How do you measure visibility in AI answers?

There is no equivalent of a search console for generated answers, and the tools that claim to supply one are producing their own measurements by a method they rarely describe. That does not make measurement impossible. It makes it manual, which is an acceptable cost given that the panel of questions that matters to a small business is usually small.

Start from questions rather than keywords. Write down the things customers actually ask before they buy — the operational questions, the comparisons, the ones about cost and timing and whether you cover their area. Twenty is plenty and ten is enough to start. These are the queries where being named as a source has commercial consequence, and they look very different from a keyword list.

Then ask them, on a schedule, and record what comes back: which sources are cited, whether you are among them, and if not, who is. The recording is the discipline. Asking once and forming an impression is worthless, because generated answers vary between runs, between users and between sessions for reasons that have nothing to do with your site. What is meaningful is a rate — in how many of thirty asks across a month did you appear — and a rate requires writing things down.

Vary who asks and from where, because personalisation and location shift results substantially. An answer produced in a logged-in session by someone who has visited your site repeatedly is not evidence about a stranger. Where the business serves a defined area, asking in a way that reflects how a local customer would phrase it matters more than any wording refinement, because the geography frequently determines which sources are considered at all.

The most useful output is usually not your own rate but the list of who else appears. Competitors that turn up consistently are worth examining for what they publish that you do not — and in practice the answer is often unglamorous: they published prices, or service areas, or a specific figure with a source on it. Directories and aggregators appearing repeatedly is its own finding, and it means the question is being answered from a listing rather than from any operator's own site, which is both a diagnosis and an opportunity.

Two cautions. Do not treat improvement over a short window as proof, because these systems change underneath you and a shift may have nothing to do with your work; hold the panel and the schedule constant so at least the measurement is stable. And do not let the panel drift toward questions you already rank for, which is the natural gravity of any self-assessment and quietly converts the measurement into a reassurance exercise.