The metrics
| KPI | Formula | What it tells you |
|---|---|---|
| Mention rate | answers naming you ÷ answers | How often AI recommends you |
| Citation rate | answers citing your domain ÷ answers | How often AI uses your pages as evidence |
| Share of voice | your mentions ÷ all tracked brands' mentions | Your slice of the recommendations in your category |
| First-mention rank | mean position among named brands, over answers naming you | Whether you are the first pick or an also-ran |
| Cited position | mean position of your page in the cited sources | How prominent your evidence is |
| Sentiment mix | positive / neutral / negative framing per mention | How the answer describes you |
| AI Overview presence | searches showing an AI Overview ÷ searches | How much of your search demand AI answers at all |
Why every metric needs n
AI answers vary between runs, so every rate is an estimate. Report it with the number of answers behind it and a 95% interval; the Wilson score interval behaves well even at small counts. With 9 answers, a 33% mention rate has an interval of roughly 12% to 65%: enough to say you are not dominant, not enough to say you fell from last month.
When a change is real
Compare this period with the last using a two-proportion test, or Fisher's exact test when any count is below 5, and call a change up or down only at p < 0.05. Otherwise the honest report is "no clear change". Mark the dates when an engine changes its model; a model change can move every number at once.
One number for the board
A visibility score combines the engines. We use the demand-weighted mention rate per engine, weighted by each engine's audience (ChatGPT 45%, AI Overviews 25%, AI Mode 10%, Gemini 8%, Perplexity 8%, Claude 4% by default) on a 0–100 scale, with its interval. Brand-direct questions ("is X legit?") are reported separately and never count toward it. The exact formula is on the methodology page.
The diagnostic KPI most teams miss
For ChatGPT, which shows both what it read and what it cited, label each question: not found, found but not cited, cited but not named, or named. The label tells you which fix to make, which a mention rate alone cannot.
See all of these for your brand in the free AI visibility check, or track them weekly with the AI visibility tracker.
Methods: E. B. Wilson (1927), score interval; R. A. Fisher (1922), exact test. Sightings methodology, October 2026.