Ranking #1 on Google no longer gets you cited in AI

For twenty years, the goal was simple: rank in the top 10. New data says that goal no longer buys you what it used to. Researchers looked at 863,000 search results and found that the pages Google cites in its AI Overviews increasingly are not the pages that rank at the top. The link and the citation have come apart.

If you run SEO, this is the finding to sit with. The comfortable assumption - "rank well and the AI will cite you" - is quietly becoming false, and there is now large-scale data to prove it.

What the study found

A 2026 study analysed 863,000 keyword SERPs and roughly 4 million AI Overview URLs. The headline: only about 38% of pages cited in AI Overviews also appear in the top 10 organic results for that query. Seven-to-eight months earlier, in mid-2025, that overlap was around 76%. In under a year it roughly halved.

Read that again, because it inverts a core belief. It used to be that if you were cited by the AI answer, you were almost certainly also ranking near the top. Now, the majority of AI citations come from pages that are not in the top 10 for the query you would have optimised for.

"The link and the citation have come apart - and being #1 no longer reliably puts you in the answer."

Curious how AI engines describe your brand right now? Get a free visibility audit and see where you stand across ChatGPT, Gemini and Perplexity.

Why it is happening: query fan-out

That research points to query fan-out as the main driver. When someone asks a question, the AI does not run that one query and cite the top results. It decomposes the question into many related sub-questions, runs them in parallel, and pulls sources to answer each piece. The final answer is stitched together from that wider pool.

The consequence is structural. A page that answers a specific sub-question precisely can get cited even if it would never rank in the top 10 for the broad head term. And a page that ranks #1 for the head term can be skipped if it does not cleanly answer the sub-questions the model actually asked. The ranking game and the citation game are now scored differently.

What this changes for your strategy

1. Stop using rank as a proxy for citation. They diverge now. A page-one ranking is no longer evidence that AI names you, and a strong AI citation can come from a page you were not even watching.
2. Cover the sub-questions, not just the keyword. Fan-out rewards content that answers the cluster of related questions around a topic clearly and self-containedly. Think in terms of the whole question space a buyer explores, not one head term.
3. Make each answer liftable. A clear, specific, standalone statement is easy for a model to retrieve and quote for a sub-question. Positioning wrapped in marketing prose is not.
4. Measure citation directly. Since ranking no longer predicts it, the only way to know whether you are in the answer is to check the answer. Track whether you are named, per engine, for your real buyer questions.

The takeaway

The collapsing overlap is not a reason to abandon SEO - ranking still helps discovery, and discovery still feeds AI. It is a reason to stop mistaking one number for the other. Being #1 was never the real goal; being in the answer is. The data now says those are separate achievements, and the brands that measure and optimise for the second one directly are the ones that will keep showing up as search keeps changing shape.

What this looks like in practice

Picture a company that sells project management software. It has spent years fighting for the top spot on "best project management tool", and it wins that ranking. Under the old rules, that would be the end of the story: rank first, get named. Under fan-out, the model asking on the user's behalf never asks that broad question in isolation. It breaks the request into pieces - "which tool handles Gantt charts", "which integrates with Slack", "which is cheapest for a five-person team", "which one do agencies prefer" - and answers each from whatever page states the fact most cleanly.

Now the ranking and the citation split apart in front of you. A comparison page that ranks nowhere for the head term, but flatly answers "the cheapest plan for small teams is X", gets pulled into the answer. The category leader, whose homepage wraps every claim in positioning language, gets skipped. The buyer reads a synthesised answer that names three tools, and the one that was #1 on Google is not among them. Nothing about the ranking changed. The route to the buyer did.

This is the uncomfortable part for anyone who has built a content operation around head terms. The work that wins rankings - broad, authoritative pages targeting the biggest keyword - is not the same work that wins citations. The citation goes to the page that answers the exact sub-question with the least friction, wherever that page happens to sit in the rankings.

How to measure whether you are actually cited

Since ranking no longer stands in for citation, you have to watch the answer itself. That means building a small, deliberate list of the questions your buyers actually ask - not your head keywords, but the real phrasing a person types or speaks - and checking, on a regular cadence, whether the AI answer names you.

• Start from buyer questions, not keywords. Write down the twenty to thirty things a prospect genuinely asks before they buy. These are the prompts worth tracking, because these are the prompts the models are fanning out from.
• Check per engine, not in aggregate. ChatGPT, Claude, Perplexity, Gemini and Google's AI surfaces draw from different source pools and weight them differently. Being named in one is not being named in all, so track each separately.
• Record presence, not just position. The question is binary before it is anything else: were you in the answer or not? Then, if you were, was your point represented accurately, and were you the primary source or a footnote.
• Re-run over time. Answers shift as models retrain and as sources move. A citation you had last month can quietly disappear, and you will only know if you keep looking.

Done consistently, this turns a vague worry - "are we showing up in AI?" - into a number you can move. It also tells you exactly which sub-questions you are losing, which is the same list that tells you what to write next.

What this does not mean

It is easy to over-read a finding like this, so it is worth being precise about the limits. The collapsing overlap does not mean rankings are worthless. Ranking still drives the human clicks that AI has not replaced, and pages that rank well are still part of the pool models draw from. A page nobody can find is not a page an AI is likely to cite either. Discovery and citation are related; they have simply stopped being the same measurement.

Nor does it mean you should chase citations by gaming them. Stuffing pages with question-shaped headings you cannot actually answer, or manufacturing thin pages for every conceivable sub-query, tends to backfire. Models corroborate claims across sources, and a page that states something no one else supports is a page they learn to distrust. The durable move is to be genuinely, verifiably right about the things your buyers ask, and to say so plainly.

"Discovery and citation are related. They have simply stopped being the same measurement."

And it does not mean the number is fixed. The overlap has moved fast in one direction, and there is no reason to assume it settles where it is today. That is the real argument for measuring citation directly rather than inferring it from rank: whatever the relationship becomes next, you will be watching the thing that actually decides whether a buyer hears your name.