Google Finally Told Us How AI Search Works. Here Is What Actually Matters

Google's official guide, "Optimizing your website for generative AI features on Google Search," says three things that matter. First, AI Overviews and AI Mode are retrieval-augmented generation over the ordinary Search index, expanded by query fan-out. Second, the eligibility bar for appearing in AI features is the same boring foundation as always: indexed, snippet-eligible, and crawlable by Googlebot. Third, the tricks the AEO industry has been selling, llms.txt files, content chunking, and special AI schema, do nothing. Everything else is detail. Here is the detail.

The mechanism: RAG plus query fan-out

When someone asks an AI-powered question on Google, the model does not answer from memory. It issues multiple concurrent related searches over the ranked Search index, a process Google calls query fan-out, and pulls a wider, more diverse set of pages into a single generated answer, citing the sources it leaned on. One buyer question becomes many hidden queries: comparisons, follow-ups, persona variants, regional phrasings.

Two practical consequences follow. Pages that would never have ranked position one for the head query can still be pulled in and cited through a fan-out branch, which is why long-tail and comparison content earns citations it never earned clicks for. And optimizing for a single keyword misses the mechanism entirely: the unit of competition is now the whole fan-out set around a buyer question, not one query.

There is also a measurement consequence, and it is the part we care most about. Because fan-out happens per question, the honest way to study it is to sample the same buyer question repeatedly across engines and watch which pages get cited across runs. Citation patterns are stable enough to act on when the same domains keep winning the same branch of a question, and noisy enough that a single run proves nothing. Google publishing the mechanism does not hand anyone the internal signals, and the guide is explicit that no third party has them, but it does confirm that the observable layer, the answers and their citations, is the right thing to instrument.

The eligibility bar is deliberately boring

Google's stated bar for generative AI features has three parts. Indexed: if a page is not in the Search index, it cannot be retrieved, so noindex directives and pages stranded from your sitemap are disqualified before the AI ever sees them. Snippet-eligible: AI features assemble answers from snippet-eligible content, which quietly promotes your snippet controls into AI visibility controls. A nosnippet robots directive removes you from consideration, max-snippet limits with small values restrict what can be used, and data-nosnippet attributes carve out the exact passages the AI cannot quote. Crawlable: Googlebot has to be able to fetch the page.

One correction worth internalizing, because getting it backwards actually hurts sites: blocking Google-Extended in robots.txt does not remove you from AI Overviews. Google-Extended governs use of your content for Gemini model training elsewhere. The crawler that gates AI features on Search is Googlebot itself, so a well-meaning "block the AI bots" sweep that disallows Googlebot removes you from Search and AI answers at once, while blocking only Google-Extended changes nothing about your AI Overviews eligibility.

What Google debunked

• llms.txt: Google does not use it. Publishing one for Google's benefit is a no-op.
• Content chunking and AI-specific rewriting: no effect. Google's advice is to write for people, and the retrieval layer handles the rest.
• Special AI schema: there is no secret markup that unlocks AI features. Standard structured data remains useful for what it always did, but no schema type buys AI inclusion.
• Manufactured brand mentions: chasing inauthentic buzz to impress the models is called out as a non-strategy.
• Internal-metrics tools: any vendor claiming to use internal Google ranking signals is claiming something Google states no third party has.

What Google's own reporting shows, and hides

In June 2026 Search Console shipped a Generative AI performance report, so you can now see aggregate impression and click behaviour from AI experiences. It is UI-only, with no API, and it does not answer the questions that decide strategy: it does not break out visibility per prompt, it does not show which competitors appear beside you in answers, and it says nothing about ChatGPT, Perplexity, or Claude, engines Google will never report on. Treat it as a complement: Google tells you the rules of the game and a slice of your traffic, and you still need the scoreboard, which pages get cited, in which answers, next to whom, across engines.

The practical checklist

1. Confirm your money pages are indexed: covered by your sitemap, returning 200, and free of stray noindex directives.
2. Audit snippet controls: find every nosnippet, restrictive max-snippet, and data-nosnippet on pages you want cited, and remove the ones you cannot justify.
3. Check robots.txt for Googlebot specifically, and separately decide your policy for AI crawlers like GPTBot, ClaudeBot, and PerplexityBot, which gate other engines.
4. Stop paying for llms.txt generators, chunking rewrites, or AI schema packages. Reallocate that budget to content that answers buyer questions directly.
5. Cover the fan-out set around each buyer question: comparisons, alternatives, pricing questions, persona and region variants, not just the head term.
6. Keep standard structured data clean, and if you sell products, treat your product feed and Merchant Center data as an AI visibility surface.
7. Open the GSC Generative AI report for the traffic slice Google gives you, and measure the answers themselves for everything it does not.

The frame that survives this guide: the inputs to AI search are SEO, and Google just published the rulebook. What the rulebook does not tell you is whether ChatGPT recommends you or your competitor tonight. The inputs are rules. The outputs are a scoreboard, and the scoreboard is measurable.