LLM SEO: how language models decide what to cite

When someone asks an assistant a question it cannot answer confidently from memory, it searches. Behind the scenes the model rewrites the user's question into several search queries (a process called query fan-out), runs them against a search index, and fetches the top results. Perplexity, ChatGPT search, and Google's AI Overviews all work this way.

The entry requirements follow directly:

• You must be findable by search engines. The candidate set comes from search indexes, so pages invisible to classic SEO are invisible to retrieval. LLM SEO does not skip the search step; it inherits it.
• You must be fetchable by AI crawlers. OAI-SearchBot, PerplexityBot, and their peers must be allowed in robots.txt and not blocked by bot-protection firewalls.
• You must cover the fan-out, not just the keyword. A question like "what should a small team use for X" fans out into comparisons, pricing, and how-to queries. Content that covers the neighboring questions gets retrieved for more of the fan-out.

Retrieved pages are not fed to the model whole. Pipelines split them into chunks (passages of a few hundred words, usually cut at headings and paragraph boundaries) and rank those passages by how well they answer the question. Your page is not one contestant; it is thirty, and each one competes alone.

This is the stage most content silently loses:

• Sections must stand alone. A passage that begins "As we saw above, this approach also fails" carries no meaning without its neighbors, and it will be judged without them. Restate the subject at the top of each section.
• Heading-plus-answer is the winning shape. An H2 phrased as the question, answered completely in the first sentence or two beneath it, produces exactly the self-contained chunk selection favors.
• Lists and tables survive chunking intact. Structured content keeps its meaning when extracted; a beautiful five-paragraph argument often does not.

With ranked passages in hand, the model synthesizes an answer and attributes claims to sources. Between two passages saying roughly the same thing, the observable patterns favor:

• Directness. The passage that states the answer plainly beats the one that circles it. Hedged, qualified prose is harder to quote.
• Specificity. Numbers, named entities, dates, and concrete examples give the model something attributable. "Migrations complete in 5-10 minutes" is quotable; "migrations are fast" is filler.
• Source-of-truth signals. For claims about a product, the product's own documentation. For data, the site that generated it. Original information gets cited because citing anything else would be second-hand.
• Consistency with the wider web. Models weigh agreement. If your site says one thing about you and the rest of the web says another, the answer follows the consensus, which is why brand mentions and reviews are part of LLM SEO.

Not every answer triggers retrieval. For stable, well-known topics the model answers from its training data: a compressed memory of the web as it stood months ago. You influence this path on a slower clock, by being consistently described across the pages models train on: your site, review platforms, comparison articles, documentation, and community threads.

The practical split: retrieval visibility responds to work within days and is won page by page; memory presence responds over quarters and is won by reputation. Do the retrieval work first, because those same cited pages become the training data of the next model generation. Today's citations are tomorrow's memory.

Everything above compresses into a working checklist:

• Keep classic SEO healthy; retrieval inherits your search visibility.
• Allow AI retrieval crawlers in robots.txt and verify your firewall is not silently blocking them.
• Serve full content in raw HTML; most AI fetchers do not execute JavaScript.
• Structure every page as question-shaped H2s answered in the first sentence, with sections that stand alone.
• Load passages with specifics: numbers, names, dates, and original data worth attributing.
• Keep your brand's facts identical everywhere machines read, and publish llms.txt plus schema markup so those facts arrive typed.
• Measure monthly: AI referral traffic, assistant spot-checks on your buyers' questions, and discovery surveys.

For the strategy layer around these mechanics (crawler taxonomy, citation tactics, measurement), the eight-chapter GEO guide is the parent of this page, and the ChatGPT citation guide applies it to the assistant your buyers use most.