AI Search Citations Broke Away From Google Rankings. Here Is What Replaced Them

Ahrefs ran the same study twice, and AI search citations moved a long way. In July 2025, 76.1% of Google AI Overview citations came from pages ranking in the top 10 for that query. By March 2026 it was 37.9%. The cause is mechanical, not mysterious.

What the two studies actually measured

Do AI search citations still follow Google rankings? Partly. Ahrefs measured 863,000 keyword SERPs and 4 million AI Overview URLs in March 2026 and found 37.9% of cited pages ranked in the top 10 for that query. In July 2025 the same study reported 76.1%. Roughly two in three citations now come from outside page one.

That is the cleanest before-and-after in the public record, and it is worth being precise about what it counts. The unit is a URL cited inside an AI Overview. The test is whether that same URL appears in the organic top 10 for the keyword that triggered the Overview.

Both conditions matter. A page can be cited, rank well for a hundred other terms, and still fail this test. The study is not asking whether ranking helps. It is asking whether the ranking on the visible SERP predicts the citation above it.

Same publisher, same method, double the sample

The comparison holds up better than most numbers in this category because one team ran both passes. Ahrefs doubled the sample between runs and states that it improved its parsing in the second one. That is a caveat, and it points toward the newer figure being the more accurate of the two rather than the older one being wrong.

The distribution moved as much as the headline. In the 2026 run, 31.2% of cited URLs ranked between positions 11 and 100, and 31.0% did not rank in the top 100 at all. The 2025 run put those bands at 9.5% and 14.4%.

Read the second band, not the first. A page ranking 40th being pulled into an answer is a different phenomenon from a page ranking nowhere being pulled in. The first is a relevance system reaching deeper into the same index. The second is a system consulting an index you were not looking at.

The number everyone is quoting is the weakest one available

The version of this finding circulating fastest is "overlap collapsed from 70% to under 20%". It comes from a 5W Research press release distributed on PR Newswire. The release credits third-party analysis, discloses no sample size, no measurement window and no method, and arrives attached to a recommendation that mid-market brands move 40% to 50% of their SEO budget.

I would not build a budget case on that number, and I would push back on any agency that does. Not because the direction is wrong, since two independent datasets support the direction, but because a figure you cannot inspect cannot be defended when your board asks where it came from.

Three studies, three definitions

The confusion is real and it is definitional. These are not three measurements of one quantity. They are three different questions that happen to produce percentages.

The last row is included because it is the number most people have seen. It is the only row that cannot be checked, which is the point of putting it in the same table as the ones that can.

Look at rows two and three together. BrightEdge's tracking shows overall rank overlap climbing steadily, from 32.3% in May 2024 to 54.5% in September 2025, while the top-10 slice sits at 16.7% and barely moves. Those two facts sound contradictory. They are not. They are the same fact seen from two distances.

Fan-out is the mechanism, and it is not mysterious

Google's AI Overview does not answer your query. It decomposes it. The system generates related sub-queries, retrieves against each of them, and assembles an answer from what comes back. Ahrefs attributes the 2026 shift directly to this behaviour under Gemini 3.

Once you accept that, the numbers stop conflicting. A cited page can rank first for a sub-query you have never entered into a rank tracker while ranking 40th for the query the user typed. It fails the top-10 test and passes the retrieval test.

You are ranked for a query you never tracked

This is the part that changes daily work rather than slide decks. Your rank tracker holds a list of keywords a human chose. The retrieval layer is generating its own list, per query, at answer time, and it is not showing you that list.

So a falling top-10 overlap is partly a measurement artefact of tracking the wrong query. That does not make it harmless. It means the keyword report on your wall is describing a smaller share of the surface than it used to, and it is not telling you which share.

One counter-intuitive detail from the 2025 Ahrefs run is worth keeping. Pages cited from outside the top 10 ranked for fewer keywords on average, 887 against 1,020, and appeared for slightly shorter queries. If fan-out simply rewarded breadth, that number should point the other way. The mechanism is real. Its shape is not fully understood, including by the people measuring it.

Where AI citations actually come from now

Three structural facts sit underneath the headline, and each one has a different operational consequence.

A third of citations rank nowhere at all

31.0% of cited URLs are outside Google's top 100 for the triggering query. Of those, 18.2% are YouTube URLs. YouTube is 5.6% of all AI Overview citations and grew 34% in six months, making it the most-cited domain in the dataset.

That is a format finding disguised as a ranking finding. A video transcript is a retrievable document with almost no competition from the pages you have been optimising. If your category has no serious video presence, that is a gap a competitor can fill in a quarter.

The engines barely agree with each other

Treating "AI search" as one channel is the second mistake. The August 2025 Ahrefs cross-assistant study ran 15,000 long-tail queries and found only 12% of AI-cited URLs ranked in Google's top 10 for the original prompt. Perplexity aligned with Google rankings at 28.6%. ChatGPT, Gemini and Copilot clustered near 8%.

Perplexity behaves the most like a search engine. ChatGPT behaves the least like one. Optimising once and expecting four engines to respond is the same category error as the one in the argument that the dashboard is dying as an interface. The assumption in both cases is that a single surface still mediates the whole relationship.

