AI visibility is how often a brand gets named or cited in the answers AI systems generate for its buyers’ questions. It is counted in citations and mentions inside those answers. It is not a ranking. It is not traffic. No single tool measures it completely.
That last sentence is the useful one. It is the part the tools selling you AI visibility tend to leave out.
Measurement is one part of a wider subject. Answer visibility sets out the rest of it: the four competing labels, how much of this overlaps with ordinary search results and what decides whether a model quotes a page at all.
What it is not
It is not a ranking. There is no position one in a generated paragraph. There is named, named late, or absent.
It is not traffic. Most AI answers are read and closed. A brand can be the single named recommendation in thousands of answers and see almost nothing in its analytics, which is exactly why the measurement problem exists.
It is not a score. Several vendors publish a proprietary “AI visibility score” out of 100. Each one is computed from a prompt set that vendor chose, weighted by a formula that vendor wrote. Two of them will not agree, because they are not measuring the same thing.
The five ways it gets measured
Every method in commercial use is one of these. Each sees something real. Each misses something big.
| Method | What it sees | What it misses |
|---|---|---|
| Run the prompts yourself | The actual answer, who is named, in what order, with the sources under it | Only the prompts you thought of, on the day you ran them |
| A third-party citation index (Ahrefs, Semrush and similar) | Citations observed across a large sample of real AI answers, split by platform | Whose questions those were, or whether they resemble your buyers’ |
| Search Console’s generative AI report | Impressions of your pages inside Google’s AI Overviews and AI Mode, from Google itself | Clicks, position, any surface that is not Google, plus the split between the two features |
| Referral traffic in GA4 | Sessions arriving from chatgpt.com, perplexity.ai, gemini.google.com and friends |
Every answer where you were named and nobody clicked, which is most of them |
| A vendor dashboard on your own prompt set | Movement over time, on a fixed set, across several platforms at once | Nothing, if you picked the prompts you already win |
Two of those deserve more detail, because they are the two that changed recently.
Search Console now reports it, partially
Google added a generative AI performance report to Search Console in June 2026. It covers AI Overviews and AI Mode.
Read the limits before you build anything on it. It reports impressions only: no clicks, no click-through rate, no average position. AI Overviews and AI Mode are combined in one view rather than split. It excludes Search Labs experiments. Not every property has it yet, which Google states plainly. Data is assigned to the canonical URL. The usual 1,000-row limit applies.
Outside that report, AI Overviews and AI Mode traffic sits inside the ordinary Web search type in the Performance report, mixed in with everything else. So a year-on-year drop in organic clicks tells you nothing about which surface took them.
A citation index gives you a hard number and no context
Ahrefs, Semrush and others now index the sources cited across a large sample of AI answers and report how many of them point at a given domain. That produces a count per platform, which is the closest thing this field has to an objective figure.
It is also the figure most easily misread. The sample is whatever questions those tools observed, not yours. A domain with a large content estate will collect citations across topics that have nothing to do with what it sells. And the platforms disagree with each other constantly, so a single total hides more than it shows.
Why two honest measurements disagree
A worked example, both figures public.
Tilio, an Exeter agency selling AEO and GEO, reports on its own site that it holds 27.1% visibility and 5.9% share of voice, ranking first among the agencies it tracks, as of July 2026. Ahrefs’ citation index, queried on 10 August 2026, records zero citations for tilio.co.uk across ChatGPT, Perplexity, Google AI Overviews, Gemini, Copilot and Grok.
Both readings can be accurate. They are measuring different things: one is share of voice on a chosen prompt set, the other is observed citations across a third party’s sample. Neither is lying. They simply have no common unit.
That is the state of the field. It is why a number on its own is worth very little. We put the same measurement across every agency the UK SERP currently names for GEO, which is the fastest way to see how wide the spread gets.
Our own numbers, since we are asking you to check everyone else’s
Ahrefs citation index, findable.digital, 10 August 2026:
| Platform | Citations | Cited pages |
|---|---|---|
| ChatGPT | 0 | 0 |
| Perplexity | 0 | 0 |
| Google AI Overviews | 0 | 0 |
| Gemini | 0 | 0 |
| Copilot | 0 | 0 |
| Grok | 0 | 0 |
Zero, on every platform. The site had no articles on it until this month.
We publish that table again every month, from this baseline, whichever way it moves. An agency that sells AI visibility and will not show its own is asking you to take the one thing it can prove on trust.
How to get a reading you can act on
Four steps, in order, none of which need a subscription.
- Write the fifteen questions your buyers actually ask. Category, comparison, alternatives, situation. Your company name appears in none of them.
- Run each one three times, in a fresh logged-out session, across ChatGPT, Perplexity, Google’s AI Overviews and Gemini. Three runs because the models are probabilistic and one answer proves nothing.
- Record four columns: were you named, where in the answer, who else was named, what was cited. Seventy-five rows is enough to see a pattern. Fewer is not.
- Read the citation column, not the answer column. The same small set of pages will do most of the work: a directory, one comparison post, a trade title, a competitor’s own page. That set is what you have to change.
The full version of that method, with the prompt shapes and the recording format, is in how to check whether AI recommends you. It takes an afternoon.
What a reading does not tell you
It does not tell you what the prompts are worth. Fifteen questions written at your desk are a hypothesis about your market. Some are worth thousands a month and some are worth nothing. A visibility score treats them identically.
It does not tell you why you are missing. Absent because nobody has written the answer, absent because your page is a brochure, absent because the cited sources do not list you: three different problems with three different costs. Generative engine optimisation is the label for the work; which of the three you have decides what that work actually is.
And it does not tell you what it is costing. That needs the commercial read underneath it: what those questions are worth, what you are already paying to answer them through paid search, where the money leaks after the click.
Measure first. The reading is cheap. Acting on the wrong one is not.