AI share of voice is your portion of all brand mentions across the answers to a tracked prompt set. It is read comparatively — a percentage means nothing without knowing the competitor set it is measured against.
Reliability depends on three variables: how many prompts are tracked, how many engines are swept, and how many times each prompt is repeated. Generative models return different answers to the same question, so a single scan is not a trend; three repeats is a sensible floor.
The calculation should separate answers produced by live web search from answers produced from model memory. A report that collapses both into one number hides the part you can actually improve.
How is AI share of voice calculated?
The formula is simple: divide the number of answers in which your brand appears by the total number of questions in your set. Appear in 12 answers across 40 questions and your share is 30%.
The difficulty is not the formula but the conditions. The same question can produce different answers at different times, so a single measurement means little; share of voice is the average of repeated measurements. The question set must also stay fixed — adding questions breaks comparability with history.
The number matters less than its movement
An absolute percentage tells you little on its own: whether 30% is good depends on how hard your question set is and how many competitors exist. Two comparisons are meaningful — against your own baseline, and against competitors measured on the same set.
Which is why the first step in any measurement programme is recording a baseline. Without one, "visibility improved" is a claim nobody can verify.
Frequently asked
How many questions does a set need?
At least 30-40 to keep volatility manageable. In smaller sets a single changed answer moves the percentage sharply and the trend becomes unreadable.
Who should choose the questions?
They should come from real user questions — Search Console queries, questions the sales team receives, support tickets. Questions containing your own product name belong in a separate group: those measure awareness, while category questions measure competition.
Should each engine be reported separately?
Yes. ChatGPT, Perplexity and Google AI Overviews rely on different indexes; you can hold a high share in one and zero in another. A single blended average hides which side needs the work.
