AI share of voice is the percentage of brand mentions in AI answers that belong to you, measured against every competitor named in the same category prompts. You calculate it by running a fixed set of queries across ChatGPT, Perplexity, and other engines, counting how often each brand appears, and dividing your mentions by the total.
That one number tells you whether AI engines treat you as a default answer or skip you entirely. Below is what the metric means, how to compute it, the supporting metrics that make it useful, and how to set up tracking you can trust week over week.
What AI Share of Voice Means
Traditional share of voice measured your slice of paid impressions or organic rankings against rivals. AI share of voice (AI SOV) moves that idea into the answer box.
When someone asks ChatGPT "what's the best project management tool for agencies," the model names a handful of brands. AI SOV asks a simple question: across all the prompts your buyers actually type, what fraction of those named brands are you?
It matters because the answer box is often the only impression you get. Many buyers read the answer and never click through, so if the model does not mention you, they never see your site at all. Being absent from the answer is closer to being invisible than being ranked tenth on a results page.
AI SOV is a competitive metric by design. A brand mention rate of 30% sounds healthy until you learn the category leader sits at 70%. Share of voice puts your number in context by dividing the total pool of mentions among everyone who shows up.
If you want the foundational concepts behind getting mentioned at all, start with what answer engine optimization is.
Share of voice vs share of model
Share of voice is your slice of brand mentions across all engines; share of model is the same calculation inside one engine. Report both, because a healthy blended number can hide an engine where you barely exist.
Three terms that are not interchangeable
Brand visibility is absolute. It is the share of your prompt set where you appear at all, regardless of who else is named, and it answers "am I present."
Share of voice is relative. It is your mentions divided by all brand mentions in the category, and it answers "am I winning." Visibility can hold steady while share of voice falls, because a competitor started appearing more often and the denominator changed.
Share of model is per engine. It is share of voice measured inside one engine's answers rather than across a blended set, and it answers "where am I winning or losing."
A worked share of voice and share of model table
Take one 30-prompt set read on three engines, with your brand and two rivals. The counts are illustrative, to show the arithmetic.
| Engine | Your mentions | Rival A | Rival B | Total mentions | Your share of model |
|---|---|---|---|---|---|
| ChatGPT | 30 | 50 | 20 | 100 | 30% |
| Perplexity | 10 | 25 | 15 | 50 | 20% |
| Gemini | 40 | 40 | 20 | 100 | 40% |
| All engines | 80 | 115 | 55 | 250 | Share of voice 32% |
The blended 32% looks competitive, yet Perplexity sits at 20% and Gemini at 40%. The fix for Perplexity (fresh, structured pages) differs from the fix for ChatGPT, so the per-engine column is the one you act on.
The Formula
The core calculation is straightforward:
AI SOV = (your brand mentions / total brand mentions in category) × 100
Say you run 30 category prompts across ChatGPT and Perplexity. The models name brands 200 times in total across those answers.
Your brand appears 50 times. Your AI share of voice is 50 / 200 = 25%.
Two details change the result more than people expect:
- Define the category boundary carefully. "Total brand mentions" means the competitors you decide belong in your set. Include three rivals and your share looks strong. Include fifteen and it drops. Pick a competitor set that matches how buyers actually compare, and keep it stable so the number stays comparable over time.
- Decide what counts as a mention. A passing reference ("tools like X, Y, and Z") is weaker than a recommendation ("I'd suggest X for this"). Many teams track both: raw mention share and recommendation share. The gap between them is where a lot of the real signal lives.
Why the same formula gives you a different answer each run
The arithmetic is trivial. The input is not, because the same prompt does not return the same answer twice.
A University of St. Gallen study that re-ran the same prompts every day on ChatGPT, Gemini, Google AI Mode and Perplexity found the cited sources overlapped by a Jaccard similarity of only 0.34 to 0.42 from one day to the next (University of St. Gallen, April 2026). Two readings of an identical prompt agree on well under half their sources.
Compute share of voice from a single pass and you have measured one sample of a distribution, then reported it as a fact.
The fix is unglamorous. Run each prompt five to ten times per engine and average before you divide. If you intend to put the number in a board deck, the honest floor is higher still, because at small sample sizes a two-point move is indistinguishable from noise.
This is also why the trend matters more than the level. A single scan tells you very little. The same scan repeated the same way for eight weeks tells you almost everything.
