Six in Ten AI Overview Citations Now Come From Outside Page One

Most writing on this topic asserts that rankings and AI citations have come apart, then stops, which leaves the reader with a slogan rather than a decision. This article gives the size of the gap, and then argues against the conclusion people usually draw from it. The data does not say rankings stopped mattering. It says the citation surface got much wider while the top of it stayed valuable. Those are different claims, and confusing them leads to abandoning work that still pays.

AI visibility is whether a generative system names your business when answering a question, and whether it represents you accurately when it does.

Two properties make it behave unlike a ranking. It is not positional: there is no list, just a handful of sources folded into synthesised text, so you are present or absent rather than eighth. And it is not stable in the way a rank is. Ask the same question twice and the source list can differ, because retrieval is probabilistic rather than a fixed lookup.

Accuracy belongs in the definition. Being cited with the wrong facts attached is a distinct failure from not being cited, and it is one a ranked list never produced.

The clearest evidence comes from Ahrefs, analysing 863,000 keyword SERPs against roughly 4 million AI Overview URLs, published March 2026.

In July 2025 the top-10 share of that same measurement was about 76%. It roughly halved in under a year.

The row that should change your thinking is the last one. Close to a third of citations go to pages ranking beyond position 100 for the query that produced them. Under a purely rank-driven model of retrieval, that should be rare. It is not rare, which means the candidate pool for citation is drawn far more widely than the first page of results.

The overcorrection is as costly as the original mistake.

The same body of work finds that position 1 still carries a high probability of citation, far above any other single position. Rank has not become irrelevant; it has become insufficient. Those sound similar and imply opposite budgets. "Irrelevant" says stop doing SEO. "Insufficient" says keep doing it and add something.

The mechanism makes this coherent. Retrieval draws a candidate pool and then selects from it. Ranking well is a strong way into the pool, which is why position 1 still performs. But the pool is much larger than the top 10, which is why two thirds of citations come from outside it. Both facts can hold at once, and they do.

Read plainly: rank is neither necessary nor sufficient for citation. Necessary and sufficient are the words doing the work. A page can rank first and be passed over. A page can rank nowhere and be quoted.

This is the part most articles skip, and it is the part that determines whether you can act on any of it.

On Google, you have a first-party read. The generative AI performance reports launched in June 2026 and finished rolling out worldwide on 31 August 2026, separating AI Overview and AI Mode impressions from ordinary search impressions. That is your own data about your own site.

Everywhere else, you do not. There is no impressions report for ChatGPT, Perplexity or Copilot. Every number you will see about visibility on those engines is produced by asking them questions and recording what they cite. That is a legitimate method and it is the only one available, but it is sampling, not census, and it is noisy in a specific way: repeat the same question and the sources can change.

The practical consequence is that AI visibility should be read as a distribution across many queries over time, not as a position you hold. Anyone presenting it as a precise, stable rank-like number is presenting sampling noise as precision.

Track citation separately from ranking, because the two have measurably come apart and one no longer proxies for the other. Keep doing the ranking work, because the top of the pool is still the most reliable way into it. Expect the relationship to keep moving, since it moved by nearly forty points in eight months and nothing suggests it has settled.

And hold all of it loosely. A July 2026 critical survey of 45 studies in this field concluded that no reviewed technique shows a stable, longitudinal, cross-platform causal effect on discoverability. The decoupling above is a robust measurement of what is happening. It is not yet a validated theory of what to do about it, and the gap between those two things is where most confident advice in this field goes wrong.

For the terminology around all of this, see SEO vs AEO vs GEO vs LLMO. For why stable rankings can coexist with falling traffic, see zero-click search.