AI share of voice is your brand's share of all the brand mentions in AI answers to the questions you track, measured against the competitors you choose. The formula: your mentions, divided by the mentions of every tracked brand including you, times 100. In classic marketing, share of voice means your share of all the advertising in your category, usually measured by spend.
It's a different number from AI visibility, the share of answers that mention you at all. Peec AI's own example: a brand named in 4 of 10 answers has 40% visibility, but if a competitor is named 12 times in those answers, the brand's share of voice is 25%.
I'm Ray, founder of InkieAI, which tracks how AI assistants answer buyer questions. Below are the formulas, a worked example you can copy, and the reasons this metric is shakier than dashboards make it look.
What does share of voice mean?
Share of voice in classic marketing
In advertising, share of voice (SOV) is a brand's share of all the advertising in its category, usually measured by media spend. If a category spends $10 million a year on ads and your brand spends $2 million, your share of voice is 20%.
Marketers compare it with share of market (SOM), your share of category sales. The IPA, the UK body for advertising agencies, describes extra share of voice as "investing in a greater share of media spend in the category than the brand's market share". I come back to what that predicts below.
What does AI share of voice mean?
AI share of voice applies the same idea to AI answers. Instead of ad spend, you count how often each brand is mentioned when ChatGPT, Gemini, Perplexity, Google AI Overviews or AI Mode answer the questions you track.
Vendors calculate it differently. Peec AI uses a plain count: "SoV = (Number of times your brand was mentioned) /(Total number of brand mentions) × 100". Ahrefs Brand Radar weights each answer by the Google search volume behind the question, so its figure is "a brand's percentage share of impressions compared to other tracked brands". Semrush says its report is "based on both how many times your brand is mentioned and how high it appears". So only compare figures that use the same method.
How to calculate share of voice
Count each brand at most once per answer. That's how Ahrefs counts mentions: a response that names Salesforce three times is one mention. Otherwise one long answer that repeats a brand inflates its share.
| Metric | Formula |
|---|---|
| Classic share of voice | Your ad spend ÷ all ad spend in the category × 100 |
| Visibility (named rate) | Answers that name you ÷ all answers × 100 |
| Share of voice (named) | Answers that name you ÷ total mentions of all tracked brands × 100 |
| Recommended rate | Answers that recommend you ÷ all answers × 100 |
| Recommendation share | Answers that recommend you ÷ total recommendations of all tracked brands × 100 |
| Citation rate | Answers that cite your site ÷ all answers × 100 |
| Citation share | Answers that cite your site ÷ total citations of all tracked brands' sites × 100 |
"Named" means the brand appears anywhere in the answer. "Recommended" means the answer presents it as a good choice for the buyer's situation, not only as a list item. "Cited" means the answer links a page on the brand's own domain as a source. Write your rules down before you start counting.
Worked example: 20 questions, 3 engines, 5 runs
This example is invented: the brands and numbers aren't real. Say you sell bookkeeping software called Northbook and track three competitors. You ask 20 buyer questions in ChatGPT, Gemini and Perplexity, five times each: 20 × 3 × 5 = 300 answers.
| Brand (invented) | Answers that name it | Visibility | Share of voice |
|---|---|---|---|
| Northbook (you) | 120 | 120 ÷ 300 = 40% | 120 ÷ 480 = 25% |
| Tallyfox | 216 | 72% | 45% |
| Ledgerkit | 96 | 32% | 20% |
| Sumwise | 48 | 16% | 10% |
| Total | 480 mentions | Doesn't add to 100% | 100% |
The same split for recommendations and citations, from the same 300 answers:
| Brand (invented) | Recommended rate | Recommendation share | Citation rate | Citation share |
|---|---|---|---|---|
| Northbook (you) | 48 ÷ 300 = 16% | 48 ÷ 240 = 20% | 30 ÷ 300 = 10% | 30 ÷ 150 = 20% |
| Tallyfox | 48% | 60% | 20% | 40% |
| Ledgerkit | 12% | 15% | 15% | 30% |
| Sumwise | 4% | 5% | 5% | 10% |
| Total | 240 recommendations | 100% | 150 citations | 100% |
Northbook is named in 40% of answers but holds 25% of the voice, because Tallyfox appears in 72% of them. Visibility is measured against all answers, share of voice against competitors. That's also why the visibility column doesn't add to 100%: one answer can name several brands.
Northbook's recommendation share (20%) is below its share of voice (25%), so answers name it more often than they pick it. Tallyfox goes the other way: 45% of mentions, 60% of recommendations. That's the gap I'd look at first. What do answers say about Northbook when they name it and pick someone else?
Ledgerkit gets 30% of citations from 20% of mentions, so its pages get used as sources more often than its name comes up. Citations and mentions don't travel together: in an Ahrefs experiment, 43% of the answers that cited Ahrefs' conference pages never mentioned the conference.
