Measuring your visibility in AI assistant answers

1. Choosing the questions

Start from the questions customers actually ask, in their own words. A question invented to flatter the brand measures nothing: it measures the question.

Questions are split into three intentions that are never mixed. Discovery names no brand — it is the only one that tells you whether a company is found by someone who does not know it. Brand puts the company name in the question: it tells you how the assistant describes the company, not whether it is discovered. Review submits the site itself to the assistant.

Each question is frozen with an identifier and a language. A reworded question is a different question: comparing the two would compare two different measurements and present the result as a trend.

2. Recording the answers

Each question is asked on an identified interface — a named API, a named model — and the whole answer is kept.

With it: the date, the model queried, the sources cited, the search mode requested in the call and the state of the search actually performed. That last point matters as much as the answer: a search can be requested and not happen. Three states are distinguished — confirmed, absent, and unknown when the interface does not say.

An interrupted answer is marked as not measured. It enters no total: a failure is neither a presence nor an absence, and counting it in the denominator would lower a rate for a technical reason.

3. Computing the indicators

Three distinct findings, never added together: is the brand named in the answer, is it recommended, and does the answer cite a page from the site. A mention is not a recommendation, and a page citation is not a mention.

Each is written with its denominator: the number of measured answers in the same segment. A segment is the combination of an interface, a language, an intention and a search state. Two segments are not added together.

An entirely fictitious example, given for the shape and not as a result: in a segment of 12 measured answers, 4 name the brand, 1 recommends it, 2 cite a page from the site. We write “4 out of 12”, never “33% visibility”.

A share of voice is computed against a list of competitors we wrote ourselves: a player missing from that list weighs nothing in it. It is therefore not a market share.

4. Comparing two campaigns

Compare the same questions, on the same interface, in the same search mode, and show both denominators. If one of the parameters changed, the comparison is presented as not established rather than as a trend.

A total accumulated over several campaigns is not the denominator of the next one: it would describe a longer period than the one being judged.

A before/after comparison is not, on its own, enough to attribute a variation to a change made to the site. Platforms change their models and their sources without telling us, and the same question can return two different answers on the same day.

5. What remains out of reach

The cause. A frequently cited page is not the demonstrated reason for a recommendation, and no figure in this method is a probability of being cited.

Platforms decide on indexing, ranking and citations. No position and no citation can be guaranteed, by anyone.

An API answer is not a consumer-application answer: the application takes into account an account, a history and a personalisation the API does not have. The two are measured separately and are not compared.