stale ones; and named, consistent entities beat ambiguous ones — if your brand is described differently everywhere, engines hedge by citing someone else's description of you.

What doesn't work

Keyword stuffing does nothing for synthesis. Blocking AI crawlers and expecting mentions anyway confuses the two layers — you can be described (training data) while being uncitable (no retrieval). And unverifiable superlatives ("the leading platform") are exactly what engines drop or launder into someone else's framing.

Verify, don't guess

Every claim above is testable against your own brand: run your buyer prompts on a schedule and watch which pages get cited, on which platforms, for which questions (prompt monitoring). Ship a restructured page and watch its citation rate. The engines will tell you what they like — if you're measuring.