EN
Free audit

What is AI share of voice, and how do you measure it?

Updated 4 min read

Short answer

AI share of voice is the percentage of AI answers to your customers' questions that mention your brand, compared with competitors. Measure it on a fixed set of 15–30 questions, across several models, at least weekly, and read trends over several runs, because single answers vary.

What is share of voice in AI answers?

In advertising, share of voice is your part of all the advertising in a category. In AI search it means something simpler and more useful: out of all the answers AI assistants gave to the questions your customers ask, in how many were you named?

If you put 20 questions to 4 models, you get 80 answers. If 24 of them name your brand, your AI share of voice is 30%. If your main competitor appears in 48, theirs is 60%.

Which metrics should you track?

MetricWhat it showsHow to calculate
Share of voiceHow often you are namedAnswers mentioning you ÷ all answers
Average positionHow prominentlyOrder of first mention among tracked brands
Share by modelWhere you are weakShare of voice per assistant
Search vs memoryWhy you are weakCompare answers with web search and without
SentimentHow you are describedPositive, neutral or negative mention
Own-site citationsWhether AI trusts your siteAnswers citing your domain as a source
Missing questionsWhere to actQuestions where competitors appear and you don't

How do you build the question set?

The question set is the most important design decision, because every number depends on it.

  • Use 15–30 questions. Fewer gives noisy results; more costs time and API quota without much new information.
  • Cover the main intents: category ("best online pharmacy in Romania"), problem, comparison, price and local questions.
  • Never include your brand name in the question.
  • Write in the language and phrasing of your customers, for each market separately.
  • Keep the set stable. If you change questions every week, you can't compare weeks.

How to collect real phrasing is covered in how to check if AI mentions your brand.

Why do results change between runs?

Language models generate text with some randomness, assistants with web search may pick different sources each time, and providers update models without notice. The same question can name four brands today and three tomorrow. That's why:

  • one answer is an anecdote, not data;
  • compare percentages over the whole question set;
  • look at three or four weekly runs before drawing conclusions;
  • treat changes of a few percentage points as noise unless they persist.

What is the difference between search answers and memory answers?

Assistants that search the web before answering (Gemini with Google Search, ChatGPT search, Perplexity) reflect the current web: your pages, reviews, rankings. Models answering from memory reflect what was written about you before their training cutoff.

Comparing the two tells you what to fix:

  • Low in search answers, fine in memory: your current pages or mentions are the problem. Check crawler access and the cited sources.
  • Fine in search answers, absent from memory: you are new or rarely written about. Keep building mentions; memory will follow with new model versions.
  • Low in both: start with the basics: AI-readiness checklist.

How do you find out why competitors win?

Look at the sources cited in answers where competitors are named and you are not. These are usually review platforms, comparison articles, marketplaces and forums. Grouping them by domain gives you a source gap list: the specific sites where being mentioned would change the answers. Read more in where AI assistants get their answers.

How often should you measure?

Weekly is a good rhythm for most businesses: frequent enough to catch changes after fixes or publications, rare enough to stay within free API quotas. Before and after major changes (a site redesign, a new robots.txt, a PR campaign), run an extra check.

Can you measure ChatGPT directly?

The consumer ChatGPT app has no official API for its exact answers, and automating the app is against its terms. Monitoring tools therefore use models available through APIs as indicators: for example Gemini with Google Search grounding for search-based answers, and models such as Llama or Mistral for answers from memory. The numbers are not identical to what each ChatGPT user sees, but the trends move in the same direction, and they are measured the same way every week.

Rekomo calculates all the metrics above every week, compares them with the previous week and sends a short report with actions. Start with the free audit or see plans.