AI share of voice is the slice of brand mentions in AI answers that belong to you. Ask ChatGPT, Gemini or Perplexity the questions your buyers ask, count every brand each answer names, and divide your count by the total. The formula fits on one line. What the formula hides is the part that decides whether the number means anything: which denominator you used, how many answers sit behind it, and how wide the margin of error is. This post works one example through from raw answers to a range you can defend in a report.
What AI share of voice measures
Share of voice is an old advertising idea: your share of all the ads, or all the media coverage, in a category. Paid search has a strict version. Google Ads defines impression share as “the percentage of impressions that your ads receive compared to the total number of impressions that your ads could get”[11]. AI share of voice applies the idea to the answers assistants write, with one difference: there is no auction log of every eligible impression, so the total has to be sampled. The question it answers is not “does ChatGPT know us?” but “when ChatGPT answers a buyer’s question, how much of the recommendation goes to us, and how much to the rivals we care about?”
It is a relative number, and that is its strength and its trap. It rises when you are named more often, and it also rises when a rival is named less often. It depends on which rivals you chose to count. And because there is no fixed results page behind an AI answer, it is always an estimate from a sample of answers, not a reading of a list that exists before you look. In the SparkToro and Gumshoe.ai study, there was less than a 1 in 100 chance that ChatGPT or Google’s AI would give the same list of brands in any two responses to the same prompt[1]. A share of voice built on a handful of answers inherits all of that churn.
The formula, and the two you will confuse it with
Three ratios get called “share of voice” or “visibility” in AI tracking tools. They use the same answers and different denominators:
| Formula | Question it answers | |
|---|---|---|
| Mention rate (visibility) | Answers that name you ÷ all answers | How often do we come up at all? |
| Share of voice | Your mentions ÷ mentions of all tracked brands | How much of the conversation do we hold against named rivals? |
| First-place share | Answers where you are named first ÷ answers where any tracked brand is first | When an answer leads with someone, how often is it us? |
None is wrong. They answer different questions, and a report that switches between them without saying so will show movement that is only a change of denominator. Pick one as the headline, label it, and show the others beside it.
How the tools define it
The vendors who rank for this term agree on the basic share-of-mentions formula and then drift apart. Semrush gives “AI share of voice = (your AI mentions ÷ total AI mentions across all brands in your category) × 100”[5]. Nightwatch gives the same ratio over “total brand mentions across tracked prompts”[6]. HubSpot’s glossary uses both ratios in one entry: 25 of 100 mentions is a 25% share, and appearing in “30 out of 100 relevant responses while a competitor appears in 50” is described as 30% against 50%[7]. The first is share of voice; the second is mention rate. Ahrefs takes a third route: its Brand Radar sums the Google search volume of prompts where a brand appears, and defines AI share of voice as “a brand’s percentage share of impressions compared to other tracked brands”[9].
Two of those vendor notes are worth keeping. Ahrefs warns that the metric “can change a lot depending on how you’ve set up the entities for you brand and competitor brands”[9]: the rival list is part of the formula. And Nightwatch draws the line between the two ratios cleanly: a visibility score “measures the percentage of prompts where your brand appears at all. AI share of voice adds competitive context”[6]. What none of these pages gives is a margin of error, which is where the worked example below goes next.
How to calculate AI share of voice, step by step
Two rules keep the counting honest. Count a brand at most once per answer, otherwise an answer that repeats one name in every paragraph counts as several recommendations. And keep each engine separate. Profound tracked 15,155 brand and category pairs daily through May 2026 and found a median gap of 8 percentage points in a brand’s visibility between Gemini, AI Overviews and AI Mode, with a gap of more than 10 points for over a third of brands[3]. An average across engines hides that.
A worked example, with the error bars
A software company tracks 40 unbranded buyer questions on one engine, once a week for 4 weeks: 160 answers. It counts four brands: itself and three rivals. Across the 160 answers, the four brands are named 170 times in total, counting each brand once per answer.
| Answers naming the brand | Answers where it is named first | |
|---|---|---|
| Your brand | 46 | 22 |
| Rival A | 58 | 41 |
| Rival B | 41 | 38 |
| Rival C | 25 | 19 |
Step 1: share of voice
Your brand was named 46 times out of 170 tracked mentions. 46 ÷ 170 = 27.1%. Rival A, the most-named brand, has 34.1%, so on the point estimates you trail by 7.0 points.
| Your brand | 27.1% |
|---|---|
| Rival A | 34.1% |
| Rival B | 24.1% |
| Rival C | 14.7% |
Step 2: the confidence range
Here is where most reports stop, and where they go wrong. 170 mentions is a sample. Ask the same 40 questions for another four weeks and the counts will differ, because the answers do. A 95% confidence interval says how far the true share could plausibly sit from the one you measured.
