Writing for AI search: how to be cited by ChatGPT, Perplexity, and Google's AI Overviews

The shift you cannot ignore

Until recently, the SEO question was: does my page rank for this query? Today an equally important question is: does my page get cited when an AI engine answers this query? Perplexity, ChatGPT search, Claude's web mode, and Google's AI Overviews all do the same basic thing — retrieve a small set of relevant pages and synthesise an answer from them, with footnoted citations. The user reads the answer; they may or may not click the citation. For information-seeking queries, citation rates are now the more important metric for many publishers, and the writing practices that earn citations are not identical to the ones that earn rank.

How AI engines actually choose what to cite

The pipeline is roughly: query rewriting, retrieval over an index of crawled pages, re-ranking by relevance to the rewritten query, extraction of candidate passages, synthesis into an answer, and citation back to the source passages used. Each step has implications for how you write. Retrieval favours pages with semantic relevance to a wide set of phrasings, not just exact matches. Re-ranking favours pages whose key claims are concentrated rather than scattered. Extraction favours self-contained passages that can stand alone outside their surrounding article. Synthesis favours sources that take clear positions rather than hedge every claim. Citation favours sources that are easy to attribute — named authors, clear publishers, dated content. The writers who get cited disproportionately are the ones who happen to write in a way that aligns with this pipeline. The good news is that the alignment is teachable.

Practice one: lead with the answer

The biggest predictor of citation is whether the first paragraph after each major heading directly answers the question the heading implies. AI extractors look for self-contained passages; an extractor that opens with 'There are several factors to consider...' is less useful than one that opens with 'The three factors that determine X are A, B, and C.' Write the answer first, develop it second. This is also better human writing — the inverted pyramid was a good idea long before AI search — but the discipline now compounds.

Practice two: own concrete claims

AI engines pass over hedged prose because synthesis cannot extract a position from a paragraph that takes none. Compare 'Some studies suggest that X may have an effect' with 'X reduces Y by roughly 30% across the studies we reviewed.' The second can be pulled directly into an answer; the first cannot. Hedging where genuine uncertainty exists is responsible — softening every claim by reflex is what to fix. The goal is not removal of nuance; it is willingness to commit to specific claims when you have grounds for them.

Practice three: structure for extraction

Lists, comparisons, and definitions extract well. Long discursive paragraphs extract poorly. Where the underlying material is genuinely list-shaped (criteria, steps, options, examples), use a list. Where it is genuinely comparative, use a table or a clear paired structure. Where it is genuinely a definition, lead with a one-sentence definition followed by elaboration. Do not contort prose into lists where prose is the right form, but recognise that a lot of writing reaches for paragraphs out of habit when the underlying shape is structured.

Practice four: write headings as questions or claims

Vague nominal headings ('Considerations,' 'The Landscape,' 'Background') are useless to retrieval. Headings that name what the section delivers — either as a question the section answers or a claim the section defends — are far more retrievable. 'Why does AI search favour list-shaped content?' beats 'Content Structure.' 'AI search engines do not pass through hedged prose' beats 'On Hedging.' This is not clickbait; it is honest signage.

Practice five: cite your sources visibly

AI engines weigh source quality. Pages that themselves cite primary sources, link to original research, and quote experts by name appear to be treated as more authoritative — both because the citations supply context for re-ranking and because they signal the kind of editorial care quality raters reward. The practical implication is that the inline citations you include in your writing become part of how downstream AI engines treat your page.

Practice six: keep dates and authorship explicit

Citation requires attribution. Pages with a clear publication date, a clear last-updated date, and a named author are easier to cite — and AI engines preferentially surface content that is recent and attributable. The cost of adding a 'Published' and 'Updated' line plus a real author bio is trivial; the citation lift is not.

Practice seven: cover the question, not just the keyword

Keyword-targeting tactics that work for traditional SEO often miss the mark for AI search because the engine's retrieval query is not the user's keyword — it is a rewritten, expanded version of the underlying intent. A page that comprehensively covers the question, including the natural follow-ups a reader would ask, is favoured over a page that hits the primary keyword and stops. The shift is from keyword density to question coverage.

Practice eight: write for the citation, not the click

A meaningful share of AI-search 'traffic' will read the answer and never click through. This is uncomfortable but real. The right response is to write content that does its job in the citation — that conveys the brand, the authority, the willingness to take a position, even when read as a quoted passage. The pages that win in this environment are pages whose excerpts read like the publication they came from, not like generic content that happens to be on a branded domain.

Case study: a guide that was rewritten for AI search

We took a 2,800-word guide on a SaaS topic that was ranking on page two of Google for its primary keyword and rewrote it to the practices above. The structural changes: every H2 became a question or claim; the first paragraph after each H2 led with a direct answer; six paragraphs of hedged commentary were tightened to specific claims; three opinions that had been buried in qualifying language were stated clearly; the author bio and updated date were made prominent; and a 'Related' section linked to three deeper pieces. Organic Google traffic moved up only modestly. Citation appearances in Perplexity and ChatGPT, tracked over the following ninety days, increased substantially — the page became the canonical citation for several of the underlying questions even though it never reached page one in classical SEO. The lesson is not that classical SEO no longer matters; it does. The lesson is that the writing changes that earn AI citations are different from the writing changes that earn rank, and they compound — well-extracted, well-attributed content tends to earn both over time.

What does not work

Stuffing pages with FAQ schema on questions the page does not actually answer. Faking author bios. Adding 'updated' dates without updating. Writing definitions in robotic encyclopedia style to seem extractable. AI engines, like search engines, are tuned against these patterns and they do not last. The durable advantage is to be the page that is most worth citing — not the page that mimics citation-friendly form without the substance.

How to start

Pick your three highest-value pages. For each, ask: does the first paragraph after each H2 directly answer the question the heading implies? Are claims specific or hedged? Is the author named and dated? Is the structure extractable? Make one editing pass focused only on these properties. Re-publish. Then watch citation appearances, not just rank, over the following sixty to ninety days. The signal you are looking for is not a spike — it is a slow climb in the share of AI-engine answers in your category that name your page.

Frequently asked