AI citations produced the most-quoted SEO statistic of the year in July 2026: only 38% of pages cited in Google AI Overviews also rank in the top 10 for the same query, down from 76% a year earlier. The conclusion drawn everywhere is that ranking no longer buys AI visibility.
The direction is probably right. The number is not as solid as its reuse implies, and the reason is buried in the source: the instrument changed between the two readings. That is worth understanding before you rebuild a strategy on it.
Where the AI citations number comes from
The figure is Ahrefs, from a study of 863,000 keyword SERPs and 4 million AI Overview URLs published in March 2026. It is frequently attributed to a different piece of research, so it is worth being precise.
It is often bundled with the Previsible 2026 State of AI Discovery Report, published 6 July 2026 via Search Engine Land, which analysed 6.77 million sessions across 166 GA4 properties over 19 months and produced the widely cited claim that ChatGPT accounts for 92.4% of trackable standalone-LLM referral traffic. Those are two separate studies, measuring two different things, by two different organisations. Worth noting on that second one: the Search Engine Land article is written by Previsible’s own co-founder and chief product officer, so it is vendor research rather than independent analysis.
That distinction matters more than a footnote. The Ahrefs figure is about Google AI Overview citations specifically. It says nothing about ChatGPT, Perplexity, or any other assistant. "AI citations have decoupled from ranking" is a much broader claim than the data supports; "AI Overview citations have" is the supportable one, and the difference is not pedantry when you are deciding where to spend effort.
The methodology caveat almost nobody reports
Here is the part that changes how much weight the number carries.
Ahrefs’ parsing methodology improved between the July 2025 study and this one. The newer analysis detects more of the citations appearing in AI Overviews than the older one could. So some portion of the apparent collapse from 76% to 38% reflects better measurement of citations that were previously being missed, rather than a change in how Google selects sources.
The sampling frame changed too, and that matters as much. The 2025 study analysed only the three most visible citations in each AI Overview. The 2026 study counts every citation. Prominent citations are precisely the ones most likely to be strong, top-ranking pages, so widening the frame pushes the percentage down mechanically, independent of any behaviour change at Google.
This is not a criticism of Ahrefs, who disclosed the parser change and hedged their own comparison. It is a criticism of everyone quoting the comparison as though both readings came off the same instrument. Search Engine Journal put it plainly: the two datasets are not directly comparable, and some of the drop could reflect better detection rather than a change in how Google selects sources. If your net gets finer between two trawls, you cannot attribute the whole difference in your catch to the fish.
There is a harder version of this point available. Measuring the same phenomenon, BrightEdge found only 16.7% of AI Overview citations came from top-10 results while overall organic overlap rose to 54.5%, which is the opposite directional story. Three vendors, three incompatible numbers. That is the strongest evidence that this metric is instrument-dependent, and it is a better argument than the one the headline makes.
What that means practically: treat 38% as a reasonably solid current measurement, and treat "down from 76%" as directional rather than quantitative. The gap almost certainly narrowed. By how much is not established.
Two things do make the current figure credible. Ahrefs ran the analysis twice, once counting all result types including ads, featured snippets, People Also Ask and video packs, and once counting only standard organic listings. The results were close: 37% in the organic-only run against 38% overall. A robustness check that survives contact with a different definition is worth more than a headline.
The finding that deserved the headline
The distribution is more interesting than the top-line number, and it is the part being left out.
The roughly 62% of citations that do not come from top-10 pages split almost evenly:
- 31.2% from positions 11 to 100
- 31.0% from beyond position 100
So this is not a story about citations drifting from position 5 to position 15. Nearly a third of AI Overview citations come from pages that effectively do not rank for the query at all. That is a different phenomenon from a loosening correlation, and it means AI Overview source selection is running on signals substantially independent of the ranking system sitting next to it on the same page.
One more detail sharpens it. Among citations that did not rank in Google’s top 100 for the keyword, 18.2% were YouTube URLs, and YouTube accounted for 5.6% of all AI Overview citations in the dataset. A meaningful slice of the non-ranking citation pool is not a text-SEO opportunity at all.
What this actually changes
If the correlation between ranking and citation is genuinely weaker, the optimisation question shifts from position to something harder to game.
Being quotable matters more than being first. A model assembling an answer needs a passage it can lift with confidence: a clear claim, stated plainly, near a heading that matches the question.
Entity clarity matters more than keyword coverage. Being unambiguously identifiable as the source on a topic is a different job from ranking for its keywords, and it is the substance behind what our generative engine optimization primer describes. Structured data is a partial lever here, and our guide to schema markup for AI Overviews covers what it does and does not buy.
