What is AI share of voice?

The name comes from advertising, where "share of voice" compares how loudly a brand is heard against its competitors. Applied to AI search, the question is similar: of all the times an assistant answers the questions in your market, how many times are you in the answer and how many times are the others? You will find it in tool dashboards and agency proposals, almost always as a percentage. For a store, the number is only useful if you know what is behind it, because it is not a metric with a single definition the way a search impression is in Search Console, but a count that everyone makes their own way.

How it works

Any AI share of voice is a proportion: your presence on top, the presence of a group of brands on the bottom. There is no standard formula, and every piece of that proportion is a decision by whoever calculates it:

• The queries. Which questions go into the panel, who chose them, in what language and from what country.
• The assistants and the repetitions. Which ones they are run in, how many times each query is run and on what dates. Answers change from one run to the next.
• What counts as presence. An AI citation, with a link to your page; a mention of your brand without a link; or the amount of the answer's text that talks about you.
• Who is in the denominator. The competitors added to the total. With three brands you get one number and with ten, another.

What search engines give you is about you, not your competitors. The AI Performance report in Bing Webmaster Tools counts your site's citations in AI-generated answers, the average number of your pages cited per day, the phrases the AI used to retrieve your content and the citations for each URL.⁠^{[1]Microsoft, Webmaster Tools blog. Introducing AI Performance in Bing Webmaster Tools Public Preview} Google's generative AI performance report gives the impressions of links to your site in AI Overviews and in AI Mode, by page, country, device and date.⁠^{[2]Search Console Help. Generative AI performance report (Search)} That is first-hand data, and both cover only the property you have verified. The bottom of the proportion, everyone else's presence, is built by each tool running its own queries.

The closest published definition is academic. The paper that named GEO (Generative Engine Optimization) measured a source's visibility inside an answer as the share of the answer's words that sit in sentences citing it, with a variant that gives more weight to sentences that come earlier, because they are more likely to be read. The authors normalize those measures so that the sum across all the sources cited in an answer is one.⁠^{[3]arXiv (Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande; KDD 2024). GEO: Generative Engine Optimization} In practice, it is a split of the answer among its sources. But it is the metric of an experiment, on one answer at a time; how it is aggregated across many queries is up to whoever uses it.

Compare it with a share that does have a formula. In Google Ads, impression share is the percentage of impressions your ad receives out of the total it could have received: impressions divided by eligible impressions. Google estimates those eligible impressions using many factors, including internal auction data.⁠^{[4]Google Ads Help. About impression share} In AI answers nobody publishes that denominator, and Google warns that no third-party tool has access to its internal ranking or AI systems.⁠^{[5]Google Search Central. Optimizing your website for generative AI features on Google Search} An AI share of voice is a sample somebody designed, not a count of the total.

In a store

A backpack store runs ten buying queries, three times each, in a single assistant: thirty answers. In each one it notes which brands appear, how many times and whether they carry a link. It is a fictional example: the brands and the counts are made up to show that the way of counting changes the result.

Three counts over the same thirty answers, and three different shares: a little under a third in the first two rows, a fifth in the third. None is the right one or the standard one; each answers a different question. The first row says how many conversations you are in; the second, how much you are talked about; the third, how many times the answer leads to your site. For a store, the third is the one that can end in a visit, and it is also the one that drops the most when the store appears a lot but rarely as a source.

How to check

1. Ask for the list of queries. Who chose them, in what language and from what country. A panel built from the questions you already show up for gives a high number that says nothing.
2. Ask for the assistants, the repetitions and the dates. A single run per query is a sample of one.
3. Ask what counts as presence: linked citations, mentions without a link or amount of text. With the same data, each option gives a different number.
4. Ask who is in the denominator and why those brands and not others.
5. Cross-check with your first-hand data: citations per page in Bing Webmaster Tools and AI impressions in Search Console.⁠^{[1]Microsoft, Webmaster Tools blog. Introducing AI Performance in Bing Webmaster Tools Public Preview}⁠^{[2]Search Console Help. Generative AI performance report (Search)} They do not give the proportion, but they do tell you whether the trend you are being shown matches what the search engines see.

Related concept: AI visibility

AI visibility is your presence in AI answers measured on its own: how many times you appear, on which queries, with which pages. Share of voice puts it in relation to your competitors'. The difference matters because the two can move in opposite directions: if the answers in your category start naming more brands, you can appear more often than before and still end up with a smaller share of the total. Visibility says whether you appear; share of voice, how much of the space you take.

To keep going, the measurement and verification reading path lays out the rest of the terms, and the entry on LLM attribution explains how to tell whether that presence ends up as visits and orders.