Guide
What Is AI Share of Voice, and How Do You Actually Measure It?
Ask a marketer whether their brand is winning in AI search and most will describe a feeling, not a figure. They typed something into ChatGPT last Tuesday and their name came up, or it didn’t. A screenshot is not a measurement. AI share of voice is the percentage of AI-generated answers in your category that mention your brand, calculated from a fixed set of prompts run across multiple assistants rather than one query typed once. Semrush’s 2026 AI Share of Voice guide states the formula plainly: (brand mentions / total category mentions across a prompt set) x 100. Run forty prompts through four AI engines, count every brand named in the results, and the CEO gets a number instead of a shrug. Sourcive runs exactly this kind of check for a single company: enter a URL, and it puts real buyer-intent questions to the OpenAI, Anthropic and Perplexity APIs to see whether and how that company gets mentioned.
What does the AI share-of-voice formula actually count?
The formula has two halves. The numerator is how many times your brand gets named across a defined prompt set, run on a schedule, not sampled at random. Every brand mention in that same category, yours included, makes up the denominator.
Take a worked example. A marketer builds a 40-prompt set: “best CRM for a 10-person sales team,” “alternatives to Salesforce for startups,” and 38 more like them. Run across ChatGPT, Perplexity, Google AI Overviews and Claude, that’s 160 answers. Across those 160 answers, every competitor mentioned by name adds up to 340 brand mentions total. The marketer’s own brand is named 51 times.
51 / 340 x 100 = 15%.
Fifteen per cent is a number. A hunch is not. It also tells the marketer something a single ChatGPT check never could: whether they’re ahead of or behind the next brand on the list, and by how much.
What are the five metrics behind the single number?
A single AI share-of-voice percentage is a rollup. Underneath it, Semrush’s guide tracks five component metrics per engine, because “mentioned” and “cited” and “recommended” are not the same event, and treating them as one hides which lever actually needs pulling.
| Metric | What it counts | Worked example |
|---|---|---|
| Citation rate | Share of answers where the brand’s own URL appears as a linked source | 12 of 160 answers, 7.5% |
| Mention rate | Share of answers where the brand name appears anywhere in the text, linked or not | 51 of 160 answers, 31.9% |
| Recommendation rate | Share of answers where the brand sits inside an explicit shortlist, not a passing name-drop | 9 of 160 answers, 5.6% |
| Citation absorption | Of the answers that cite the brand’s page, the share where the answer’s wording actually draws on that page rather than the link sitting unused | 6 of 12 cited answers, 50% |
| Sentiment | Split of brand mentions into positive, neutral or negative framing | 40 neutral, 9 positive, 2 negative |
Mention rate is usually the highest of the five, because it counts every name-drop. Citation rate is usually the lowest. Most engines paraphrase, not link. A brand can have a healthy mention rate and a poor citation absorption score at the same time. That gap is the whole diagnosis: the brand gets named, but the engine isn’t actually pulling from its site.
How is AI share of voice tracked across ChatGPT, Perplexity, Google AI Overviews and Claude?
Four engines, one prompt set, run the same way every time. Consistency is the whole point. It’s what makes the resulting number comparable month to month.
The engines don’t behave alike. Google AI Overviews draws heavily on the same organic index that ranks its ordinary search results, so a page already ranking on page one has a head start. Perplexity leans on live web citations shown alongside the answer, which is why its citation rate tends to run higher than the other three. ChatGPT blends training data with browsing results when a query needs current information, so the same prompt asked in different sessions can pull different sources. Claude’s answers vary depending on whether web search is switched on for that session, which changes whether a brand mention comes from training knowledge or a live page. Sourcive automates this exact workflow for a single company, running the same fixed prompt set against the OpenAI, Anthropic and Perplexity APIs each time, rather than leaving it to a marketer’s memory of what they typed last week.
Run the same 40-prompt set across all four monthly and the totals move: 160 answers becomes 160 answers again next month, but the 51 mentions might become 58, or 44. That movement is the trend. A single check can’t produce it, because it changes which prompt gets typed as often as it changes which engine gets asked.
Common questions
Is AI share of voice the same as traditional media share of voice? No. Traditional share of voice assumes a fixed slot, won by one advertiser at a time. AI answers are synthesised, so a single response can name three or four brands at once, and often does. The denominator in the AI formula is every mention across every brand in that answer, not one slot per query.
How many prompts need to be in the set before the number means anything? Semrush’s guide and most practitioners treat 20 to 50 prompts per category as the floor for a stable percentage. Below that, one viral prompt or one odd answer can swing the whole result. A 40-prompt set, covering variations like “best CRM for a small sales team” and its close cousins, sits comfortably inside that range. Sourcive’s own audits run against real buyer-intent questions for exactly this reason: a single prompt is a screenshot, not a measurement.
How often should the prompt set be re-run? Monthly is the common cadence. The answer engines themselves change roughly that often, so a source cited in July can drop out by September without any change on the brand’s side. Re-running the identical prompt set is what makes month-over-month comparison mean anything. Anything less is a screenshot with extra steps.
Sourcive measures whether ChatGPT, Claude, Perplexity and Gemini mention your business when your buyers ask, and gives you the files to fix it. Run the free audit.