The three numbers
"Are we showing up in AI answers?" is really three questions, and each needs its own number.
- Mention rate: answers that mention your brand, divided by all answers. How often you are in the conversation at all.
- Cited URL rate: answers that cite a page on your domain, divided by all answers. How often your own content is the source, rather than someone else writing about you.
- Share of voice: mentions of your brand, divided by mentions of all the brands you track, in the same answers. How much of the category's attention is yours.
Report them separately. A brand can be mentioned often and cited rarely, which usually means third parties are describing it. A brand can hold its mention rate while losing share of voice, which means a competitor is gaining. One blended score hides both. The glossary entry on share of voice has the short definition.
Why single answers mislead
AI answers are not stable. SparkToro ran 2,961 prompts across ChatGPT, Claude and Google's AI Overviews, falling back to AI Mode where no Overview appeared, with each prompt run 60 to 100 times per platform. The same list of brands came back less than once in a hundred runs, and the same list in the same order less than once in a thousand.
Rand Fishkin's conclusion was blunt: "any tool that gives a 'ranking position in AI' is full of baloney." What held up better was how often a brand appeared across many runs. That is the basis for everything below: measure frequency across a set, never position in an answer.
Step 1: fix the prompt set
Choose the questions your buyers ask an assistant on the way to a decision, not your brand name, which almost always returns your brand. Aim for a stable set, around 50, spread across the stages from problem to purchase. Our glossary entry on prompt discovery covers what makes a good prompt.
Then freeze it. If you change the prompts between periods, the trend measures your edits, not your visibility. Add new prompts as a separate cohort and report them separately until they have their own history.
Step 2: name the competitors
Pick five to eight brands a buyer would realistically shortlist against you. For each, list the names an answer might use: the company, the product, common abbreviations. Count a mention when any of them appears. Decide the list before you look at the answers, so you are not choosing the competitors that make the numbers look best.
Step 3: run it more than once
Because a single answer is a sample of one, run each prompt more than once in each period, at similar times, and record the engine, the date and the full answer with its sources. More runs cost more, but they are the only thing that turns a noisy measure into a usable one.
Step 4: count per engine
Engines do not draw on the same sources. In Profound's analysis, Wikipedia was 47.9% of ChatGPT's top-10 cited sources while Reddit was 46.7% of Perplexity's. A brand strong on one and absent from another will look average in a blended number and be neither. Calculate all three numbers for each engine, and only then summarise.
Step 5: put a margin on it
A mention rate is a proportion, so a standard margin applies. For a rate p measured on n answers, the 95% margin is about 1.96 × √(p × (1 − p) ÷ n).
Two consequences are worth remembering:
- The margin shrinks with the square root of the number of answers, so quadrupling the answers halves the margin.
- A change smaller than the margin is not yet a change. Wait for it to hold over another period, or collect more answers.
A worked example
These numbers are illustrative, for a fictional project-management company called Larkspur. It tracks 50 prompts against four competitors on one engine, and runs each prompt four times in a month: 200 answers.
- Larkspur is mentioned in 60 of the 200 answers. Mention rate: 30%.
- A Larkspur page is cited in 22 answers. Cited URL rate: 11%.
- Across the 200 answers, the five brands are mentioned 240 times in total, 60 of them Larkspur. Share of voice: 25%.
The margin on the 30% mention rate is 1.96 × √(0.3 × 0.7 ÷ 200), which is about 6.4 points. The true rate is plausibly anywhere from roughly 24% to 36%.
The next month, Larkspur's mention rate reads 34%. That is inside the margin, so it is not yet evidence of improvement. If Larkspur ran each prompt 16 times instead of 4, it would have 800 answers and a margin of about 3.2 points, and a four-point move would start to mean something.
Share of voice is harder to put a simple margin on, because one answer can mention several brands. Treat only large or sustained moves in it as real.
What moves the numbers
The levers are the ones covered in How to Get Cited in ChatGPT, Perplexity, and Google AI Overviews: let the engines' search crawlers in (see the AI crawler audit), publish pages that answer the prompts directly with sourced claims (see how to audit content for unsourced claims), and earn mentions in the third-party sources each engine leans on.
Reporting it
A report worth reading has, for each engine: the period, the number of answers, the three rates, the margin on the mention rate, and the change from the last period with a note of whether it exceeds the margin. Show at least three periods before calling a trend, and annotate the content changes you made, so movement can be linked to cause.
AI visibility tracking in Veritas records mentions, citations and share of voice against the competitors you name across ChatGPT, Perplexity, Gemini and Google AI Overviews. Our comparison of AI visibility tools covers the alternatives, including how often each re-runs your prompts, which directly affects how quickly the margins shrink.
Want to see which AI answers mention your brand?Start a 7-day trialSee pricing