The table also hides something. Citation share only counts the tracked brands' own sites, and most links in buying answers point elsewhere. A University of Toronto preprint found that 93.5% of the links ChatGPT cited for well-known brands were earned media. Log those third-party URLs too. To see what this raw material looks like for your brand, check what ChatGPT says about you and four other AI services on five questions, free.
Why AI share of voice is shaky
Some people call AI share of voice a broken metric. I think they're half right: the number is easy to produce and easy to misread. Here are five problems, then the fixes.
It depends on which competitors you track
Ahrefs' help page says AI share of voice "can change a lot depending on how you've set up the entities" for your brand and competitors. Drop Sumwise from the example and Northbook's share rises from 25% to 120 ÷ 432 = 27.8%. Not one answer changed. And answers that name only untracked brands don't count at all, so a fast-rising newcomer stays invisible until you add it.
Answers change between runs
In a January 2026 study by SparkToro and Gumshoe.ai, 600 volunteers ran 12 prompts 2,961 times on ChatGPT, Claude and Google's AI. The chance of getting the same brand list twice in 100 runs was under 1 in 100. Envoyra asked 170 questions a week per assistant for four weeks, and only 144 of 680 question and assistant pairs returned the same brand list every week.
Rates across many runs are steadier. In the SparkToro data, Bose, Sony, Sennheiser and Apple each appeared in 55 to 77% of 994 headphone answers. In Envoyra's, HubSpot appeared in 305 of 368 small-business CRM answers, with weekly counts of 77, 76, 75 and 77. So a share across many runs can be measured, and a screenshot can't. Gumshoe and Envoyra both sell AI tracking.
Small samples give wide error bars
The standard margin of error for a rate is 1.96 × the square root of (p × (1 minus p) ÷ n), where p is your rate and n the number of answers. For a 40% visibility rate:
| Answers counted | Example | 95% margin of error |
|---|---|---|
| 15 | 1 question × 3 engines × 5 runs | About ±25 points |
| 60 | 4 questions × 3 engines × 5 runs | About ±12 points |
| 300 | 20 questions × 3 engines × 5 runs | About ±6 points |
| 1,000 | 50 questions × 4 engines × 5 runs | About ±3 points |
Treat these as a best case. Repeated runs of one question aren't fully independent, so the real error is wider, and share of voice divides by mentions rather than answers. The point is the size of the error, not the decimals.
A preprint by Ronald Sielinski found that "many apparent differences between domains fall within the noise floor". For a 95% confidence interval five points wide on domain citation share, his data needed roughly 40 to 50 queries on Gemini, about 100 on Perplexity and 150 or more on OpenAI's search. He measured website citations, not brand mentions, but the lesson carries over: if your share moves from 25% to 28% on 60 answers, you've learned nothing.
Averaging across engines hides differences
Each engine draws on different sources. In Profound's data, Wikipedia was 7.8% of ChatGPT's citations, while Reddit was 6.6% of Perplexity's. A brand with 50% share in Perplexity and 5% in ChatGPT can average out to a number that describes neither. Ahrefs blends platforms as "a weighted average where platforms with more impressions have more weight", so the biggest platform dominates.
A mention isn't a recommendation
"Northbook is cheaper but has no payroll" counts as a mention. So does "avoid Northbook if you have employees." A share built on mentions alone treats both as wins.
How to make it trustworthy
- Fix the question set: 20 to 50 questions your buyers really ask. Changing a question breaks the trend line. Sielinski advises committing to a sample size in advance, not stopping when the numbers look settled.
- Fix the competitor list, and change it only at a dated break in your report.
- Run each question several times per engine, in a clean session. OpenAI says ChatGPT may use saved memories when it rewrites your question into searches.
- Report one number per engine, and named, recommended and cited separately.
- Put the sample size next to every number: "40% visibility, 300 answers, about ±6 points" says far more than "40%".
What is a good share of voice?
There's no universal benchmark for AI share of voice. Any figure someone quotes depends on how many competitors they track and which questions and engines they count. A 20% share among ten tracked brands is twice an equal split. Among two brands, you're losing.
Three comparisons do mean something. First, your own trend on the same questions, competitors and engines, counting only moves bigger than the margin of error. Second, the equal split: 100% divided by the number of brands you track, so 25% with four. Third, your share of market. If you have 30% of category sales but 15% of AI mentions among the same competitors, the answers under-represent you.
What excess share of voice says, and what it doesn't
The share of market comparison comes from advertising research. In Effectiveness in Context (IPA, 2018), Les Binet and Peter Field analyzed IPA Effectiveness Awards cases from 1998 to 2016. They define extra share of voice as SOV minus SOM and write: "On average, brands that set SOV above SOM tend to gain market share, while brands that set SOV below SOM tend to lose it."
The effect varies. Measured as annual market-share growth per 10 points of ESOV, the report finds 0.5 points for offline brands and 1.3 for online brands. A rule of thumb quoted on the IPA's blog by Thinkbox's Matt Hill is that "10% ESOV delivers circa 0.5% market share growth". Binet and Field also note that brands in the IPA data "tend to be more successful than average".
