What Google actually says about AI search optimization
Google has published its own guidance on how websites appear in AI features: AI features and your website. It is short, it is specific, and much of the GEO advice sold today contradicts it. Before paying for an AI optimization playbook, read the primary source. This page walks through what it says.
What Google says you do not need
The guide is unusually direct about the tactics that do nothing:
- No llms.txt. Google Search ignores llms.txt files. Adding one neither helps nor harms your visibility in Google's AI features. Vendors selling llms.txt generation as a Google optimization are selling a file Google does not read.
- No special markup or AI files. There is no AI-specific markup, tag or file format that improves selection for AI features. Standard crawling and indexing are the whole interface.
- Structured data is not required. Structured data has its own uses in search, but Google states it is not required for content to appear in generative AI features.
- No chunking or AI-specific rewriting. You do not need to break pages into fragments or restructure prose so a model can read it. Google's systems process normal pages written for people.
- No page per query variation. Generating a page for every phrasing of a question is not an optimization. It falls under the scaled content abuse policy, covered below.
What Google recommends instead
The positive guidance is brief: publish unique, non-commodity content with first-hand perspective. Content that says what fifty other pages already say, in different words, is exactly what a generative system does not need another copy of. Content grounded in your own experience, data and perspective is what gets selected.
The guide also explains the mechanics. Google's generative responses use retrieval augmented generation grounded in core ranking: candidate pages are retrieved through the same ranking systems as classic search. For a single question, the system uses query fan-out, issuing multiple related sub-queries and gathering pages across all of them. Two practical consequences follow. First, if your page cannot rank, it cannot be retrieved or cited, so ordinary SEO fundamentals still carry the load. Second, fan-out means a page can be pulled in through a related sub-query it answers well, so covering a topic with genuine depth matters more than matching one exact phrasing.
The spam policy warning
One popular GEO tactic is explicitly prohibited. Google's spam policies include scaled content abuse: mass-producing pages targeting variations of the same query. Because fan-out already retrieves your page across related sub-queries, generating hundreds of near-duplicate pages for phrasing variants adds no coverage and puts the whole site at policy risk. If a vendor's pitch is volume across query permutations, the pitch is a spam policy violation with a dashboard.
Measurement you can actually trust
Search Console includes a generative AI performance report, showing how your content performs in Google's AI features, measured by Google's own systems. That is first-party data. Google also states that no third-party tool has access inside its systems, so any external tool's numbers about your AI visibility on Google are outside-in estimates, whatever the marketing implies. How to run defensible measurement anyway is covered in measuring AI search visibility honestly.
Why the primary source wins
The pattern across the guide is consistent: there is no separate machinery to optimize for, and the levers that exist are the unglamorous ones, useful content, first-hand perspective, clean indexing, no scaled spam. When advice you have been sold conflicts with the document above, the document wins. It costs nothing to read and takes ten minutes.