16% shared sources, 55% shared brands: AI engines read different pages and name the same companies
Two AI engines barely consult the same pages, yet agree far more on the companies they recommend. What the published research shows, and what it changes for your visibility.
16 - 59%
the sources two engines have in common on the same question (BrightEdge)
In this article
Short answer: answer engines agree markedly more on the brands they name than on the sources they read. On the same question, two engines share between 16% and 59% of their sources, but between 36% and 55% of the brands they name. Put differently: the page that gets you cited changes from one engine to the next, the reputation that gets you recommended changes far less. That asymmetry is what should decide your strategy.
36 - 55%
the brands those same engines have in common - a far tighter spread
up to 40%
the visibility gain from working on content, in the founding GEO study
The engines disagree about where they read, and agree far more about the name they say. So you do not optimise a page, you build a consensus.
Traffic collapses, value moves
The first fact is brutal and well documented. On queries where Google shows an AI Overview - the written summary at the top of the page - Seer Interactive measures organic click-through falling from 1.76% to 0.61%, a 61% drop. Ahrefs, using a different method, measures a 58% fall on the top-ranking page. On those same queries, the share of searches producing no click to any website reaches 80 to 83%.
The second fact drew less attention, and it reverses the reading. What little traffic survives converts far better. Adobe Analytics reported in April 2026 that shoppers arriving from an AI converted 42% better than others on US retail sites. Microsoft Clarity, across more than 1,200 sites tracked over eight months, measures a 1.66% sign-up rate for visitors from an assistant, against 0.15% from classic search.
| Measure | Classic search | With an AI answer |
|---|---|---|
| Organic click-through, queries with an AI Overview (Seer, Sept 2025) | 1.76% | 0.61% |
| Paid click-through, same queries (Seer) | 19.7% | 6.34% |
| Sign-up after the visit (Microsoft Clarity, 1,200+ sites) | 0.15% | 1.66% |
The two facts do not contradict each other: they describe a funnel that narrows at the top and hardens at the bottom. Someone clicking through from an AI answer has already read a comparison, retained a name and decided to go and look. They are not arriving to research, they are arriving to verify. Counting sessions therefore measures what is disappearing; what remains worth measuring is the qualified arrival.
The engines do not read the same pages
Here sits the most useful result, and the least intuitive. BrightEdge compared, sector by sector, the top hundred sources cited and the top hundred brands named by five engines - ChatGPT, Perplexity, Gemini, Google AI Mode and AI Overviews - across ten sectors, measuring the overlap by Jaccard similarity.
| Engine pair | Sources in common |
|---|---|
| Google AI Mode and AI Overviews | 59% |
| Gemini and ChatGPT | 39% |
| Gemini and AI Overviews | 34% |
| Gemini and AI Mode | 27% |
| Floor measured across all pairs | 16% |
Sources therefore overlap by 16% to 59% depending on the pair: a 43-point spread, which is near-total instability. Brands overlap by 36% to 55%: a spread of only 19 points. The disagreement is about the library, not about the recommendation.
The practical consequence is clear. Chasing a citation on one particular site because it « shows up in ChatGPT » optimises the unstable half of the system: that same site counts for much less on Gemini. Being described the same way by enough different sources that every engine, working from its own library, lands on the same name - that is the half that holds.
The corpus is not your website
That leaves the question of where the engines read. An analysis of more than 150,000 assistant citations puts Reddit first, cited in 40.1% of cases, ahead of Wikipedia at 26.3% and YouTube at 23.5%. LinkedIn appears in 14.3% of ChatGPT Search answers. None of those four is a company website.
| Domain | Share of citations |
|---|---|
| reddit.com | 40.1% |
| wikipedia.org | 26.3% |
| youtube.com | 23.5% |
| linkedin.com (ChatGPT Search answers) | 14.3% |
This is the mechanical reason behind the asymmetry above. An engine leaning on forums, an encyclopedia and videos does not read the pages read by an engine wired into the trade press. But if your name recurs in those different places with the same description, all of them end up producing it. The consistency of what is said about you matters more than control of any one channel.
What the research says actually works
The founding study on GEO - generative engine optimisation, published by researchers at Princeton and Georgia Tech - tested content changes against a benchmark of queries and sources. Its headline result, as the paper states it: these methods boost visibility by up to 40% in generative engine responses. The levers tested are editorial rather than technical: citing sources, quoting people, giving figures, writing clearly.
That point deserves emphasis, because it cuts against the intuition inherited from SEO. There is no tag to add here, no domain score to win. What moves visibility is the shape of what is written: a text that asserts without evidence is quoted badly, a text that gives a figure and its source is quoted well. An engine repeats what it can attribute.
These figures come from tool vendors, not a public institute.
Seer, Ahrefs, BrightEdge, Adobe and Microsoft measure across their own client or site estates. They are the best public data available, they are not official statistics, and we have not reproduced them ourselves.
The methods are not comparable with each other.
Ahrefs measures click-through on the top-ranking page, Seer a series at brand level: reading « -58% » and « -61% » as two confirmations of one number would be a mistake. They are two different measurements pointing the same way.
Nobody agrees on the size of the phenomenon.
Datos and SparkToro measure all AI tools at 3.2% of US desktop searches in the fourth quarter of 2025, while other publications put the share far higher. The gap comes from what is being counted: usage, queries, or sessions.
One widely shared figure did not survive checking.
It is often written that engines share only 0.09% of their sources while agreeing 98% of the time on brands. We could not find it at source with any publisher; the figures BrightEdge publishes give 16% to 59% and 36% to 55%. We publish those, less spectacular and verifiable.
What this changes
Aim for consistency, not for the page.
A citation won on one particular channel counts with one engine and little with the others. What travels across engines is a name described the same way in several places.
Measure engine by engine.
A single score averaging ChatGPT, Gemini and Perplexity adds up disjoint libraries. Always ask which engine the number you are shown refers to.
Feed the corpus, not only the site.
The four most cited domains are a forum, an encyclopedia, a video platform and a professional network. Your site is only one of the sources the engine consults.
Count qualified arrivals, not sessions.
Traffic falls and converts better: a dashboard tracking volume alone will record a catastrophe where a shift is taking place.