Position 11 to 100 is no longer dead ground. Under the old model, a page outside the top 10 was invisible. On these numbers it has roughly the same chance of being cited as a top-10 page. That is an argument for depth and reinforcement across a topic rather than aggressive pruning of anything not ranking, which is the approach we have taken with our own coverage.
And it is one surface among several. These numbers describe Google AI Overviews. If most of your assistant-referred traffic is arriving from ChatGPT, as the separate Previsible data suggests is common, then AI Overview citation behaviour is not the thing to optimise against. Our comparisons of GEO and SEO and our notes on measuring GEO cover how thin the measurement tooling still is.
The wider caution
Three of the numbers in circulation this month share a weakness worth stating plainly.
AI Overviews are reported on 48% of queries, but that is BrightEdge’s figure for queries it tracks, and it sits at the high end of the field: Ahrefs measured around 16% for US keywords and seoClarity around 30% for US desktop. On click-through, the numbers being quoted together do not belong together. Ahrefs now models a 58% reduction for position one on AI Overview keywords, having revised its earlier 34.5% figure upward in February 2026. Seer Interactive separately measured absolute organic CTR falling 61% over time, but in the same dataset queries without AI Overviews fell 41%, so the share attributable to AI Overviews is closer to twenty points than sixty. Those are third-party tracking estimates measuring different quantities, not Google data and not a range. The 92.4% ChatGPT share comes from 166 sites, which is not the web, and referral attribution from AI surfaces is unreliable in ways that are hard to quantify.
It is also worth knowing whether the assistants can reach you at all before optimising for how they cite you, which our AI crawler access audit walks through. None of this means the studies are worthless. It means the honest posture is to treat them as instruments with known error bars rather than as measurements of a settled reality, and to be suspicious of any single number that has travelled far enough to appear in a listicle without its methodology attached. On the martech side the same discipline applies: Knak’s survey of 333 enterprise marketing decision-makers found 88% saying AI-generated content needs moderate or substantial editing. That is a self-report about perception rather than a measurement of output quality, which is a distinction worth keeping, but it is still a useful corrective to the idea that any of this is automatic.
Frequently Asked Questions
Do AI citations still come from top-ranking pages?
Less often than they did. Ahrefs found 38% of pages cited in Google AI Overviews also rank in the top 10 for the same query, across 863,000 keywords and 4 million AI Overview URLs. The remainder splits almost evenly between positions 11 to 100 at 31.2% and beyond position 100 at 31.0%.
Is the drop from 76% to 38% reliable?
The direction is, the magnitude is not. Ahrefs’ parsing methodology improved between the July 2025 study and the current one, so it now detects citations it previously missed. Some of the apparent drop reflects better measurement rather than changed behaviour. Treat 38% as a solid current reading and the comparison to 76% as directional.
Is the 38% figure from the Previsible study?
No, and the two are frequently conflated. The 38% figure is Ahrefs. The Previsible 2026 State of AI Discovery Report, published 6 July 2026, analysed 6.77 million sessions across 166 GA4 properties and produced the separate claim that ChatGPT drives 92.4% of trackable standalone-LLM referral traffic. Different organisations, different datasets, different questions.
Does this apply to ChatGPT and Perplexity too?
Not demonstrably. The Ahrefs data measures Google AI Overview citations only. Other assistants use different retrieval and selection, and no comparable public dataset exists for them. Be wary of coverage that generalises this finding to “AI citations” without qualification.
Should we stop caring about rankings?
No. Ranking still drives organic clicks, and ten pages capturing 38% of citations remains enormous over-representation against the number of pages that could rank for a query. What has changed is that ranking is no longer close to sufficient for AI visibility, so it is one input rather than the whole strategy.
Why would a page that does not rank get cited?
Because AI Overview source selection appears to run on signals partly independent of ranking: passage-level relevance, how quotable a specific statement is, entity associations, and content type. The YouTube finding illustrates it, with 18.2% of citations that did not rank in the top 100 being YouTube URLs.
What should we do differently?
Write passages a model can lift with confidence, with clear claims under headings that match real questions. Build unambiguous entity associations on your topics. Stop treating pages outside the top 10 as dead weight, since on this data they are cited at a broadly similar rate. And identify which AI surface actually sends you traffic before optimising for any of them.
How much of AI marketing output needs human review?
Knak reported that 88% of AI marketing output still requires human editing. That figure is a useful counterweight to the automation framing around this whole category, and it is consistent with what the citation data implies: these systems are becoming more influential in distribution while remaining unreliable enough to need supervision.
