How to measure AI visibility without fooling yourself
AI answers change from run to run, so one screenshot proves nothing. How to sample prompts, which numbers to track, and how many answers you need before a change is real.
Ask ChatGPT "what's the best tool for X?" twice in a row and you can get two different shortlists. Models generate text with some randomness, and when they search the web the results they read shift too. A screenshot of the run where you were named first says very little about what your next buyer will see.
Measuring AI visibility means treating each answer as a sample and tracking rates.
Start with the right prompts
Use the questions your buyers ask, in their words. Sales calls, support tickets and the queries in Google Search Console are better sources than brainstorming. Most of the list should be category questions without your brand name ("best invoicing software for freelancers"), because those are where you win or lose new buyers. Keep a few brand questions ("is YourBrand free?") to check what assistants say about you.
Twenty-five to a hundred prompts covers most companies. Track them per engine, because ChatGPT, Gemini, Perplexity and Claude search and cite differently.
Ask each prompt more than once
One answer per prompt per week gives you a coin flip each week. Three to five answers per prompt per run is a practical minimum, and you get more precision by pooling answers across prompts and weeks.
The table shows why. It gives a 95% confidence range (Wilson interval) for a mention rate of about 50% at different sample sizes:
| Answers | Observed rate | Likely true range |
|---|---|---|
| 3 | 2 of 3 | 21% to 94% |
| 10 | 5 of 10 | 24% to 76% |
| 30 | 15 of 30 | 33% to 67% |
| 60 | 30 of 60 | 38% to 62% |
| 120 | 60 of 120 | 41% to 59% |
With ten answers, a move from 40% to 60% can be noise. With a hundred and twenty, it is very likely real. Report rates with their sample size beside them, and be suspicious of any dashboard that shows a percentage without one.
The numbers worth tracking
Mention rate is the share of answers that name you. It is the headline number and the one most affected by GEO work.
Average rank is your position in the recommendation order when you are named. Fifth of five and first of five both count as a mention, but buyers act on the top of the list.
Citation rate is the share of answers that link to one of your pages as a source. It tells you whether your own content is doing the work or whether you are being named because of other people's pages.
Fact accuracy is the share of claims about your brand that match reality. Wrong prices and missing features in AI answers cost you buyers who never visit your site to find out.
Separate the engines and the samples
Pooling every engine into one number hides the story. A rise on Perplexity and a fall on ChatGPT can cancel out. Keep each engine separate and look at a prompt-level view when the totals move, so you can see which questions changed.
Keep the raw answers, too. When a number drops, reading the answers tells you why: a new competitor in the list, a cited article that dropped you, a search query that now returns different pages.
Connect changes to what you shipped
Note the date you publish a page, earn a listing or fix a crawler block, then watch the prompts it should affect for the next two to three weeks. Answers that rely on live web search often respond within days of the page being indexed. Answers that rely on model memory change only with new model releases.
FAQ
How often should I re-test?
Weekly is enough for most teams. Daily runs cost more and mostly add noise unless you track a very large prompt set.
Can I measure AI visibility manually?
You can for a handful of prompts, but you need fresh sessions without your history, several runs per prompt and a consistent way to record ranks and sources. Beyond ten prompts it becomes a part-time job, which is the gap tools like Promptvane fill.
What is a good mention rate?
It depends on how crowded your category is. Compare yourself with the competitors named in the same answers rather than with an absolute target, and watch the trend.