Share of voice (in AI answers)
Your brand mentions as a proportion of all brand mentions across a tracked prompt set. Useful against your own history, meaningless as an absolute number.
Share of voice in AI answers is your brand's mentions as a proportion of all brand mentions across a tracked set of prompts. If five brands appear across your prompt set and one fifth of the mentions are yours, your share of voice is 20%.
How it is calculated
Take a prompt set. Send it to several assistants on a schedule. Count every brand name that appears in the replies. Divide yours by the total.
Some tools weight by position, so a first mention counts more than a fourth. Some count only the brands you told them to track, which makes the denominator smaller and your share larger. Check which one you are looking at before you celebrate a number.
Where it is useful
Against your own history. With a fixed prompt set, a rising share of voice means something changed. That is a real signal.
Against named competitors, inside the same set. If you and three rivals are tracked on the same prompts, the relative ordering is meaningful.
Where it is not useful
As an absolute number. The score is a function of the prompts you chose. Pick prompts your product is obviously right for and the number goes up without anything improving. There is no scale.
Compared across tools. Different products parse mentions differently, cover different engines and refresh at different rates.
As a proxy for adoption. This is the important one for a developer tool.
Why it does not predict adoption
Share of voice measures a person reading an answer. A large and growing share of developer tool adoption happens somewhere else: a coding agent inside a repository, choosing and installing.
Those two things come apart, measurably:
| Finding | Why share of voice cannot see it |
|---|---|
| Same request: one winner in 111 of 111 JavaScript sessions, 48 of 132 in TypeScript | A chat prompt has no repository |
| A written instruction lost 2 of 3 once a rival was installed | A chat prompt has nothing installed |
| The three coding agents disagreed on the leader in 10 of 18 categories | A chat model is a fourth decider again |
| Claude Code searched the web in 1.6% of decision runs | Content is absent from most of its decisions |
The equivalent metric for that surface is install share.
How to move it
The tactics are ordinary and they work.
- Answer the tracked questions directly, in the first 40 to 60 words under a question heading.
- Be specific. One dated number beats ten adjectives.
- Keep pages current. AI answers favour recency more strongly than classic search did.
- Earn third party sources that describe you the same way you do. Models weight consensus over self-description.
- Stay fetchable: no JavaScript-only rendering, no gates, AI crawlers allowed.
The honest summary
Track it. It is cheap and it tells you what a person hears when they ask about your category.
Do not report it to a board as though it were market share, and do not let it stand in for the number that ends in installed code.
Common questions
What is share of voice in AI answers?
Your brand's mentions as a proportion of all brand mentions across a tracked set of prompts sent to AI assistants. If five brands appear across your prompt set and you are a fifth of the mentions, your share of voice is 20%.
Is share of voice a reliable metric?
It is reliable against itself and unreliable against anything else. Because the number depends entirely on which prompts you chose, it cannot be compared with a figure from another company, another tool, or another prompt set.
How is it different from install share?
Share of voice counts mentions in answers a person reads. Install share counts sessions where a coding agent actually installed the product. One is attention, the other is adoption.
How do you improve share of voice?
Publish content that answers the tracked questions directly and specifically, keep it current, and earn third party sources that describe you the same way. Models weight consensus across sources more than they weight your own claims.
Where this comes from
Armature ran 5,292 judged sessions with Claude Code, Codex and Cursor inside 51 realistic codebases, and published every run. The numbers on this page come from that work.