Ask the engines the questions your buyers ask, many times, and count how often your company is named, as N of M samples against named competitors. Then rerun the identical set after a change. One answer proves nothing: in our scans, two identical runs agree on only about 36% of the brands they name.
That is the whole method, and most of what is sold as AI visibility tracking skips at least one step. A screenshot of a single answer, a dashboard with a share-of-voice percentage, a monthly score with no saved answers behind it: each looks like a measurement and none of them can be checked. This piece describes a measurement that can be checked, and what it costs.
one answer is an anecdote
AI models do not return the same answer twice. Ask the same question in the same engine back to back and the list of brands changes. The figure we publish on our scan pages is that two identical runs agree on about 36% of brands. A founder who screenshots one good answer, or one bad one, has learned almost nothing about where the company stands.
The fix is sampling. Our paid scan asks 40 questions across four engines, ChatGPT, Claude, Gemini and Perplexity, with three samples per question. The free scan asks five questions with ten samples each, on ChatGPT only. The numbers differ because the purpose differs, but the principle is identical: you read a count over many samples, never a single reply.
write the questions the way buyers type them
The most common mistake is asking the engine about your own brand. A prompt that names your company tests whether the model has heard of you, which is the easy case. Buyers do not start there. They describe a problem and ask for options, and the real question is whether your company appears in the answer.
So the question set is problem-shaped and written in buyer language, agreed with your team before any sampling starts. The test for each question is simple: would a sharp prospect plausibly type this sentence at the moment they begin looking? If a question only makes sense to someone who already knows your product, it does not belong in the set.
count appearances, not share of voice
This is where we disagree with a lot of the field. Share of voice borrows a metric from advertising, where the total audience is a known pie. AI answers have no fixed pie. The number of brands per answer varies, the samples are small, and a percentage implies a precision the data cannot support.
We report appearances instead: named in N of M samples. It is a count anyone can check, because every answer is saved as a receipt, the full engine text and not a summary. Our scan pages say it plainly: this is not share of voice. If a vendor hands you a percentage and cannot show you the answers underneath, you are being asked to take it on trust.
read the answers, not only the count
The count tells you where you stand. The saved answers tell you why. When a competitor is named and you are not, the full text shows what the engine said instead, how it described the category, and which alternatives it reached for first. That is the material the fix is built from.
It is also how the third-party side gets handled. Models lean on sources you do not own, so the sprint includes a prioritized list of the third-party sources they cite for your category, with the play for each one. A count with no answers behind it can only tell you that you have a problem, and never what to do about it.
name the competitors
A reading only means something relative to someone. The paid scan runs the same questions for your company and three competitors you choose, then reports them side by side, so the result reads as named in N of M samples for you and for each of them. Nothing in our fixed-price tests is scored against an average or an index, because an average of no one in particular is not something you can make a decision against.
set a baseline, change something, rerun the identical set
Measurement earns its keep when it is repeated. The AI Positioning Foundation runs a baseline in week 0, makes fixes on the sources you own, such as site copy, schema and entity consistency, and in week 8 reruns the identical prompts. Same instrument, same questions, so the before and after is a real comparison and not an estimate.
We do not promise the count will rise. The rerun measures the change, whatever it turns out to be, and that sentence is written into the offer because a guaranteed lift in AI mentions cannot be kept honestly when the engines are not ours. For the work of earning mentions in the first place, our post on answer-engine visibility covers what goes on the page.
what it costs and what it will not tell you
The Who AI Names scan is a fixed USD 2,500 and ships in seven days: an expanded question set run against your named competitors, with recommendations. The first look is free, five questions and ten samples each, with no account and no card. The three tests on the proof tests page are each a one-time, dated read with no ongoing monitoring, and the guarantee there is that if we find nothing actionable, you do not pay for it.
A scan will not tell you whether mentions turn into revenue, and it cannot tell you how a single answer will read tomorrow. Models change, so every reading is dated. Treat it as a measurement of position in a given week, and rerun the identical set when something material changes.