We use the Wilson score interval rather than the textbook “plus or minus” formula. The NIST handbook notes that with Wilson “the lower limit cannot be negative”, which the common formula can produce for small shares[12], and a widely cited review of the options recommends “the Wilson interval or the equal-tailed Jeffreys prior interval for small n”[13]. Share of voice for a small brand is exactly that case: a low share from a modest sample.
| Share of voice | 95% range | Tied for the lead? | |
|---|---|---|---|
| Rival A | 34.1% | 27.4% to 41.5% | Yes |
| Your brand | 27.1% | 20.9% to 34.2% | Yes |
| Rival B | 24.1% | 18.3% to 31.1% | Yes |
| Rival C | 14.7% | 10.2% to 20.8% | No |
Your range runs from 20.9% to 34.2%. Rival A’s runs from 27.4% to 41.5%. The ranges overlap, so the 7.0-point gap that looked like a clear second place is inside the noise. On these numbers, Rival A, Your brand and Rival B are statistically tied for the lead. A dashboard that ranks them 1, 2, 3 is ranking noise. Rival C, on the other hand, is clearly behind you: its range does not reach yours.
A useful shorthand is the noise floor: the smallest gap between two brands that means anything at this sample size. With 170 mentions it is about 14.8 points (twice the worst-case half-width). Gaps smaller than that should be reported as ties.
Step 3: the same answers, divided differently
Now compute the other two ratios from the same counts. Mention rate divides by the 160 answers, not by mentions: your brand appears in 46 of them, 28.8% (range 22.3% to 36.2%). First-place share divides by the 120 answers where one of the four brands came first: you led 22 of them, 18.3%, against 34.2% for Rival A.
| Mention rate | Share of voice | First-place share | |
|---|---|---|---|
| Your brand | 28.8% (22.3% to 36.2%) | 27.1% (20.9% to 34.2%) | 18.3% (12.4% to 26.2%) |
| Rival A | 36.3% (29.2% to 43.9%) | 34.1% (27.4% to 41.5%) | 34.2% (26.3% to 43.0%) |
| Rival B | 25.6% (19.5% to 32.9%) | 24.1% (18.3% to 31.1%) | 31.7% (24.0% to 40.4%) |
| Rival C | 15.6% (10.8% to 22.0%) | 14.7% (10.2% to 20.8%) | 15.8% (10.4% to 23.4%) |
The story changes with the denominator. On share of voice you look roughly level with the pack. On first-place share you are well behind: you are named often, but rarely first. SparkToro saw exactly this pattern in real answers: one brand appeared in 69 of 71 ChatGPT answers but was mentioned first in only 25[1]. If buyers read the first name and stop, the first-place share is the one to worry about.
How many answers before a gap is real
The range narrows as mentions pile up, roughly with the square root of the sample: four times the mentions halves the range. This table shows the noise floor at different totals, computed with the same function:
| Gaps smaller than this are ties | |
|---|---|
| 50 mentions | 26.8 points |
| 100 mentions | 19.2 points |
| 200 mentions | 13.8 points |
| 500 mentions | 8.8 points |
| 1,000 mentions | 6.2 points |
| 2,000 mentions | 4.4 points |
Two caveats make the real picture a little worse than the table. First, mentions in one answer are not independent: if an answer names three rivals, it names fewer slots for you. The interval treats every mention as a separate coin toss, so read it as a best case. Second, answers to the same prompt in the same week tend to lean the same way. More prompts usually buy more precision than more runs of the same prompt.
Independent work points the same way. A 2026 statistics preprint that sampled Perplexity, SearchGPT and Gemini answers argues that visibility metrics “should be treated as sample estimators of an underlying response distribution rather than fixed values”. It found that “overlapping confidence intervals of this kind are the norm rather than the exception for domains that appear to differ in citation share by less than 5-7 percentage points”[4]. That study measured cited domains, not brand names, and used a different interval method, but the lesson carries over: a few points between two names is usually noise.
This is the reason SparkToro’s conclusion was not “stop measuring” but measure rates across many prompts and many runs: Rand Fishkin calls visibility across dozens to hundreds of prompts, run multiple times, a reasonable metric[1]. Our guide to choosing prompts for LLM tracking covers how many to start with.
