Skip to main content
Share of voice is the fraction of successful runs in which a model named your brand in its answer text. It is the count of successful runs where your brand was named, divided by every successful run in the window. Nothing else enters the calculation: not link counts, not answer length, not position.
Share of voice here is a rate, not a slice of a pie. It answers “in what fraction of answers were you named?” — never “what fraction of all brand mentions were yours?” It does not sum to 100% across brands. Most tools in this category use the name for the pie version; we do not, for the reason in Why ours does not sum to 100% below.

​
What counts?

A mention is your brand name appearing in the answer the model wrote. No link is required. A model can describe you at length, recommend you, and never attach a URL — that still counts. A link to your domain is a separate thing entirely, counted as a citation. The unit of measurement is a run: one prompt, sent to one model, for one country, at one point in time. Not a reading, not an answer paragraph, not a keyword occurrence. If your name appears eleven times in one answer, that run contributes exactly one mention, the same as an answer that names you once.

​
How is it calculated?

Failed runs are stored for debugging but excluded from the denominator, so a model outage lowers your sample size rather than your rate. See How we collect data for what makes a run succeed.

​
Does sampling more often change my number?

No, and this is deliberate. Reading the same prompt three times instead of once triples the runs behind the number — and roughly triples the URLs those answers carried — but both the numerator and the denominator grow together, so the rate holds and only n grows. Sampling deeper buys you confidence, not a different number. The same property is what makes windows comparable. A window read at greater depth than the one before it is not “up”; the significance gate compares two rates using both sample sizes, so a change has to clear the noise floor for the n actually involved. Two things do scale with sampling depth, and they are counts rather than rates: the number of citations collected, and the number of answers read. Read those as evidence volume, never as movement.

​
Why ours does not sum to 100%

Competitor share of voice uses the same denominator — all successful runs in the window. That has a consequence worth internalising: competitor rates do not sum to 1. Several brands can be named in one answer, or none can. A window where you sit at 40% and three competitors sit at 55%, 30% and 20% is perfectly ordinary and describes reality accurately. We could compute the version that does sum to 1 — your mentions divided by every tracked brand’s mentions. We deliberately do not, because its denominator would be the competitor list you typed. Add a competitor and your headline number falls; remove one and it rises. Nothing about your visibility would have changed. A number that moves when you edit a settings page is not a measurement, and we would rather publish one that only moves when the models do. This means our share of voice is not comparable to a number under the same name from a tool that computes the pie version. Ours will generally look larger, because it is answering a different question.

​
How do I read the number?

Every rate arrives as {value, n}. value is the rate; n is the number of successful runs behind it. A value of null means the cell was never measured — that is not zero. A value of 0 with n above zero is a real, measured zero: the model was asked and did not name you. Read n before you read value. At a sampling depth of one run per prompt per model per day, a single week gives you seven runs per prompt-model pair, and a rate built on seven runs moves in large steps. Reading the numbers covers small samples, deltas and unmeasured cells in detail.

​
What this measure gets wrong

Detecting a mention means finding your brand name in the answer text, as a whole word. That handles the ordinary cases — possessives, hyphenation, your domain inside a URL — and it will not count “architecture” as a mention of a brand called Arc. It cannot yet tell a brand from its own name used as an ordinary word. If your brand is called Linear, an answer explaining that something “scales in linear time” is counted as a mention. The error runs one way — brands whose names are common words read high, never low — and it is worth knowing about when you read your own number. We store every answer verbatim, so when the detection improves we can re-run it over history rather than starting a new series. If your brand name is a common word, open a few answers on the prompt detail view and read them before you trust the rate.