How to Rank in AI Overviews: What Google Has Actually Published

To rank in AI Overviews you have to be retrievable first. Google builds its AI features on the same index and the same core ranking systems as the blue links, retrieves pages with those systems, and then generates a response from what it retrieved. A page that cannot be retrieved cannot be cited, and there is no separate entrance. Everything else on the subject, the file formats, the chunking advice, the special markup, is either a detail or a distraction, and Google has now said in writing which is which.

I came to this from the recovery side rather than the AI side. When a client loses rankings in a spam update, they lose their AI citations in the same week, for the same reason, and they get them back in the same order. That is not a coincidence, it is the architecture. Which means the useful question is not how to optimise for AI Overviews, it is what is currently stopping your pages from being retrieved.

How AI Overviews choose what to cite

Google’s guide to optimizing for generative AI features describes two mechanisms by name, and both have practical consequences.

Grounding. Google uses retrieval-augmented generation to, in its own words, improve the quality, accuracy, and freshness of AI responses by relying on its core Search ranking systems to retrieve relevant, up-to-date web pages. The model does not answer from memory. It answers from documents the ranking system handed it, which is why classic ranking work is the input.

Query fan-out. Alongside the query the user typed, the model issues concurrent related queries to request more information and fetch additional relevant search results. This is why you can appear for a question you do not rank for directly: you were retrieved for one of the sub-queries. It is also why the tactic of building one page per predicted sub-query fails, which I come back to below.

Eligibility. Google states the requirement plainly: a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. Two things follow. A nosnippet or aggressive max-snippet directive removes you from AI features. So does anything that keeps the page out of the index, including the section-level demotions I deal with in recovery work.

What Google has published against what the market is selling

Most of what is currently sold as GEO or AEO consists of tactics Google has explicitly addressed. I keep this table because it settles about eighty percent of client arguments in one screen. The right column is Google’s published position, not my opinion of it.

Tactic	Google’s published position	What to do instead
Add an llms.txt file	Google Search itself does not use them; they will neither harm nor help	Keep it if another system consumes it. Do not bill for it as Google visibility work.
Chunk content into small blocks for AI parsing	No requirement to break content into tiny pieces	Structure for readers. Multiple topics on one page are understood.
Rewrite pages in an AI-friendly style	No need to write in a specific way just for generative AI search	Write the clearest version for a human. Synonyms and intent are handled.
Add AI-specific structured data	Structured data is not required for generative AI search, and no AI-specific type exists	Implement Article and Organization because they do useful work elsewhere.
Chase brand mentions across the web	Pursuing inauthentic mentions is not as helpful as it might seem	Earn coverage that would be worth having if AI search did not exist.
Publish a page per fan-out query variant	Violates the spam policies and is ineffective long term	One page per intent. This is scaled content abuse by definition.
Sell GEO or AEO as a separate discipline	Optimizing for generative AI search is optimizing for the search experience, and thus still SEO	Price it as SEO. The deliverables are the same ones.

I make my living partly from the GEO label, so I have an obvious incentive to tell you otherwise. I am not going to, because the alternative is selling a separate workstream that produces the same artefacts, and clients work that out eventually.

The last row is the dangerous one

Query fan-out created a specific and predictable mistake. Once practitioners learned that the model issues sub-queries, the obvious move was to publish a page answering each predicted sub-query. Agencies have shipped hundreds of these per site.

That is scaled content abuse under Google’s own definition: many pages generated for the primary purpose of manipulating search rankings rather than helping users. The policy is explicit that generative AI tools used to produce many pages without adding value fall inside it. So the tactic that is supposed to win AI citations is the one most likely to cost you the index eligibility that AI citations require. I expect this to be the dominant recovery job of the next eighteen months, in the way that borrowed-authority placements, which I covered in parasite SEO and site reputation abuse, were the dominant one of the last two years.

What actually moves AI citations

Google’s own list is short, and unglamorous, and I would write it in this order of leverage.

Fix retrievability before anything else. Indexed, snippet-eligible, no stray nosnippet, no crawl blocking, no demoted section. If a page fails here, nothing further on this list applies to it.
Publish something a model could not generate without you. Google names this directly: create valuable, non-commodity content and bring a unique point of view. Original data, first-hand outcomes, tested results. A restatement of the consensus is exactly what the model already has.
Answer the question in the opening. Not because AI needs special formatting, but because grounding works from what the page actually says, and a page that buries its answer under six hundred words of preamble is a worse document to ground against.
Cover the intent properly on one URL. Fan-out rewards a page that genuinely addresses the neighbouring questions. It does not reward seven thin pages that each address one.
Keep it current. Freshness is named in Google’s description of grounding. Stale figures are a retrieval liability, not just a credibility one.
Measure in Search Console. The generative AI performance report is the only first-party data on this. Google warns explicitly against third-party tools claiming access to internal Google metrics, because none have it.

The measurement trap

One reporting problem catches almost everyone, and it is worth naming because it produces false alarms that lead to unnecessary site changes.

AI Mode has counted toward Search Console totals since June 2025 and AI Overviews appear in the Performance report, so any comparison that straddles those boundaries is not clean. More importantly, a growing share of searches now end without a click. A page can hold its position, gain impressions and lose clicks, and that is a different diagnosis with a different remedy than a ranking loss. Report impressions and average position alongside clicks or you will misread it, and clients who misread it start making changes that destroy attribution.

This is the same discipline that separates a real penalty from a reporting artefact, which I set out in manual action penalty or algorithmic demotion.

Where AI visibility and recovery meet

If your pages were demoted by a spam update or a core update, AI visibility is not a separate project you can run in parallel. It is downstream of the recovery, because retrieval is the gate and the demotion is what closed it. In the recoveries I have run, and I documented one end to end in my spam update recovery case study, citations return alongside rankings rather than ahead of them.

The practical consequence for a site in that position is simple. Do the recovery work in my penalty recovery process first, and treat AI Overview presence as a result you will observe rather than a target you optimise toward. For operators in competitive commercial niches, the same order applies, and I set out what survived the last two years of updates in iGaming SEO after the spam updates.

Royking Niba is an SEO and GEO consultant specialising in penalty and spam-update recovery, with more than 8 million organic visits recovered for clients. Every Google position quoted here was checked against Google’s published Search Central documentation on 22 September 2026.