What Is Share of Voice in AI Search and How to Calculate It

Share of voice in AI search tells you how much of the AI conversation in your category belongs to you. If you run growth or marketing at a seed to Series B AI startup and the board has asked whether AI engines recommend you or your competitor, this is the number that answers it with more than a screenshot. The formula is a simple ratio. Almost all of the work is in choosing, and then freezing, the prompts, engines and competitors that go into it.

What does AI share of voice measure?

AI share of voice measures how often engines name your product, relative to named competitors, when buyers ask the questions that decide your category. A mention is counted once per answer. Ahrefs' AI visibility metrics documentation uses the same rule: an answer that names Salesforce three times still counts as one mention, so a long answer that repeats one brand does not distort the result.

It is a relative metric. It tells you whether you are gaining or losing ground against a fixed set of rivals, not how many people saw your name. For reach you need impression estimates, and for business impact you need referral sessions and signups, which we cover in tracking AI search referral traffic in GA4.

How is AI share of voice different from search share of voice?

Search share of voice is built from rankings, which hold still for days and have a clear order. AI share of voice is built from answers, which change on almost every run and have no reliable order. That one difference changes how you sample, what you count and how much data you need before a number means anything.

Swipe sideways to see every column

	Search share of voice	AI share of voice
What is counted	Rankings or estimated clicks for a keyword set on Google and Bing	Brand mentions or citations inside sampled AI answers
Where the data comes from	Rank trackers, Google Search Console and Bing Webmaster Tools	Prompts you or a tool run, plus first-party counts: Search Console's Generative AI performance report (Google AI Overviews and Google AI Mode impressions) and Bing Webmaster Tools' AI Performance report (Copilot citations)
Stability	A ranking usually holds for days or weeks	The same prompt returns a different list on most runs
Position	Position 1 to 10 carries real meaning	Order inside an answer rarely repeats from run to run (SparkToro found closer to 1 in 1,000 for the same order), so position in a single answer tells you little
Demand weighting	Keyword search volume	Prompt volume is estimated by tools, not published by the engines

Appearance rates were far steadier than rankings in the same study. Across 994 answers to 142 differently worded prompts about headphones, Bose, Sony, Sennheiser and Apple each appeared 55% to 77% of the time. That is the property AI share of voice relies on: any single answer is a lottery, but the share of answers you appear in settles into a usable number once you sample enough of them.

What is the formula for AI share of voice?

The core formula is your brand's mentions divided by the total mentions of every brand you track, times 100. Two companion formulas are worth keeping beside it: mention rate, which tells you how often you appear at all, and a weighted version that counts high-intent prompts for more.

AI share of voice (%) = answers that mention your brand
                        / sum of answers that mention each tracked brand
                        x 100

Mention rate (%)      = answers that mention your brand
                        / total answers sampled
                        x 100

Weighted share (%)    = sum over prompts (prompt weight x your mentions)
                        / sum over prompts (prompt weight x all tracked mentions)
                        x 100

There is no industry standard, and the main tools disagree. Semrush's AI visibility metrics page defines share of voice as the percentage of mentions your brand receives compared with competitors. Ahrefs defines it as your share of impressions, where each mention is weighted by the Google search volume behind the prompt, and the cross-platform figure is a weighted average in which platforms with more impressions count for more. Same name, different numbers.

Swipe sideways to see every column

Method	What it counts	Best for	Watch out for
Mention share (Semrush's share of voice)	Your mentions as a percentage of all tracked brands' mentions	Comparing yourself with named competitors	Adding or removing a competitor moves your number
Mention rate (SparkToro's visibility percentage)	Percentage of sampled answers that name you	Early-stage brands and trend lines	Does not show who wins the answers you miss
Impression share (Ahrefs Brand Radar)	Your mentions weighted by the search demand behind each prompt	Categories with real search demand	Prompts with no keyword volume carry little or no weight

How do you calculate share of voice in AI search, step by step?

You calculate it by sampling a fixed prompt set on every engine you care about, logging each brand named in each answer, and dividing. The steps below work in a spreadsheet or inside any tracking tool that exports raw answers.

