Visibility score, share of voice and average position: how visibility in AI answers is measured

Three metrics describe visibility in AI answers. Mention rate — in how many of the answers you were mentioned at all. Average position — when you were mentioned, in what place in the list of names you appeared, because the first is read differently from the fifth. Share of voice — what part of all the mentions in the industry belongs to you compared with your competitors. The visibility score combines the first two into a single number between 0 and 100; share of voice is reported alongside it. Each of them is measured on discovery questions only — questions that do not mention your name — because a question that includes the name is only asking the model to repeat it.

Mention rate: how many times at all?

The simplest metric, and two details in it matter.

The denominator is discovery answers only. Not all the answers — only those asked without your business name. An answer to a question that included your name is not counted here at all.

Refusals are not counted. If an engine refused to answer, the answer comes out of both the numerator and the denominator. Otherwise a particularly cautious engine would lower your score over something you have no control over.

One more detail: if you were mentioned twice in the same answer, that is still one mention. We count answers that mentioned you, not the number of times the name appeared.

Average position: being first is not the same as being fifth

When a model answers "I recommend A, B and C", the place in the list matters. The first is read, the third is skimmed, the fifth usually is not.

The position is counted by the order in which business names appear in the answer, starting from 1. If you were mentioned more than once in the same answer, the best place is taken — not the average of them. The working assumption is that the first mention is the one that gets read.

When you were not mentioned at all, there is no position. It is not counted as "last" and not as a zero in the sense of a place — there is simply no value, and the answer contributes only to the denominator.

Share of voice: the only number that says anything about the market

A mention rate of 30% is excellent in an industry where everyone is around 20%, and miserable against a competitor sitting at 70%. Without competitors beside it, an absolute number is almost meaningless.

Share of voice is your part of all the business mentions in the answers: your mentions, divided by your mentions and your competitors' together.

And one detail that is easy to miss: only you and the competitors that were defined are counted. Other businesses the engines mentioned that nobody entered do not go into the calculation. The reason is practical — otherwise every random name an engine produced would dilute your share against companies you have never competed with, and the number would fall without anything having happened in the market.

The visibility score: how two of the three are combined

The formula, in full:

score = 100 × mention_rate × ( 1 ÷ √average_position )

Rounded to a whole number. An example: if you were mentioned in 40% of the discovery answers, and on average in second place —

100 × 0.40 × (1 ÷ √2) = 100 × 0.40 × 0.707 = 28

Three things worth noticing in this formula:

Share of voice is not in it. That is not an omission. Share of voice depends on who was defined as a competitor, and that is a decision of yours — if it entered the score, the score would change when a competitor was added to the list, without anything in the world having changed. A metric that moves when you change a definition is not a measurement. That is why it is reported separately, and why it is worth looking at both.

Position enters under a square root and not linearly. The difference between first place and second is far larger than the difference between seventh and eighth, and the square root reflects that.

Anyone who was not mentioned gets a zero, not an error. When there are no mentions there is no average position, and rather than divide by something undefined, the score is defined as zero.

Why only discovery questions count

This is the one decision here that anyone might disagree with, and so it is stated explicitly.

Ask "what do you know about the Cohen firm?" and the model will mention Cohen. It was asked to. That does not mean Cohen gets found — it means the model knows how to answer a question that already contains the answer.

A question like that measures reputation: what is said about you once it is already known who you are. That is a legitimate and interesting metric, and we report it separately — how many of the questions that named you addressed you, and how many of those did so positively or negatively. It is simply not visibility.

Mixing the two is what inflates a score until it means nothing. A question bank half of which is brand questions will give every business a high score, because half the questions guarantee a mention in advance.

When we refuse to give a score

If after all the filtering not one discovery question is left, we have nothing to compute.

In that case the report says "we could not measure" and does not show a zero. The difference is not semantic: a zero says "we asked and did not find you", and that is a business conclusion. "Not measurable" says "we asked no question that could find you", and that is a statement about the check, not about you.

That is also why every score line carries the number of discovery answers behind it. A number without its denominator cannot be checked.

How the two engines combine into one score

The check runs against ChatGPT and against Gemini, and each gets a score of its own. The combined score is an average of the two scores, at equal weight — not the formula re-run over the merged pile of data.

The difference matters: averaging positions between an engine that mentioned you first and an engine that did not mention you at all is not a meaningful quantity. An average of two scores, by contrast, is exactly what "50/50" means.

An engine that could not measure anything does not enter the average and does not drag it down.

So what counts as a good score?

We do not know, and we are not going to invent a number.

Any threshold you see — "above 40 is good" — is a guess, unless someone measured it in your own industry in Hebrew. We measure in Hebrew in Israel precisely so that such data exists, and until there is enough of it, the only number worth anything is a relative one: higher than the one you had last month, and higher than that of the competitor you care about.

That is also why the measurement repeats exactly the same questions. Wording that changes between runs makes every comparison worthless.

Want to see the three numbers for your own business? The check is free and takes about a minute.