Share of voice (in AI answers)

Share of voice is a term borrowed from advertising, where it means one brand's spend or impressions as a fraction of a market's total. Carried into AI answers, the borrowing usually breaks. In almost every AI-visibility product — Attensira included — the number reported under that name is a mention rate: how often a brand was named in the answers that were actually collected. It is not a division of a fixed total between competing brands.

The distinction matters because the advertising reading invites arithmetic that the measurement does not support.

What the number actually is

A mention rate is a fraction with two halves that must be read together:

• The numerator is the successful runs in a window whose answer text named the brand.
• The denominator is all successful runs in that window.

A run is one prompt put to one model on one occasion. Failed runs are excluded from the denominator rather than counted as misses, so a day with collection problems yields a smaller sample rather than a depressed rate.

In Attensira's API a field named ShareOfVoice carries exactly this. Its published limitations say it plainly: the field is the brand's own mention rate under an older name, and it is not a share of anything. There is no metric in the product that divides a fixed total between brands.

Why the rates do not sum to 100%

Competitor rates use the same denominator as yours. Each is an independent fraction over the same set of runs, not a slice cut out of one pie.

The consequence is easiest to see with one answer. Suppose a model answers "best CRM for agencies" and names you and three competitors. That single run raises four rates at once. Nothing was taken from anyone. Add the four together and you get a total that can exceed 100% and describes nothing — there is no whole for those parts to be parts of.

This is the single most common misreading of the number, and it is the reason the advertising metaphor is worth abandoning. In advertising, one brand's gain is another's loss because the impressions are finite. In a generated answer they are not: a longer answer can name everybody.

What the rate inherits from detection

A mention rate is only as good as the mention detection underneath it, and Attensira documents that method rather than implying precision it does not have. A mention is a case-insensitive substring match of brand name or domain against the answer text. There is no entity resolution, no disambiguation against a knowledge graph, and no word-boundary check.

So a brand called Arc matches "march" and "search"; a brand called Notion matches "notional". The error this produces is systematic rather than random, which means averaging does not remove it and a longer window does not wash it out. If your brand name is a common word, read a handful of stored answers before trusting the rate — Attensira stores the verbatim answer text for every successful run precisely so that this is possible.

The same error sits inside every competitor rate you compare yourself against. If a competitor is named after a common word and you are not, the comparison flatters them, and vice versa.

Reading a rate honestly

Four habits make the number usable:

1. Read `n` before the percentage. A rate is reported as {value, n}. At small n a single different answer moves the figure a long way: one hit in three draws is 33%, and the next draw makes it 0% or 67%.
2. Distinguish null from zero. value: null means the thing was never measured. value: 0 with n > 0 means runs happened and none named you. Charting the first as the second invents a rejection that never occurred.
3. Compare whole days. Work is enqueued hourly but each prompt × model × country combination runs at most once per workspace-local day, so today's figure is partial until the day closes.
4. Do not treat movement as change unless it is tested. Attensira gates every delta with a two-proportion z-test at 95% significance and returns {real: false, value: null} when the movement does not clear it — which means the change could not be proven, never that it was zero.

What it does not tell you

A mention rate is a presence measurement. It says a model named you in some fraction of the answers collected. It does not say anyone read that answer, clicked anything, or bought anything, and no product in this category can close that gap: there is no identity join between an answer a model gave a person and a visit that person later made. Attensira states this in its limitations rather than leaving it to be inferred.

Connect the rate to commercial outcomes with your own analytics if you want to — and say out loud that you are inferring it.

See also AI visibility score for why blending this rate into a composite number loses the denominator, and AI brand mentions for what counts as a mention in the first place.