How to measure AI share of voice for your brand

The metric is only as good as the prompt set behind it, so build the set from questions buyers ask before they know your name. Pull them from sales call transcripts, support tickets, comparison threads on Reddit in your category, G2 and Capterra category pages, and the question-shaped queries in your Search Console data. A workable set is 20 to 40 prompts; fewer and one odd answer swings the score, more and manual checking becomes a chore nobody keeps up with.

Balance the set across intent. Category prompts (best expense software for agencies), comparison prompts (Brand A vs Brand B for a five-person team), and problem prompts (how do I automate invoice reminders) should dominate. Keep branded prompts under ten percent of the set, because asking an assistant about your own brand measures description accuracy rather than recommendation strength. Then freeze the set. Every prompt you swap breaks the trend line, so treat edits the way a survey researcher treats changing the questionnaire mid-study.

Assistants rarely give the same answer twice. Sampling variance means the same prompt can name you on Monday and skip you on Wednesday, and a model update or a change in retrieval behaviour can shift an entire category overnight. A one-off check therefore tells you very little. A flattering snapshot can hide a decline that started weeks earlier, and an alarming one can send a team into a fire drill over what turns out to be sampling noise.

Accuracy problems compound the variance problem. Our July 2026 study found 27 of 30 brands materially misdescribed by an AI assistant answering from memory, and 13 of 30 still misdescribed when the assistant used live web search. A single check cannot tell you whether an error like that is persistent or a bad sample; only repetition can. Re-run the same frozen prompt set weekly under the same conditions. Four weeks of data gives you a direction, and eight gives you a trend you can defend in a quarterly review.

An exec report needs three numbers and nothing else on page one. The score trend plots your share of voice per week against the previous eight, so a reader can see at a glance whether things are improving. The competitor leaderboard ranks every brand mentioned across the prompt set by mention count from the same run, which settles who is winning the category. The citation gap lists the sources the assistant cited in answers where competitors appear and you do not; that list is what to fix first.

The citation gap is the operational one, because it converts a score into a task list. If the assistant keeps citing a G2 category page you are absent from, getting listed there is the work. If it leans on a recurring Reddit thread or an outdated Wikipedia description, those become the work instead. This is the mechanism Generative Engine Optimization runs on: assistants recommend what their sources support, so you change the sources.

Ignore nearly everything else. Wording changes in a single answer and week-to-week movements of a few points are sampling noise, and so is the exact order of brands in one response. Ignore vanity prompts no buyer types, however well you perform on them. Resist pasting one good answer into a slide as evidence; a screenshot of a single response is the AI-era equivalent of celebrating a number one ranking for your own brand name.

The free way is a spreadsheet. List your prompts in column A, run each one in a fresh assistant session every week, and log which brands each answer names and which sources it cites. For a 30-prompt set this costs roughly an hour a week and nothing in cash. Consistency is the weak point: fresh sessions and identical phrasing matter, and manual processes drift, usually in the week your report is due.

Tooling removes the drift. Discoverable's free AI visibility checker produces a snapshot with no login; entering your email opens the full report, and the free plan includes one fresh re-run per month, which suits a quarterly sanity check. Monthly exec reporting needs the weekly cadence, which is what the paid tiers cover: Growth, from $79 a month billed annually, adds weekly automatic re-checks and generates citation outreach drafts, with a 14-day trial and no card required. Checks run on Claude today, with other engines rolling out. Whichever route you take, the discipline is identical: a frozen prompt set and a fixed cadence, reported as a trend rather than a snapshot.