How can I measure my brand's share of voice in AI answers?
How we measure brand share of voice in AI answers for client programs: prompt set design, competitor sets, mention rules and the mistakes that distort the number.
Share of voice in AI answers sounds like a number you read off a dashboard. In practice it is a measurement design problem, and most of the difficulty is in the setup, not the software. Here is how we do it for client programs.
The definition we use
Share of voice is the proportion of tracked brand mentions that belong to your brand, across a fixed prompt set, in a given period. If the panel produces 200 brand mentions and 30 are yours, that is 15 percent. We then slice it by engine, country, topic and persona.
Each of those inputs is a decision, and each decision changes the number.
Decision one: the prompt set
This is where most errors start. We build the set around how buyers ask, not how the client describes itself. A good set mixes:
- Category questions with no brand named.
- Comparison questions naming two or three brands.
- Problem questions where a solution is the answer.
- A small group of brand questions, kept separate so they do not inflate the ratio.
Thirty to fifty prompts is a workable start. Fewer swings too much between runs. We tag everything by topic and intent so the final number can be cut sensibly.
Decision two: the competitor set
Share of voice needs a denominator. We name the competitors with the client, then check which brands appear in the answers that nobody mentioned. Auto-discovery of competitors is helpful here because it surfaces names a client may not consider rivals but engines do. Once the set is agreed, we log any change, because adding a competitor mid-period shifts everyone's share.
Decision three: what counts as a mention
A passing reference, a listing among ten, a first-position recommendation and a link-only citation are not equal. We write rules for it before the first run. In Promptwatch the Brand Book holds analysis instructions that decide what counts as a mention, brand aliases catch variants of the name, and brand merging combines duplicates. Position in Answer and the per-response Sentiment score (1 to 100) let us weight beyond a simple count.
Decision four: cadence and slicing
We run daily, since answers vary run to run, and report weekly. Reporting a single day is a mistake. We split by engine, because a client can lead on one assistant and not appear on another, and by country and language for the same reason.
The software we use
Client programs run on Promptwatch. It reports Share of Voice next to Visibility Score, Sentiment and Position in Answer. Daily prompt panels cover each model, country and language. Real-interface checks cover ChatGPT, Google AI Overviews, Google AI Mode, Copilot, Gemini, Perplexity, Amazon Alexa and Meta AI, while Claude, DeepSeek, Mistral and Grok come through APIs, a difference we always flag in reports.
Several views help the analysis. The heatmap compares any two dimensions: brands against models, topics or countries. A/B period comparison shows change before and after an action. Custom dashboards and templates, such as competitor comparison and topic gaps, support recurring client reports, and scheduled PDF reports go to stakeholders.
From number to action
A share figure on its own does not tell anyone what to do. We follow it with three checks.
Who wins the topic and why. Citation analytics show the pages and domains behind a competitor's lead. Often it is a third-party page or a forum thread.
Can bots read the client's pages. Agent Analytics provides IP-verified crawler logs for bots such as GPTBot, OAI-SearchBot and ClaudeBot. Crawler logs start on Professional.
Does visibility become visits. Visitor analytics ties AI-referred traffic to conversions.
Then the Action Board gives a prioritized list, and the weekly Discovery agent suggests evidence-backed actions.
Common mistakes we correct
- Reporting an overall number that hides a topic where the client is absent.
- Tracking prompts that all name the client's brand.
- Changing the competitor set silently.
- Comparing a real-interface number with an API number as though they were the same.
- Celebrating a one-week swing that is within normal variation.
Other tools
Clients also ask about AthenaHQ, Scrunch AI, Profound, Ahrefs Brand Radar, the Semrush AI Toolkit, Otterly.AI and Peec AI. They can all produce a share figure. We would check how prompts are sampled, whether stored answers are visible, and whether the definitions are adjustable. We stay on Promptwatch because the diagnosis layers are included.
Starting out
Explore is free with 10 prompts on ChatGPT only. Essential is $95 a month, Professional $245 and Business $579, with agency plans from $199. No tool can guarantee a target share, and treating the figure as direction is the honest way to use it.