What correlates with getting cited

The honest answer is that nobody has published a causal account. What exists is correlation, and the strongest correlation set comes from Ahrefs measuring 75,000 brands with Domain Rating above 40.

Brand mentions beat backlinks 3 to 1

Branded web mentions correlate at 0.664. Backlinks correlate at 0.218. Branded anchor text sits at 0.527 and branded search volume at 0.392. Every off-site brand signal outranks the link count.

Ahrefs is careful about this, and I will be too. The publisher describes all of these correlations as moderate to very weak on the Spearman scale, and says outright that correlation is not causation. High-mention brands are also large brands, and large brands are cited for reasons a coefficient cannot separate.

Even so, the ordering is stable and it points somewhere specific. Being talked about across the web is a better predictor of being quoted than being linked to. That is closer to a distribution problem than a technical one. It is the same shape as the case for why the incumbent with worse AI often wins, and as the reason distribution keeps capturing the value in agent marketplaces.

The academic base is thinner than the vendor base. The founding GEO paper from Aggarwal and colleagues, published at ACM SIGKDD in 2024, showed content-side changes lifting visibility in generative engines by up to 40%, with effects varying by domain. That is a controlled result on a benchmark, not a field measurement of Google in 2026. It is the strongest peer-reviewed evidence available, and it is two model generations old.

A citation is not a click, and the gap is the whole business case

Everything above is about visibility inside an answer. The commercial question is what that visibility is worth, and here the evidence is unflattering.

Pew Research tracked 68,879 searches by 900 US adults in March 2025. On visits where an AI summary appeared, users clicked a traditional result 8% of the time. Without a summary, 15%. Links inside the summary itself were clicked on 1% of visits. Users also ended the browsing session more often after a summary page, 26% against 16%.

So the click rate roughly halves and the in-answer link is close to worthless as a traffic source. If you are funding a citation programme on projected sessions, that model breaks on contact with these numbers.

The defensible case for GEO is not traffic. It is presence in the answer at the moment the buyer forms a shortlist, and it has to be measured that way. This is the same measurement problem covered in the analysis of where measurable AI return has actually appeared. The value is real and the instrument that would prove it does not exist yet. That gap is also the core of the AI productivity paradox in the national accounts.

Where this argument is weakest

This section exists because every post in this category skips it, including the ones selling the conclusion.

The instruments belong to the people selling the cure

Ahrefs, BrightEdge and Semrush all sell AI visibility tracking. Each of their studies concludes that AI visibility needs dedicated tracking. That is not an accusation of dishonesty, and Ahrefs in particular publishes its method and its caveats more openly than most peer-reviewed work in adjacent fields.

It is still a structural conflict, and it has a specific effect. Nobody in this dataset has a commercial reason to publish a null result. The finding "your existing rank tracker is fine" would be true or false on the same evidence, and it would never be distributed.

Semrush's 2026 index analysed 126 million prompts across four systems and does not publish its sampling or weighting. Scale is not method. A very large sample drawn in an undisclosed way is still an undisclosed way.

The counter-case: rankings still explain the majority

Here is the strongest version of the opposite argument. In the 2026 Ahrefs data, 69% of AI Overview citations rank somewhere in Google's top 100. BrightEdge puts the equivalent figure at roughly 53% and rising year on year. Being in the index and ranking respectably is still the entry ticket for two thirds of citations.

On that reading, the honest instruction is narrower than the headlines suggest. Traditional search competence has become necessary and insufficient rather than obsolete. Reallocating half an SEO budget on the strength of this evidence is an overreaction to a real change.

The part nobody can settle

Every figure here is a snapshot of a retrieval system that its own operator changes without notice. The 76% to 38% shift is attributed to a model upgrade. The next upgrade could move it back, and no external party would know why. Anyone selling you a durable playbook against a moving retrieval layer is selling confidence, not method.

What to change on Monday, and what to leave alone

The practical conclusion is narrower than the headline and more useful.

The last row is the one most GEO pitches leave out. It is also the row with the most evidence behind it.

Two additions are worth funding now. The first is video, because YouTube is the most-cited domain in AI Overviews and most B2B categories have almost nothing there. The second is measured brand presence, which behaves less like a marketing metric and more like an asset, in the way described in the distinction between data moats and workflow moats.

What I would not do is rebuild the content operation. The evidence supports adding an instrument and a format. It does not support the budget reallocations being sold on the back of an unmethodologised press release.

Frequently asked questions

Where to start this week

Two things, and the first one costs nothing but an afternoon.

Write down 30 prompts a real buyer would type into ChatGPT or Google in the fortnight before they shortlist you. Not keywords, prompts. Run all 30 across ChatGPT, Gemini and Perplexity, and record whether you appear and who does instead. That is your baseline, and almost nobody has one. Do it before you buy a tool, so you can tell whether the tool agrees with what you observed.

Then check one thing about your last twelve months of coverage: how many mentions of your company across the web carry no link. If that number is invisible in your reporting, you are not tracking the signal that correlates three times more strongly than the one you are tracking. Fixing that is a spreadsheet, not a strategy.

Related reading