AI visibility score: what is inside it
Many tools show a single “AI visibility score” instead of, or next to, share of voice. A score like that is a blend of the counts above, often with position and sentiment mixed in, and each vendor weights them its own way. Semrush’s score runs from 0 to 100 and “combines two factors”: topic coverage and mention consistency[8]. HubSpot’s AI Search Grader scores five dimensions out of 100, with sentiment worth up to 40 points and share of voice up to 10[10]. Both are reasonable designs, and they measure different things. That makes scores useful for watching one tool’s trend and useless for comparing two tools.
Before you trust a score, ask four questions:
- Which counts go in? Mentions, citations, position, sentiment. A score that mixes citations (links to your site) with mentions (your name in the text) can rise when you are cited for a how-to article and named in no recommendations.
- What is the denominator? All answers, answers that name any brand, or all tracked mentions. The worked example above shows how far that moves the number.
- How many answers sit behind it? If the tool cannot tell you, it cannot give you a range either.
- Are engines blended? A single score across ChatGPT, Gemini and AI Overviews hides gaps that the Profound data shows are often wider than 10 points[3].
The AI visibility metrics worth keeping
Share of voice is one line in a report. These are the ones that, together, explain it:
| What it counts | Use it to | |
|---|---|---|
| Mention rate | Share of answers that name you | See whether you come up at all |
| Share of voice | Your share of all tracked brand mentions | Compare yourself with named rivals |
| First-place share | Share of answers that name you first | See whether you lead or trail in the list |
| Citation rate | Share of answers that link to your site | See whether your pages are used as sources |
| Consistency | Share of weekly runs a brand appears in, per prompt | Tell a held position from one that rotates |
| Description | How the answer describes you | Catch outdated or narrow positioning |
The consistency line matters most for planning. A brand at 30% share of voice that holds the same few prompts every week is in a different position from one at 30% that appears on different prompts each week. SISTRIX found the same split in the sources AI answers cite: for brand queries, 43% of brand domains were present in all 17 weeks it tracked, while the domains cited alongside them rotated at around 70% a week[2]. Our product calls the two groups Core and Carousel: core is a brand present in at least 70% of weekly runs for a prompt, emerging from 40%, carousel below that, and nothing is classified before 4 weeks of data.
Tracking it every week
A spreadsheet works for a first look. To spot-check one question first, the free ChatGPT rank tracker lists every brand an answer names, in order. Run it a few times and you will see the counts move, which is the point of this whole post. For ongoing tracking, CoreCited’s competitor tracking reports share of voice with a 95% range on every brand and marks ties instead of ranking them. The Starter plan covers 40 prompts for $69 a month; every limit is on the pricing page.
Questions people ask
What is AI share of voice?
AI share of voice is the share of brand mentions in AI answers that belong to you. Run a fixed set of buyer questions through ChatGPT, Gemini or another assistant, count every time each tracked brand is named, and divide your count by the total for all tracked brands. It tells you how much of the conversation you hold against named rivals.
How do you calculate share of voice in AI search?
Divide your brand's mentions by the mentions of every tracked brand, you included, across the same answers, then multiply by 100. Count a brand once per answer, keep the prompt set and rival list fixed, report each engine separately, and put a confidence range next to the result, because the answers change from run to run.
What is an AI visibility score?
Usually a vendor's composite number built from simpler counts such as mention rate, citation rate, position in the answer and sometimes sentiment. Each tool weights them differently, so scores from two tools are not comparable. Ask which counts go in, how they are weighted, and how many answers sit behind the number.
What is a good AI share of voice?
There is no universal benchmark: it depends on how many rivals you track and how crowded the category is. With four tracked brands an even split is 25% each. The more useful questions are whether you lead, whether the gap to the leader is bigger than the margin of error, and whether the trend holds for several months.
How many AI answers do I need for a reliable share of voice?
More than most dashboards use. With 200 brand mentions in the period, two brands need to be about 14 points apart before the gap clearly clears the noise; with 1,000 mentions, about 6 points. Below that, report brands whose ranges overlap as tied rather than ranking them.
What is the difference between AI share of voice and AI visibility?
Visibility, or mention rate, divides by answers: the share of answers that name you. Share of voice divides by mentions: your slice of all brand names in those answers. Visibility can rise while share of voice falls, if rivals gain faster than you do, so track both.