What the range on a mention rate actually means

A mention rate without a range is a number pretending to be more certain than it is. If an assistant named you in 2 of 5 answers, the honest statement is not "40%". It is "somewhere between about 12% and 77%, and we would need far more questions to say anything tighter".

Where the range comes from

Every reading is a sample. You asked some questions, not every question a buyer could ask, and the assistant gave one answer out of the several it might have given. A range is the arithmetic that turns "2 of 5" into an honest statement about the thing you actually care about, which is how often you get named in general.

The range gets narrower as you ask more questions, and it narrows slowly. This is the part people find frustrating and it is the part that matters most: going from 5 questions to 20 helps a lot, and going from 100 to 120 barely moves it.

Why this changes what you can claim

Two numbers with overlapping ranges have not been shown to be different. If last month was 1 of 5 and this month is 2 of 5, the middle moved from 20% to 40% and the ranges overlap almost entirely. The honest report is "no change", even though the headline figure doubled.

This is the single most common way an AI visibility report misleads. A tool that shows you 20% then 40% with no ranges has told you that you doubled. You did not. You ran a small sample twice.

The three other places small samples mislead

The rate is the obvious one. These three cause as much trouble and get discussed far less.

• A position average. "Average position 2.4" over six mentions is not a position, it is six numbers. One answer that happened to list you first moves it a long way.
• A per engine comparison. Splitting fifty questions across five assistants leaves ten each, and ten is not enough to say one assistant likes you more than another. The difference you are looking at is usually noise.
• A per segment breakdown. Slicing the same small sample by city, product line or question type produces confident looking numbers on four or five questions each. Every one of those slices is a guess.

The rule that follows is simple: the more ways you cut a sample, the less each piece can tell you. A dashboard with many small numbers on it is usually a dashboard with one small sample behind it.

What to do with a wide range

Three things, in order.

Ask more questions. This is the only thing that genuinely narrows it.

Read the answers, not only the rate. On a small sample the rate is noisy but the content is not. Which sources were cited, who was named ahead of you, and what was said about you are all useful at any sample size.

Watch the direction over several readings rather than the jump between two. A number that rises across four consecutive readings is telling you something even when any two of them overlap.

The arithmetic, worked through

This section is arithmetic rather than an observation from a run. It uses round numbers so the overlap is easy to see. Start with twenty questions in January, named in three. The middle of that is 15% and the honest range runs from roughly 3% to 38%.

In February it is five of twenty. The middle is now 25%, which looks like a substantial gain, and the range runs from roughly 9% to 49%. Those two ranges overlap heavily, so the correct report is that nothing has been shown to change.

March is seven of twenty, April nine of twenty. Any two consecutive readings still overlap. But four consecutive readings rising is a different kind of evidence from one jump, and at that point the honest report is that the direction is real even though no single month to month comparison proved it.

Notice what would have happened with a tool that reports 15% then 25% with no range: it would have declared a two thirds improvement in January, and then had nothing to say when the next reading came back flat.

The engines agree with this, and said so

Across our own audit in September 2026, in batch after batch, the assistant declined to name a competitor as winning most often. Its reason was that its own previous answers had not produced a comparable, live tested result that would support such a claim.

That is the clearest statement we have that a single run does not establish a ranking, and it came from the engine rather than from us. If the thing generating the answers will not draw a conclusion from one pass, a dashboard should not either.

What a range cannot fix

A range describes sampling uncertainty. It says nothing about whether you asked the right questions, and a clean range on a badly chosen question set is a precise measurement of the wrong thing.

It also cannot rescue a contaminated set. If any question names your company, the number is inflated by construction and the range around it is just as wrong as the middle.

And it does not tell you whether being named produced any business. Mentions are not revenue, and we have not seen a reliable way to connect the two.

Common questions

Why does my rate move when I changed nothing?

Two reasons, and both are normal. The assistants give different answers to the same question on different days, and your sample is small. If the change is inside the range, nothing has happened.

How many questions do I need for a number I can report?

For a rate you would defend in a board meeting, think in hundreds rather than tens. For deciding what to write next, twenty is plenty, because the useful output there is the list of cited sources rather than the rate.

Should I report the middle or the range?

Both, always together, and never the middle alone. A rate with no range invites a comparison that the data cannot support, and somebody will make that comparison.

Is a wide range a sign the tool is bad?

Usually the opposite. A tool showing you a narrow range on a small sample is either using far more questions than it told you about or is not calculating a range at all.

Can I compare my rate to a competitor's?

Only on the same question set, taken at the same time, on the same assistants. Two rates from two different sets are not comparable at all, which is worth remembering when a vendor shows you an industry benchmark.

What does this mean for the numbers on this site?

We print the range next to every rate, and when two of our own numbers have overlapping ranges we say they have not moved, even when the middle of one is higher. It makes our own reports look less impressive than they could. It is also the only version of this that is true.

What to do first

Look at whatever AI visibility number you are currently being shown and ask two questions: how many questions is it based on, and where is the range. If nobody can answer the first, the second does not exist either, and you are being shown a guess with a decimal point on it.

How we know this

We read the assistants rather than guessing what they do. In our own audit of 1,260 saved answers from ChatGPT and Gemini across the United States, the United Kingdom and India, 82.0% of every citation landed on a site the business itself owned, and review directories took 0.7%. Counted as one domain per reading, 16,516 of 20,143.

That is the whole asymmetry: the pages you own are the pages an assistant quotes back to a buyer. See the dated readings and the method on our proof page.

Find out whether assistants name your business, on your own domain and your own buying questions.

By Aditya Goel, Founder, AI Knows Us. Reviewed 1 November 2026.