Free tool
AI visibility sample size calculator
ChatGPT, Perplexity and Gemini give a different answer each time you ask. Before you test a change, find out how many prompts and runs you need to see a real difference. After the change, check whether the rise in citations is bigger than the noise. Both run in your browser.
1. Plan: how big a change can your panel show?
For example, with 10 prompts, a 20% cited rate today and 4 control prompts:
| Runs per prompt | Answers each side | Smallest rise it can show |
|---|---|---|
| 1 | 10 | 65 points |
| 2 | 20 | 48 points |
| 3 | 30 | 39 points |
| 5 | 50 | 31 points |
| 10 | 100 | 22 points |
A point is one percentage point of the cited rate: going from 20% to 35% is a rise of 15 points. If the smallest rise your panel can show is bigger than any change you could realistically make, add prompts or runs before you start, not after.
2. Read: did the cited rate really change?
The numbers filled in above double the cited rate, from 4 of 20 answers to 8 of 20. That sounds like a clear win, but the calculator reads it as a rise of 20 points with a likely range of -8 to 44 points: not clear yet. Twenty answers on each side is too few to tell a doubling from luck.
Why AI visibility needs more samples than people expect
Each answer either cites your site or it does not, so a single prompt is a coin toss with an unknown weight. The range around a rate shrinks only with the square root of the number of answers: four times as many answers halves the range. That is why one-off checks, and dashboards that ask each prompt once a week, rarely show a change you can trust. The guide prompt sample size explains how to set a budget, and change versus noise covers reading a result.
How the calculator works
Each rate gets a 95% Wilson interval. The change is the rise on the prompts you changed, less the rise on the control prompts over the same days, with the four intervals combined by the MOVER method so the range stays sensible near 0% and 100%. The result reads as a real rise or drop when the whole range is on one side of zero, and as no change when the whole range sits within 10 points either way. With fewer than 10 answers on either side it is always not clear yet.
Let RankEcho run the test
RankEcho runs this test for you. When you ship a change, SEO/GEO/AEO Intel freezes the page's numbers from Google Search Console, picks the site's unchanged pages as the comparison, and reads the result at 14 and 28 days with the same maths as this calculator. AI answers are re-asked on the same questions at 3, 7, 14 and 28 days. It is $29/month with a 3-day trial on one site. See how it works.
Want a starting rate first? The free AI visibility checker asks ChatGPT, Perplexity and Gemini real buyer questions about your site.
Frequently asked questions
Why do I need to ask the same prompt more than once?
AI assistants give different answers to the same question. One answer that cites you, or does not, says little. Asking each prompt several times turns yes or no into a rate, such as cited in 6 of 20 answers, and a rate can be compared before and after a change.
How many prompts should I track?
Enough that the smallest change you care about is larger than the noise. With a 20% citation rate today, a few dozen answers on each side can only show very large rises. Use the planner above with your own rate to see the smallest rise your panel could show.
What are control prompts?
Prompts you did not try to change, asked on the same days. If the control prompts' rate also rose, the engine itself changed, and your change gets credit only for the difference. RankEcho uses up to 4 control prompts per test.
What does the range mean?
It is a 95% range for the true change in the cited rate. If it is all above zero, the rise is unlikely to be chance. The calculator uses Wilson intervals, which stay sensible at rates near 0% and 100%.
Is my data sent to RankEcho?
No. The calculation runs in your browser.
Measuring Google instead? Use the SEO change calculator.
Turn the diagnosis into a shipped change: use the RankEcho Resources library for Find, Fix, and Prove playbooks, checklists, and templates.