Ask an AI engine which accounting software suits a small agency, and it names a few brands. Ask again and the list can change. Ask tomorrow and it may change again. Engines vary their wording, search the web afresh and update their models, so a single answer is one draw from a range of possible answers.
That makes AI visibility a rate: the share of answers that name you. And a rate read from a handful of answers is mostly noise. This article shows how much a given number of answers can tell you, using the same formulas RookAtlas uses.
On this page
One answer is a coin toss
A screenshot of one answer that names you, or leaves you out, says almost nothing about how often that happens. If an engine names your brand in three answers out of ten for a question, any single answer will show you named or not named, and both are normal.
So the useful question isn't "does ChatGPT recommend us?" but "in what share of answers to our buyers' questions are we named, and how sure are we of that share?"
What a range adds
A 95% interval gives the range of rates that fit what you observed. Narrow means the data pins the rate down; wide means it doesn't yet. RookAtlas uses the Wilson interval, which behaves well with small samples and with rates near 0% or 100%, where the textbook "plus or minus" formula breaks down. It's also why we write a range such as 22–40% rather than 30% ± 9: the range isn't symmetric around the estimate.
Here is the same observed rate of 30% read from more and more answers:
| Answers | Observed | 95% interval |
|---|---|---|
| 10 | 30% | 11–60% |
| 30 | 30% | 17–48% |
| 100 | 30% | 22–40% |
| 300 | 30% | 25–35% |
| 1,000 | 30% | 27–33% |
The first answers buy a lot of certainty and later ones buy less: the width shrinks with the square root of the number of answers, so halving it takes four times as many.
Repeat runs are not new answers
The table assumes every answer is independent. Real tracking isn't built that way: you choose a set of prompts and ask each one again and again. Answers to the same prompt tend to agree with each other, because the engine finds the same pages and leans the same way. Asking a prompt a fifth time tells you less than asking a new one.
Statisticians handle this by treating each prompt as a cluster. The result is an effective number of answers, smaller than the raw count whenever answers to the same prompt agree. With equal runs per prompt it follows a simple rule: effective answers = answers ÷ (1 + (runs − 1) × ρ), where ρ is how strongly answers to the same prompt agree.
Take 100 answers from 20 prompts asked 5 times each, 30 of them naming the brand. If five prompts name it every time, five name it once and ten never do, those 100 answers carry about as much information as 23 independent ones:
Same 30%, a much wider range. Nothing went wrong; the second row is simply honest about where the answers came from. This is also why adding prompts usually helps more than adding runs, once each prompt has been asked a few times.
Reading a change between two periods
Visibility moved from 28% to 34%. Is that real?
Check two things before you celebrate. First, did anything change besides your work: the prompts tracked, the engines asked or the markets? A pooled rate across engines depends on the mix, so compare the prompts asked on the same engines in both periods, each counted once, so that a prompt answered more often in one period can't move the figure on its own. Second, is the change bigger than run-to-run noise? Setting the two periods' intervals side by side isn't the test: the difference itself has an interval, computed on those paired prompts with prompts as clusters, and the change is more than noise only when that interval excludes zero. With fewer than about 10 paired prompts, that interval can't be trusted, so treat the change as observed, not tested.
Detecting small changes takes more answers than people expect. To detect a 10-point change around 50% with the usual 80% power takes roughly 390 effective answers in each period; a 5-point change takes roughly 1,570. A weekly report built on a few dozen answers can show a trend, but it can't separate a 5-point move from noise.
Rules of thumb
- Report a rate with its range, never a bare percentage.
- Under 10 effective answers, don't report a number at all. The range covers most of the scale.
- Under 30 effective answers, call it low confidence. Useful for direction, not for decisions.
- Under 10 prompts, call it low confidence too, however many answers each has: a few prompts can't speak for a whole panel.
- Prefer more prompts over more runs once each prompt has a few runs.
- Compare like with like: the same prompts, engines and markets in both periods.
- Call a change real only when the paired difference's interval excludes zero and the change holds for a second period.
How RookAtlas applies this
These are the rules the product follows, not just advice. Visibility and citation rates carry a 95% Wilson interval, computed with prompts as clusters. A rate under 30 effective answers or 10 prompts is marked low confidence, and under 10 effective answers or 3 prompts it gives way to the words "Too little data". A change between periods is computed on the prompts asked on the same engines in both, each counted once; from 10 such prompts it is tested against run-to-run noise, and it counts as a change only if the next period agrees. Share of voice and average position carry a 95% interval too, computed with prompts as clusters, and their changes take the same paired test on answers measured against the same competitors; a position change counts only prompts where you were named in both periods. The method page has the formulas.
Questions
Why a range like 22–40% instead of 30% ± 9?
Because the interval isn't symmetric around the estimate, especially for small samples and rates near 0% or 100%. Writing the two ends states it exactly; a ± would misstate one side.
What is an effective answer?
The number of independent answers your data is worth. Answers to the same prompt tend to agree, so 100 answers from 20 prompts can carry as much information as about 23 independent ones. Intervals are computed on the effective number.
How many prompts should I track?
Enough that each number you act on has at least 30 effective answers in the period you report. A visibility rate pooled across engines gets there quickly; one engine in one market needs more prompts, since it sees only its own answers.
Does asking the same prompt more often help?
Up to a point. A few runs per prompt show how much its answers vary. Beyond that, new prompts add more information than more runs of the same ones, because repeat answers tend to agree.



