How do you optimise for AI Overviews, according to Google?
- AI Overviews pull pages from the regular Google index: if you don't rank, you won't be cited.
- llms.txt, chunking, AI rewrites and AI schema do nothing for Google.
- You can measure it in Search Console, with the report for generative AI.
For a year now, findability in AI answers has kept getting new names: GEO, AEO, AI SEO. Each name comes with a list of tricks, and each list with a tool that measures it for you. Google has now written down for itself how it works, in the guide Optimizing your website for generative AI features on Google Search. Anyone who wants to know how to go about optimising for AI Overviews gets a remarkably down-to-earth answer: do SEO. Google literally calls GEO and AEO other names for the same work.
We have written about GEO, llms.txt and AI SEO on this blog before. After reading the guide, we updated those articles wherever they contradicted Google. Below is what Google says, what you can drop and where the story is different for ChatGPT and Perplexity.
How AI Overviews and AI Mode find your page
Google describes two mechanisms. The first is grounding, also known as retrieval-augmented generation: Search pulls relevant pages from its own index, the language model reads them and the answer shows links to the pages it relies on. So the model doesn't invent sources; it chooses from what the regular search engine has already found.
The second is query fan-out. For a single question, the model sends several related searches into the index at once, to cover the user's broader question. A question about a kamado for six people can thus also trigger searches about size, charcoal use and price.
The consequence is simple. A page has to be indexed and eligible for a snippet to end up in an AI answer. If you block snippets with nosnippet, or your page isn't in the index, it doesn't exist for AI Overviews. And because the same ranking and quality systems decide which pages take part, there is no separate AI ranking to win.
What you can drop, according to Google
The most interesting part of the guide is the list of things that aren't needed. A good share of the past year's GEO advice falls under it.
llms.txt. Google Search doesn't use it. The file can stay and may be useful for other services, but it does nothing for AI Overviews and AI Mode.
Markdown versions and AI files. Separate machine-readable versions of your pages aren't needed for Search. Google reads ordinary HTML.
Chopping content into chunks. Google extracts the relevant passage itself from a page that covers several topics. According to the guide there is no ideal page length; write for your reader.
Rewriting for AI. Search understands synonyms and intent. You don't need a separate text for every long-tail variant. In fact, mass-producing pages for query variations or fan-out questions can fall under the policy against scaled content abuse.
Special schema markup. There is no schema.org type that gets you into AI Overviews. Structured data is still useful for rich results, so keep the markup you had.
Buying or arranging mentions. Inauthentic "mentions" to steer what AI says about your brand don't help. The same spam systems that stop manipulation in the regular results apply here.
What does count when optimising for AI Overviews
What Google does recommend will be familiar to anyone who takes SEO seriously, with one emphasis that carries more weight than before: originality.
Original, non-commodity content. Google asks for content from first-hand experience, with its own point of view, rather than summaries of what already exists. A language model can write that summary itself; your measurements, customer questions and case studies it can't.
Writing for people. Clear structure, paragraphs and headings. Writing answer-first is still good, not because it is an AI trick, but because it helps a reader quickly.
Images and video. Relevant, high-quality images and video support the text.
A sound technical setup. Crawlable, including content loaded via JavaScript, as long as it isn't blocked. Semantic HTML where practical, without demanding perfection. Good page experience, little duplicate content, and on large sites attention to crawl budget.
Feeds and profiles. For online shops and local businesses, Google points to Merchant Center feeds and the Google Business Profile. Product data that is correct there can also be used by Google in AI answers.
Measuring: Search Console, and scepticism about tools
For measurement, Google points to Search Console, which has a performance report for the generative AI features in Search and Discover. It shows how your content appears in those features. So verify your site in Search Console if you haven't already.
The guide is also notably direct about third-party tools: be sceptical of tools that promise rankings or claim insight into internal Google metrics. Nobody has access to those systems. Check a tool's advice against the official documentation. That applies to us too: an AI findability score on a page says something about how readable and citable it is, not about what Google decides internally.
And ChatGPT, Claude and Perplexity?
The guide is only about Google. Other assistants have their own crawlers and their own ways of choosing sources. There, an llms.txt can help, and it matters that GPTBot, ClaudeBot and PerplexityBot are allowed to read your site. For those systems too, many crawlers don't run JavaScript, so server-side readable text matters more there than with Google.
The practical advice therefore stays simple. For Google, build on plain SEO and original content. Allow AI bots from other companies, and keep an llms.txt if it stays accurate automatically. Don't spend hours on tricks that only exist to game an AI system; Google now says in black and white that it doesn't look at them.
What we did with it, and what you do tomorrow
We put our own articles next to the guide. The twelve SEO tips said that ChatGPT and AI Overviews read your site differently from Google; for AI Overviews that isn't true, so we changed it. The article on AI SEO listed llms.txt as technical help for AI Overviews; it is now explicitly something for the assistants outside Google. Nothing changed in the platform: a Neuramerce site generates llms.txt automatically, and the rest of the basics (titles, metas, internal links, server-side HTML) was already the work Google now confirms.
For your site: check in Search Console whether your most important pages are indexed and allowed to show snippets. Run the free site scan to check the technical basics. And strike everything from your to-do list that only exists to please AI. That time is better spent on one page with something only you know.