When five AI engines answer the same question, they name the same brand 30.3% of the time and cite the same exact page 6.8% of the time. A gap of 4.5 times. Since July, that number has underpinned a recommendation you now read everywhere: brand travels better than the page, so invest in the entity and stop chasing citations.
In August, the team that produced the number ran it against a random baseline. Measured against what chance would produce, the order inverts.
A permutation test compares one engine's source list for a given question against another engine's list for an unrelated question, holding list lengths fixed. It measures the overlap you would see with no shared understanding of the topic at all. Three independent shuffles, agreeing to the third decimal.
| Level | Real agreement | If picking at random | Multiple |
|---|---|---|---|
| Exact page | 6.8% | 0.16% | 42 × |
| Website | 10.9% | 1.0% | 11 × |
| Brand | 30.3% | 1.8% | 17 × |
In raw percentages, brand dwarfs page. Measured against chance, page becomes the stronger signal. The authors put it plainly themselves: they would no longer say brand "travels better" as though it held some superior property. It simply has a smaller field of candidates. An AI answer names three brands and picks from thousands of possible pages. Co-occurring on a brand is easy because there are so few places to take.
Why the correction matters more than the number
Any overlap rate rises mechanically as the pool of candidates shrinks. That is arithmetic, not behaviour. A raw percentage in this context therefore measures the size of the field as much as the way the engine chooses.
The same correction, applied to a second finding in the study, destroyed it. In raw terms, Google's two surfaces — AI Overviews and AI Mode — agree on 24.5% of websites, against an average of 7.3% for two engines from different companies. An apparent premium of 3.4 times, which people read as "engines from the same builder choose alike." Measured against chance, the premium falls to 1.06 times. Worse: Gemini paired with a Search surface scores higher than the two Search surfaces score with each other. What brings AI Overviews and AI Mode together is not how they choose. It is the pool of sources they draw from.
A raw number does not say what you think it says
- Brand converges more than page — but because there are fewer brands, not because it carries better.
- What separates two engines sits upstream of your page, in the pool of sources they consider at all.
- A finding published without a baseline is an observation, not a result.
What does hold, holds very well
The study covers 22.7 million citations, 1.15 million questions and 441,946 websites, between January and June 2026. Across the 531,889 questions where all five engines answered and cited at least one source, 79.6% of cited websites appear on one engine only. That result did not move under pressure.
Better: it strengthens when the test gets stricter. Match on the exact page rather than the website and it rises to 85.1%. Without cleaning up web addresses at all, 93.2%. And it stayed within a band of 77.7% to 81.0% in every one of the six months. This is not one odd batch of questions. It is how the system behaves.
ChatGPT stands furthest apart: no other engine touches 76.3% of what it cites. Perplexity follows at 69.9%, Gemini at 63.7%, Google's two surfaces around 50%. Here too the correction qualifies the reading — measured against chance, ChatGPT's pairings sit inside the common range. The gap describes the slice of the web it draws from, not a more unpredictable way of choosing. The practical conclusion is unchanged: you cannot infer presence in ChatGPT from presence in the other four.
The market's other studies carry the same exposure
BrightEdge, on 24 April 2026, reports 36% to 55% overlap across the top hundred brands named, against 16% to 59% across the top hundred sources cited. The direction is the same, and the work is serious. But those figures are raw. No baseline has been published. They are exposed to exactly the critique that brought down the Wellows interpretation.
Kevin Indig, in The Consensus Gap in May 2026, reports 91% of citations appearing on one engine only — across three engines only, which mechanically leaves fewer chances to overlap. Ahrefs, separately, measures 13.7% citation overlap between AI Overviews and AI Mode, against a semantic similarity of 0.86 between their answers. Two surfaces saying much the same thing while leaning on different sources.
That last observation is the defensible version of the original idea: engines can converge on substance while diverging on visible evidence. It remains true. What does not hold is the strategic use made of it — "work the entity, it carries better."
A contradiction nobody settles
Two recent studies say the opposite of each other about who owns the evidence.
Dmitrij Zatuchin, in an arXiv preprint dated 24 June 2026, analyses 167,551 URL citations covering 128 brands, 12 markets and 13 languages. He finds 85.7% of citations pointing to third-party domains against 14.3% to brand-owned domains. Wikipedia is the most-cited domain in 11 languages out of 12, and 80% of citations come from about 18% of domains.
Yext, across 6.8 million citations drawn from 1.6 million responses, concludes that 86% of citations come from sources brands control or strongly influence — their own sites, listings, reviews. Brand websites alone account for 44% of citations.
Both figures are defensible because they do not count the same thing. Zatuchin sets owned against non-owned by domain ownership. Yext files listings and review platforms on the brand-influence side. A professional directory where you fill in your own entry is a third-party domain to one and an influenced surface to the other. Neither is wrong. But quoting one without the other means choosing your conclusion before counting.
The useful question is not what share of the evidence you own, but which evidence you can create and which you have to earn.
What we take from it for measurement
Three rules, applicable tomorrow morning to any AI visibility dashboard.
Two lines, never one. Page citation and brand mention are two distinct measures that move independently. Adding them into a single score destroys the information you would need to decide. If no engine cites your pages but your name appears in the answers, you have a page problem, not a brand problem — and the remedy is different.
One engine, one column. With 79.6% of sources exclusive to a single engine, a blended score averages five surfaces that barely overlap. It can rise while the engine your buyers actually use goes dark.
A baseline, or nothing. That is the rule this episode installs. A raw gap between two percentages only becomes a result once it is measured against what chance would produce. We now apply it to our own readings, and we will ask for it from any study we cite.
The limits of what you have just read
The Wellows study is vendor research on its own customer data, not a sample of the web. The questions are all in English, spread across 27 markets with 84% in the United States, and come from brands already watching their AI visibility. Each question was collected once per engine, so part of what is measured as disagreement between engines is in fact a single engine's run-to-run variation — the authors name this as the largest open question in their work. Five engines are covered, with no Claude, no Copilot, no Grok.
The correction itself was only applied to the pairwise agreement figures. The convergence trend — from 8.49% in January to 11.25% in June — and the by-category differences were not tested. Finally, the Zatuchin preprint comes from a researcher affiliated with an AI visibility software company, which he declares. The same reservation applies to him as to Wellows, BrightEdge and Yext. It applies to us too.
So what to take away is not a new number to repeat. It is a requirement: ask to see the baseline before building a strategy on a gap.
Sources
- Wellows — AI Citation Overlap Study, 29 July 2026, updated 6 August 2026. 22,749,707 citations, 1,146,483 questions, January–June 2026.
- BrightEdge Research — Why AI Engines Cite Different Sources but Recommend the Same Brands, 24 April 2026.
- Dmitrij Zatuchin — How Large Language Models Source Brand Reputation Across Languages and Markets, arXiv:2606.25787, 24 June 2026. Preprint.
- Ahrefs — AI Overviews vs AI Mode. Citation overlap and semantic similarity.
- Kevin Indig — The Consensus Gap, Growth Memo, May 2026.
- Yext, via eMarketer — analysis of 6.8 million citations.