Calculating AI share of voice

Swipe sideways to see every column

Brand	Answers naming the brand	Share of voice	Mention rate
Brand X	168	16.7%	20.0%
Competitor A	420	41.7%	50.0%
Competitor B	294	29.2%	35.0%
Competitor C	126	12.5%	15.0%
Total	1,008 mentions in 840 answers	100% (rounded figures sum to 100.1%)	Not additive

Split the scorecard by engine before you read it. Profound's June 2025 analysis of 680 million citations found Wikipedia made up 7.8% of ChatGPT's citations, while Reddit made up 6.6% of Perplexity's. Engines that read different sources recommend different brands, so a 16.7% blended share can be 30% on one engine and close to zero on another.

How do you choose the prompts?

Choose prompts by buying intent, not by phrasing. SparkToro's volunteers wrote prompts with an average semantic similarity of 0.081, almost nothing in common, yet the engines still returned similar brands for the same underlying need. So cover every intent a buyer passes through, write each one the way a buyer would, and keep brand-name prompts out of the share of voice calculation, because they name you by definition.

• Category prompts, such as "best AI captioning tool for podcasts".
• Alternatives prompts, such as "alternatives to Competitor A for a small video team".
• Comparison prompts, such as "Brand X vs Competitor B for YouTube creators", tracked separately from the share figure.
• Problem prompts, such as "how do I add accurate captions to a 60 minute podcast video automatically".
• Constraint prompts, such as "cheapest captioning API with speaker labels for developers".

Good raw material sits in sales call notes, support tickets, the questions in Google Search Console and Bing Webmaster Tools query reports, and the People Also Ask boxes on your category terms. If you have never checked whether engines name you at all, start with the one-hour method in checking whether ChatGPT and Perplexity recommend your product without paid tools, then turn what you learn into the frozen set.

How do you choose the competitor set?

Choose the brands the engines actually name, not only the ones your sales team worries about. Run the prompt set once without a fixed list, record every brand that appears, and keep the ones that show up repeatedly plus an all-others bucket. Then freeze the list, because the denominator is part of the metric.

Here is why freezing matters. If Brand X adds a Competitor D that collects 200 mentions, its share of voice falls from 168 / 1,008 = 16.7% to 168 / 1,208 = 13.9%, with no change at all in how often engines name Brand X. Add new competitors at the start of a quarter, restate the baseline, and note the change in the report.

How do you benchmark AI share of voice over time?

Benchmark against a fixed baseline: same prompts, same engines, same competitors, same counting rules, measured on the same schedule every month. Read three months of trend before reacting to one, and report share of voice beside mention rate, citation counts and referral sessions. The measurement side of AI visibility work is mostly this discipline, applied every month without shortcuts.

Annotate anything that changes the measuring stick. Tools add engines and widen coverage, and engines change their interfaces: Similarweb reported in July 2026 that after ChatGPT made brand names clickable in May 2026, homepage visits went from roughly a quarter of its referrals (26% to 29%) to about 62% to 63%. A jump in your numbers the month a tool or engine changes is a measurement event until proven otherwise.

What is a good AI share of voice?

There is no universal good number, and we have not seen reliable cross-category benchmarks for AI startups. What a share can reach depends on how many brands the engines choose from. In Gumshoe's runs for the SparkToro study, top headphone brands reached 90% to 100% visibility, while the top brand design agencies for a coffee shop sat in the 30s and 40s, because that category has far more candidates.

Judge your number against two things instead: the gap to the category leader on the same prompts, and your own trend against the baseline. Until reliable category benchmarks exist, your own baseline is the comparison that counts.

When is share of voice the wrong metric?

Share of voice is the wrong headline metric in three situations. When you start at zero, it stays at 0% until you appear, so track mention rate on narrow prompts and citation counts first. When engines describe you wrongly, a mention still counts, so read the answers for accuracy and sentiment alongside the count. And in a two-brand niche, a share figure adds little over a simple mention rate.

It is also not a revenue metric. A rising share should show up later as more AI referral sessions and more signups from them, and the evidence on how that traffic converts is mixed, as we set out in whether traffic from ChatGPT and Perplexity converts better than Google traffic. If you want a tool to run the sampling for you, our comparison of AI visibility tracking tools shows how each one counts.