on this page
Most firms that invest in search have the same working assumption: if the site ranks, the AI will find it. Rank well on Google, and the assistants that increasingly sit between a client and a shortlist will pick you up as a matter of course.
That assumption was defensible two years ago. It is now measurably wrong, and the number that shows it is not subtle.
The number that settles it
In mid-2025, 76% of the citations appearing in Google's AI Overviews came from pages already ranking in Google's top 10 results. By early 2026, that figure had fallen to 38%, according to Ahrefs.
The systems halved their agreement in under a year.
This matters because of what the original 76% implied. If answer engines were simply reading the search index and summarising the winners, that overlap would have held steady or grown as the technology matured. Instead it collapsed. Whatever decides which sources an AI cites, it is no longer mostly the same thing that decides which pages rank.
For a firm that ranks well, this is the uncomfortable part: the ranking is still real, still valuable, and still earning traffic. It has simply stopped being a reliable proxy for the other thing.
Google's own products do not agree with each other
If the divergence were only between Google and its competitors, it could be explained away as a difference in engineering. It is not.
Only 13.7% of citations overlap between Google AI Overviews and Google AI Mode — two AI features built by the same company, drawing on the same index, released within months of each other. That figure is also from Ahrefs, measured in December 2025.
A firm cited in one is, on the evidence, unlikely to be cited in the other. There is no single Google position to win any more. There are several surfaces, and they behave differently enough that being visible on one says very little about the rest.
The same question, asked twice, returns different firms
There is a further problem, and it is the one that most often surprises firms when they first see it measured.
AI recommendations are not stable. Asked the same question repeatedly, ChatGPT or Google's AI has less than a 1 in 100 chance of returning the same list of brands in any two of those responses, according to SparkToro research from January 2026.
This has a direct consequence for how visibility can honestly be described. A firm is not simply "in" or "out" of an AI's answer. It has a probability of appearing, and that probability moves. A single check on a single day tells you almost nothing — it is one sample from a distribution.
This is why AI visibility has to be measured as a rate across a fixed bank of questions, asked the same way, repeatedly, over time. Anything less is anecdote.
Cited is not the same as named
One more distinction matters, because the two are routinely conflated.
Being cited means a link to your page appears among an answer's sources. Being named means the assistant says your firm's name in the answer itself. These happen at very different rates. Google AI Mode cites sources 76.3% of the time but mentions brand names only 37.6% of the time. AI Overviews sits closer together, at 61% mentions against 84.9% citations. Both figures come from Growth Memo research published in April 2026.
For a professional firm, the second number is usually the one that carries commercial weight. A client asking which firm to approach is looking for a name. A footnote they never open is worth considerably less than a sentence that says who you are.
What this changes for a firm that already ranks
None of this argues for abandoning search. Rankings still bring qualified visitors, and the technical foundation that earns them is the same foundation an answer engine needs in order to read a site at all.
What it argues against is treating the ranking as evidence. If a firm's position on AI platforms has never been measured, its current standing is unknown — regardless of how the site performs in search. The two used to move together closely enough that one could stand in for the other. At 38% overlap, and falling, it cannot.
The practical starting point is unglamorous: establish where the firm actually stands. Build a fixed bank of the questions clients genuinely ask, run them across the platforms that matter, and record how often the firm is cited, how often it is named, and who appears instead. That baseline is what turns AI visibility from an impression into a number that can be tracked.
Once there is a baseline, the work has a target. Without one, there is only the assumption — and the assumption is the thing the data has stopped supporting.
Sources
- Ahrefs, share of AI Overview citations drawn from pages ranking in Google's top 10, and citation overlap between AI Overviews and AI Mode (2026)
- SparkToro, consistency of brand recommendations across repeated AI queries (2026)
- Growth Memo, citation and brand-mention rates across the major AI engines (2026)