Measuring your visibility in AI answers: share of voice, cited or merely retrieved

Visibility in Google has been measurable for over ten years, through positions, impressions and clicks that Google supplies itself. Visibility in generative engines has no equivalent: neither ChatGPT, nor Perplexity, nor AI Overviews hand a publisher the list of questions it was cited for. So you have to build the measurement yourself, and build it reproducibly, or you are only measuring noise. Here is the method we use, with its limits.

Why one question is not enough

The same question asked twice of the same engine can produce two different lists of sources. Asking once and checking whether you appear tells you nothing. You have to ask a set of questions that covers how users actually phrase their need, and repeat the measurement over time.

Google has described how its AI mode breaks a question into several sub-queries run in parallel, a technique it calls query fan-out in its AI Mode announcement. The measurement has to do the same thing in reverse: start from the central question, spread it into variants, and query each variant.

Building the question set

From a starting question, you generate six families of variants:

Thirty variants per central question is a reasonable order of magnitude. Below that, the measurement depends too much on phrasing; above it, collection costs rise without gaining precision. Write the first set of variants by hand, then use a language model for later sets, provided you read them: a variant that matches no real question distorts the measurement as much as a missing one.

Cited or merely retrieved: the distinction that counts

When a generative engine answers, it retrieves a set of candidate pages, then cites a few of them in its reply. A page can be retrieved without being cited: it served as raw material, and the user will never see it. Some engines expose both lists, others only the citations. An honest measurement separates the two and tracks the second above all, because that is what produces awareness and visits.

The distinction also sharpens the diagnosis. A site frequently retrieved but rarely cited has a quotability problem: its passages are not sharp, sourced or attributable enough to be picked up. A site never retrieved has an access or indexing problem. The remedies differ: the first belongs to writing, covered in "Answer first, qualify second"; the second belongs to robots.txt, sitemaps and logs.

Choosing the engines and the cadence

Three engines already cover much of European usage: ChatGPT with its built-in search, Perplexity, and Google's generated answers. Each is queried through its interface or an API, with the same variants. The cadence depends on the stakes: a monthly measurement is enough to follow an underlying trend; a weekly one helps during a production campaign, to see whether new content is retrieved and then cited. What does not change is the protocol: same questions, same context, same counting method. Changing the question set halfway through means starting over.

Collection costs little against what it prevents: thirty variants across three engines, once a week, comes to ninety queries a week, and they can be automated. The analysis needs a human eye, particularly to read how competitors are presented and in which words.

Share of voice

The summary metric is a share of voice: across all the answers obtained for a question set, the share of answers in which a given source is cited, set against its competitors. You compute it by engine, by question family and by period, and each of those axes reads differently:

• By engine: a site can be well cited by Perplexity and absent from ChatGPT, because the indexes and the robots differ. It happens often, and it is frequently a robots.txt matter, as we explain in "robots.txt and AI".
• By question family: being cited on comparisons but not on definitions shows you where to produce.
• Over time: only the time series means anything. A one-off measurement is a point, not a trend.

The limits, stated plainly

This measurement has limits, and they are better accepted than hidden:

1. Answers vary from one run to the next. You do not compare two measurements, you compare two averages across several runs.
2. Engines change without notice: a shift in index or model moves the whole series. Date every measurement and record the changes you know about.
3. Personalization and location affect answers. Measure from a neutral, stable context, and document it.
4. The link to traffic is indirect. Being cited does not guarantee a click. Share of voice measures a presence, not revenue; set it separately against the real traffic arriving from AI tools, where your analytics can tell it apart.

What the measurement changes in production

A share of voice tracked by question family says exactly what to write next: the questions where competitors are cited and you are not, the families where nobody is cited consistently, the engines where you are absent for a technical reason. That is our production loop: understand what the market asks, produce in quotable passages, measure, pick up what slips. Without measurement you write blind; with a one-off measurement you tell yourself a story. Only a series by family and by engine lets you steer.

One practical piece of advice: keep the raw answers, not only the scores. Reading back, six months later, how an engine phrased a reply and who it cited is often more instructive than a curve. It is also the only way to check, when a number moves, whether the site changed or the engine did.