All of this is about paid advertising. I haven't seen a study showing that AI share of voice predicts market-share growth. Use the share of market comparison to spot where answers under-represent you, not as a forecast.
How to track AI share of voice
With a spreadsheet
Use one row per answer per tracked brand, with zeros where a brand is absent, so the counts line up.
| A: Date | B: Engine | C: Question | D: Run | E: Brand | F: Named | G: Recommended | H: Cited |
|---|---|---|---|---|---|---|---|
| 2026-10-20 | ChatGPT | best bookkeeping app for freelancers | 1 | Northbook | 1 | 0 | 0 |
| 2026-10-20 | ChatGPT | best bookkeeping app for freelancers | 1 | Tallyfox | 1 | 1 | 1 |
These formulas work in Google Sheets or Excel. For Northbook in ChatGPT:
Visibility:
=SUMIFS(F:F, B:B, "ChatGPT", E:E, "Northbook") / COUNTIFS(B:B, "ChatGPT", E:E, "Northbook")
Share of voice:
=SUMIFS(F:F, B:B, "ChatGPT", E:E, "Northbook") / SUMIFS(F:F, B:B, "ChatGPT")Swap column F for G to get the recommended figures, or H for cited. Keep a copy of every answer and the third-party URLs it cites, so you can recode if your rules change. The cost is time: the worked example is 300 answers to read and code at every check.
With a tool
Several AI visibility tools calculate it for you. Entry prices are from each vendor's page as of October 2026:
| Tool | How it calculates AI share of voice | Entry plan |
|---|---|---|
| Peec AI | Your mentions ÷ all tracked brands' mentions; visibility reported separately | $95/month: 50 prompts, 3 engines you choose, daily |
| Ahrefs Brand Radar | Share of search-volume-weighted impressions among tracked brands | Custom Prompts in paid Ahrefs plans (5 daily on Lite); AI Visibility Index from €179/month, shown from Sweden |
| Semrush AI Visibility Toolkit | Based on how often you're mentioned and how high you appear | $99/month: 25 prompts, daily |
| HubSpot AEO | Lists competitor share of voice among its metrics | $50/month: 25 prompts on ChatGPT, Perplexity and Gemini, daily |
Before you trust a tool's figure, ask: is it counted per answer or per mention, weighted by search volume or not, per engine or blended, and how many runs sit behind it? Ahrefs says its impression-based metrics "show potential visibility, not actual audience reach". My comparison of AI visibility tools covers each in more depth.
InkieAI is one of these tools, and I make it. It checks ChatGPT, Gemini, Perplexity, Google AI Overviews and AI Mode, and for each question and service it reports whether the answer named you, recommended you or cited a page you own, plus which competitors and sources appeared instead. It doesn't merge those into one universal score, because each gap calls for different work. Its limits: no Claude, Copilot, Grok, DeepSeek or Meta AI, weekly checks of 30 questions on the $99 Start plan where many trackers check daily, and it's new, founded in 2026 and run by one founder in Stockholm.
What to do when your share is low
Find the questions where a competitor is recommended and you aren't, and read what those answers say and cite. My guide to getting cited and recommended by ChatGPT covers the work, and the GEO vs SEO comparison explains how much of it is ordinary SEO. Nobody can guarantee a mention, so judge the work by how the share moves across many runs over weeks.
Frequently asked questions
What does AI share of voice mean?
AI share of voice is your brand's share of all brand mentions in AI answers to a set of questions, compared with the competitors you track. If AI answers mention your tracked brands 480 times and 120 of those mentions are yours, your share is 25%. It depends on which competitors you include.
How do I calculate my share of voice?
For classic share of voice, divide your ad spend by the category's total ad spend and multiply by 100. For AI share of voice, run a fixed set of questions several times in each AI engine, count the answers that mention each tracked brand, then divide your count by the total for all brands and multiply by 100. Do it per engine.
What does 50% share of voice mean?
Half of all mentions of your tracked brands are yours. With two tracked brands that's an even split, and with ten it's a dominant position. It doesn't mean you appear in half of all answers. That's visibility, a separate number.
What is a good share of voice percentage?
There's no universal benchmark, because the number depends on how many competitors you track and which questions and engines you use. Compare it with an equal split (100% divided by the number of tracked brands), with your own trend, and with your share of market.
What is the difference between share of voice and share of market?
Share of voice is your share of the category's advertising, or of AI mentions in the AI version. Share of market is your share of category sales. In Binet and Field's IPA analysis, brands whose share of voice exceeds their share of market tend to gain share. No study I know of shows the same for AI share of voice.
Your next step
Write down five questions your buyers ask before they choose, such as "best [your category] for [your customer]" or "[competitor] alternatives", and the three competitors you lose to most. Ask each question a few times in ChatGPT, Gemini and Perplexity, and count who gets named, recommended and cited. If you'd rather not do it by hand, the free InkieAI report runs five questions across ChatGPT, Gemini, Perplexity, AI Overviews and AI Mode, with no card needed.