You do not get into an AI Overview by ranking first any more. In the most recent large study of the question, 38% of the pages cited in AI Overviews ranked in the top 10 for that same query, against 76% seven months before. The rest were quoted from page two, page ten, or from nowhere in the top 100. That single change is what makes this a different job from the one it looked like last year.
Where cited pages actually rank
Ahrefs published the numbers on 2 March 2026, from 863,000 keyword SERPs and 4 million AI Overview URLs collected in January 2026.
Where the cited page ranks for that query Share of citations
Top 10 37.9%
Positions 11 to 100 31.2%
Not in the top 100 31.0%
Filtered to plain organic listings only, with ads and snippet features excluded, the picture gets slightly worse for the received wisdom: 37.1% in the top 10 and 36.7% outside the top 100.
One caveat belongs here rather than in a footnote, and Ahrefs raise it themselves. They improved how their parser detects citations between the July 2025 run and this one. Some part of a fall from 76% to 38% is therefore better detection rather than changed behaviour, and anyone quoting the drop as a pure measurement of Google is overreaching. The direction survives the caveat. The precision does not.
Query fan out is the mechanism
The reason a page that ranks nowhere gets quoted is that Google is not answering your question. It is answering several questions it wrote itself.
Given one query, Google generates a set of narrower related searches, runs them concurrently against its index, and composes the overview from what comes back. Your query was the prompt. The searches that actually retrieved the sources were never shown to you. A page can therefore be cited on a query it has never ranked for, because it ranks for a sub question inside it.
Work an example. Somebody searches for the best project management tool for small teams. The fan out underneath is closer to a list: which tools have free tiers, what do teams under ten people complain about, how do the pricing models compare per seat, which integrate with the tools they already run. A comparison page optimised for the head phrase competes with everybody. A support page that answers the fourth question precisely, on a domain nobody would call an authority on project management, wins that retrieval outright.
This is why the old instinct to consolidate everything into one long definitive page now works against you in this specific channel. Fan out rewards a surface that answers many adjacent narrow questions well. It does not reward one page that answers the broad one beautifully.
If you want to see the fan out for your own topic before you write anything, the query fan out prompt reconstructs a plausible set and marks which of them you currently answer.
What Google says, which is less than people claim
It is worth reading the official position, because a large amount of advice sold in this category contradicts it.
Google’s documentation on AI features states that there are no additional requirements to appear in AI Overviews or AI Mode and no other special optimisations necessary. It goes further on the point people most want to be wrong: you do not need to create new machine readable files, AI text files, or markup, and there is no special schema.org structured data that you need to add.
Take that at face value and two things follow. There is no secret to buy. And the eligibility floor is ordinary: the page has to be indexed and allowed to show a snippet. A noindex, a nosnippet, an over aggressive max-snippet, a data-nosnippet wrapped around the part worth quoting, or a robots rule blocking the crawl, and no amount of content work matters. Those are the first things to check and they are quick: the indexability checker and the robots.txt checker between them cover the whole floor.
Believe Google on what is not required. Do not read it as a claim that nothing you do makes a difference, because the citation data plainly shows some pages getting quoted far more than their rankings explain.
The four things that actually move it
Be legible before you are clever
An assistant that cannot parse your page cannot quote it. If the content only exists after JavaScript runs, a retrieval pass may see an empty document, which is the failure we find most often and the one nobody notices because the page looks fine in a browser. Check it with the render gap checker, and run the whole page through the AI content readiness checker to see whether a clean passage can be lifted out of it at all.
Write the answer as a liftable unit
The shape that gets quoted is a self contained passage of forty to sixty words that answers one question without depending on the paragraph above it. That is not a style preference. A retrieval system is choosing a chunk, and a chunk that needs three sentences of context to make sense is a worse candidate than one that does not. Sub headings phrased as the actual question, answered immediately underneath, are the cheapest structural change available.
Cover the sub questions, not just the topic
Once you accept fan out, the content plan changes shape. Instead of one page targeting the head term, you want genuine coverage of the adjacent narrow questions, each one answerable in a paragraph, each with its own heading. A frequently asked questions block is the crudest version of this and it works. A set of properly separated pages is the better version, provided they are actually different pages and not the same page with the noun swapped.
Be somewhere other than your own site
The single most useful detail in the Ahrefs data is buried in the tail. Of the citations that came from outside the top 100, 18.2% were YouTube URLs, and YouTube accounted for 5.6% of every AI Overview citation in the sample. Fan out retrieves across source types, so a brand with a decent presence across a video, a forum thread, a documentation page and a review platform is quoted more than a brand with one excellent article and nothing else. Our own reading of the wider evidence, written up in where AI answers really come from, points the same way.
What the click is worth now, stated honestly
Three studies, three methods, one direction.
Pew Research Center installed a browsing tracker on more than 900 consenting adults and recorded 68,879 real Google queries in March 2025. Users clicked a traditional result on 8% of visits where an AI summary appeared, against 15% where none did. They clicked a link inside the summary itself on 1% of visits.
Seer Interactive analysed 3,119 informational queries across 42 organisations, covering 25.1 million organic impressions from June 2024 to September 2025, and published on 4 November 2025. Organic click through rate on queries with an AI Overview fell from 1.76% to 0.61%.
And the number that should temper the panic, from the same Seer dataset: queries with no AI Overview at all fell 41% year over year. Something larger than this one feature is moving, and attributing every lost click to AI Overviews is the kind of reasoning that produces a bad strategy.
Against that, the finding that justifies the work: Seer measured that being cited in the overview was associated with a 35% higher organic click through rate than not being cited on the same query. The pie is smaller. Being named is how you keep a slice of it.
How often does this even happen to you
Published prevalence figures for the first half of 2026 range from roughly a quarter of queries to roughly 60%, depending on whose keyword set, which country, which devices and how the tracker detects the feature. The spread is the finding. Averages assembled from somebody else’s keyword list tell you nothing actionable about yours.
Your own Search Console data does. Look for queries where impressions are steady and clicks have fallen, and you have found where this is happening to you specifically, without needing anybody’s industry benchmark. The Search Console keyword analysis prompt does that split on an export.
What not to bother with
An llms.txt file, for this purpose. It has its arguments and Google’s stated position is explicit that no AI text file is required. We wrote up the wider evidence in llms.txt, and whether it works.
Blocking Google-Extended and expecting to stay in AI Overviews. That directive governs training and other Google AI products, and getting the robots directives confused here is a way to remove yourself from a surface you were trying to win. The AI crawler checker shows what each agent is currently allowed.
Any product promising placement. There is no placement to sell. There is eligibility, which is technical and checkable, and there is being the best available source for a question you never see, which is editorial.
Publishing twenty near identical pages to cover twenty sub questions. Fan out rewards coverage, and a template with the noun swapped is not coverage. It is the doorway pattern, it is a known penalty, and it is the most likely way this advice gets misapplied.
The order to do this in
Confirm eligibility. Indexable, crawlable, no snippet suppression on the part worth quoting. Nothing else matters until this passes.
Fix the render gap if there is one, because it invalidates everything downstream.
Reconstruct the fan out for your five most valuable queries and list the sub questions you do not currently answer.
Restructure your existing strong pages so each section answers one question in a liftable passage. This is editing, not writing, and it is the highest return hour on the list.
Only then write new pages, one per genuinely distinct question.
Establish presence on at least one surface you do not own.
A free audit covers steps one and two in a couple of minutes, which is the part that is mechanical. Steps three onwards are the actual work and no tool does them for you.
Top comments (0)