What AI share of voice actually measures

What is a share-of-voice number actually counting?

Two different questions get sold under one label. The first is a question about you: across the questions your buyers ask, how often does an engine name you at all. The second is a question about the field: of all the company names an engine produced across those questions, what portion were yours. The first is a citation rate. The second is share of voice. They move independently, and a brand can look strong on one while being marginal on the other.

The gap is easy to demonstrate. A small set of questions where you are named most of the time gives a high citation rate on a thin sample. If each of those answers also names four competitors, your share of the naming is small, and the shortlist a buyer reads has four other companies on it. Share of voice is the harsher of the two numbers because it counts the competition into the denominator rather than leaving them out of frame.

Neither number is meaningful on its own. Both are ratios computed over a set of prompts that somebody chose, run some number of times, at some point in a month. Change any of those three and the number changes without a single answer having changed. That is the whole subject of this essay: not what the metric is called, but what has to be published beside it before it means anything.

Why a rate and a union are different numbers

An engine does not return the same answer twice. The same prompt, asked again an hour later, can name a different set of companies. So any figure has to decide what to do with repetition, and there are two honest choices that produce very different numbers.

A per-run rate asks: of every prompt run we executed, in what fraction were you named. It converges as you add runs. It goes down when you are absent and up when you are present, which is what a reader expects a visibility figure to do.

A union asks: on how many prompts were you named at least once, across all the runs we did. That quantity can only rise. Every additional run is another chance for a hit and no chance for a miss, so a union climbs toward one hundred percent as a function of how long the tool has been running rather than as a function of anything you did. It is a useful number, correctly labelled. It describes reach: the breadth of questions where you can appear at all. Labelled as visibility, it flatters, and it flatters more every month.

The same capture, read three ways

Take a hypothetical capture. None of the figures in this paragraph describe any real account; they exist to show that one set of answers supports several defensible numbers. Twelve prompts, each asked three times of one engine, is thirty-six prompt runs. Suppose you are named in fourteen of those thirty-six, and that those fourteen fall across nine of the twelve prompts. Suppose the thirty-six answers named two hundred and ten companies in total, counting every mention.

Three numbers follow, and all three are true of the same capture. The per-run rate is fourteen of thirty-six, or 38.9 percent. Reach, the union, is nine of twelve prompts, or 75 percent. Share of voice is fourteen of two hundred and ten mentions, or 6.7 percent. A provider who wants the chart to look good reports 75 percent. A provider who wants it to be comparable next month reports 38.9 percent and says how many runs it rests on.

We shipped that error in our own product. Until 2026-06-02 our leaderboard computed the union across all runs and labelled it visibility, which meant the headline figure for an internal account read 75 percent where the corrected per-run rate was 38.2 percent. Nobody had gamed it. The metric simply answered a different question than its label claimed. We renamed the union to reach, made the per-run rate the headline, published both definitions and the sample size beside them, and the regression that proves the fix is that the rate now holds steady across added runs while the old union climbed with every one.

Why the denominator is the whole argument

Every one of these figures is a fraction, and the numerator is the part a provider does not control. The denominator is the part they choose. That asymmetry is where most of the dishonesty in this category lives, and almost none of it requires anyone to lie.

Drop the prompts you never win and share rises. Add prompts that contain your brand name and it rises again, because an engine asked about you will usually mention you. Restrict the field to three named competitors instead of every company the answers mentioned, and your share of voice can double while the answers a buyer reads stay exactly the same. None of those moves alters a single captured answer. All of them alter the number.

The defence is to fix the set before you measure and to treat any change to it as a new baseline rather than a continuation of the old line. Freeze the prompts, version the set, stamp every run with the version it was executed under, and compute each month inside the set that was live when those runs happened. A figure computed against today's prompt list while scanning last month's runs is not a trend. It is a rewrite, and it is the subject of the next essay in this set.

What to ask before you accept one

Six questions separate a measurement from a chart. What is the prompt set, is it frozen, and can I see it. How many times was each prompt run. What is the window, stated as dates rather than as a month name. Is this figure a rate or a union. Which engines are covered, and which are not covered and why. Below what movement do you call a change noise rather than progress.

A provider who can answer all six is not necessarily right, but their number can be checked, and a number that can be checked is a different class of object from one that cannot. Ask us the same six. Our own answers are on the method disclosure at the foot of every page on this site, including this one, and the last question has a specific answer: we hold a figure back rather than publish it when it has not run enough times to separate a real reading from run-to-run variance.