AI share of voice: the formula, a worked example and its limits

Formula

share of voice = entity mentions ÷ all entity mentions in scope × 100

The denominator is what makes share of voice useful, and what makes it fragile. It counts every brand the answers name, yours included, so your share moves when a competitor gains ground, even if nothing changed for you.

Two counting rules decide the result. The first is what counts as a mention: every time the answer names the brand, or only once per answer. Counting every mention rewards the brands an answer discusses at length, which is usually what you want to measure. The second is which brands count. An open scope includes every brand the engines name; a closed scope keeps only the competitors you list. A closed scope reads more cleanly over time, and misses a new entrant until someone adds it to the list.

The definition used across this site settles both: an entity's mentions divided by all entity mentions in scope, every mention counted, aliases merged first^[1].

Here is how the two split on a sample of ten answers to an agency prompt. Each answer recommends Cadence and Northbeam over several sentences, with three and two mentions, and names Loopwork once at the end. Loopwork's visibility is 100, since every answer names it. Its share of voice is one mention in six, about 17%.

The reverse happens too. A brand named in only a third of the answers can hold a larger share of voice than one named everywhere, if the answers that name it discuss it at length. Northbeam could be that brand on prompts about time tracking: absent from most answers, and the main subject of the few that name it.

Report the one that matches the question. A board asking whether AI engines know the brand wants visibility, covered in the guide to measuring AI search visibility. A sales team asking who wins the comparison wants share of voice, read topic by topic.

The marketing definition is a plain ratio: a brand's impressions divided by the impressions of every brand in its category, times 100. The same calculation works on mentions, so 500 mentions out of 1,000 make a 50% share^[2].

The IPA analysis covered 700 brands in 83 categories. It placed each brand's equilibrium where share of voice equals share of market: brands above that line tended to gain share, and brands below it tended to lose it^[3].

AI answers change the mechanics. In advertising, share of voice follows media budget. In an AI answer, it follows the pages the engine reads and the way it summarises them, which ties it to content and PR.

The engines are starting to publish share metrics of their own. Microsoft added a Citation Share to Bing Webmaster Tools in 2026: a site's citations as a percentage of all citations for the same grounding query, which Microsoft calls an observational metric rather than a ranking^[4]. The research paper that introduced generative engine optimization also scores each cited source as a share of the answer, rescaled so that the shares of all sources in one answer add up to one^[5].

None of this shows that the growth link measured in advertising carries over to AI answers. Treat AI share of voice as a competitive reading, and check it against your own pipeline before reading growth into it.

1. Step 1

   Fix the scope

   Choose the prompts, engines, countries and dates, then list the brands that count. Keep that list unchanged for as long as you compare periods.

2. Step 2

   Run the prompts and keep every answer

   Ask each prompt on each engine in the chosen countries and save the full answers. Run each prompt more than once, since answers vary between runs.

3. Step 3

   Count mentions per brand

   Count every time each brand is named, after merging its aliases. Add up the counts of all brands to get the denominator.

4. Step 4

   Divide, then automate

   Divide each brand's mentions by the total and multiply by 100. The AI Visibility API returns share of voice per entity, per engine and per country, so the count never has to be redone by hand. See the AI Visibility API.

• The brand list sets the score. Add two competitors to the scope and every other share drops, though no answer changed.
• Repetition inflates it. An answer that compares Cadence feature by feature can name it six times and hand it six mentions.
• Tone is invisible. A mention that warns buyers away counts as much as a recommendation.
• Long lists dilute it. An engine that lists ten tools spreads its mentions thinner than one that lists three.