The Core Metrics to Track
AI SOV is one number, and on its own it hides too much. These seven metrics give you the full picture. Track each one per platform and over time, never from a single answer, because results shift with every model update.
| Metric | What it measures | How to read it |
|---|---|---|
| Brand mention rate | Share of answers where your brand appears at least once | The baseline. Compare it with your own earlier scans, not a cross-industry average |
| Recommendation rate | Share of prompts where you are actively recommended, not just named | Often a stronger pipeline signal than raw mentions. 100 prompts with 35 mentions but 12 recommendations = 12% |
| Prompt coverage | Share of your defined prompt library where you show up at all | Tells you how broad your presence is across topics |
| Share of voice | Your mentions divided by total brand mentions in the category | The competitive view. Read it against your locked competitor set and your own trend |
| Model-specific visibility | The same metrics broken out by engine | Reveals where you win and where you are missing entirely |
| Visibility volatility | How much your numbers swing between scans | High volatility means your position is fragile and needs reinforcing |
| Sentiment | The tone the model uses when describing you | A brand can be mentioned often but framed poorly, flagging missing features or stale information |
Why sentiment needs a closer look
The sentiment metric deserves attention. An engine might cite you in half its answers and still describe you as "more expensive than alternatives" or "limited on integrations." Frequent and unfavorable is its own kind of problem, and you only catch it by reading the actual language, not just the mention count.
Why Your Score Differs Across Platforms
You will not get one AI share of voice. You will get a different one on every engine, and the spread is wide.
You might sit at 40% inside ChatGPT and 15% inside Perplexity for the same prompt set. That is normal, and it reflects how differently these systems build answers.
The engines retrieve from different pools of sources and weigh them differently, so the domains one engine cites are often not the ones another cites, and the same brand can be cited heavily on one platform and barely at all on the next.
These are not the same channel with cosmetic differences. They are separate surfaces with separate rules.
How each engine builds answers
A few reasons for the divergence:
- ChatGPT leans on parametric knowledge. It builds a lot of its answers from patterns baked into training rather than live web results, so playbooks that move the needle on citation-heavy engines often do little here. See how to get cited by ChatGPT for the specifics.
- Perplexity rewards freshness and structure. New editorial pieces with clean schema can appear in Perplexity citations within days. Our guide on how to rank in Perplexity covers what it favors.
- Google AI Overviews track classical ranking. AI Overview citations tend to follow where you already rank in regular Google search. Without the underlying organic work, optimizing the overview alone goes nowhere. See how to rank in Google AI Overviews.
The practical takeaway: a single blended score across all engines averages away the detail you need. Track each platform on its own, then decide where the opportunity is largest. The GEO metrics guide defines the supporting KPIs.
Benchmarks
Published industry averages for AI mention rates rarely compare cleanly with your own numbers. They come from different prompt sets, engines and competitor lists. Three reference points are more reliable:
- Engine-specific baselines. Engines differ sharply in how often they name brands at all, so the same brand can look dominant on one engine and absent on another. The St. Gallen researchers recommend setting a separate visibility baseline per engine rather than one threshold for all of them.
- Your locked competitor set. Share of voice only means something against the rivals you chose. Beating the two brands buyers actually shortlist you against matters more than any cross-category average.
- Your own history. The same prompts, engines and run counts each week give you a baseline nobody else's data can.
Treat these as orientation, not targets. A niche B2B category with three real players behaves nothing like consumer software with fifty. Your own trend line, measured the same way each week, is more useful than any industry average.
How to Set Up Tracking
You do not need a platform to start. You need a consistent method. The discipline is in keeping the inputs fixed so the outputs stay comparable.
- [ ] Define a fixed prompt set of 20 to 30 queries. Mix three clusters: brand prompts ("is [your brand] any good"), category prompts ("best [category] tool for [use case]"), and comparison prompts ("[you] vs [competitor]"). Write them the way buyers actually ask.
- [ ] Lock your competitor set. Decide which rivals count toward the "total mentions" denominator and keep that list stable. Changing it mid-stream breaks your trend line.
- [ ] Run every prompt across each engine. Cover ChatGPT, Perplexity, Google AI Overviews, and Claude at minimum. Run each prompt five to ten times per engine and average, since a single pass samples a distribution rather than measuring it.
- [ ] Record four things per answer. For each prompt and engine, log whether you were mentioned, your position relative to competitors, which sources were cited, and the sentiment of the description.
- [ ] Scan on a fixed cadence. Run a full audit weekly or at least monthly, and spot-check your top prompts more often. Citation distributions shift within weeks as models update and indexes refresh.
- [ ] Compute the metrics and watch the trend. Calculate mention rate, recommendation rate, prompt coverage, and share of voice per platform. The single most useful output is the direction of travel, not any one snapshot.
Doing this by hand across 30 prompts, five engines, and several runs each is a few hundred queries per scan. That is fine for a one-time baseline.
It gets heavy fast as a weekly habit, which is where tooling earns its place. See the best AEO and AI visibility tools for a comparison, and how to measure AI visibility for the full scoring method.
From measuring to moving the number
Measuring share of voice is only half the job. A dashboard that reports the number every week tells you where you stand, but it does not move you up the list.
The brands that grow share of voice fastest close the loop: they read what the data shows, which prompts they lose, which sources win the citation, and how each engine frames them, then update content to match, then re-scan to confirm the number moved.
This matters more with every model launch, because a new ChatGPT or Gemini release can reshuffle citation patterns within days, and content that won last month may not win this one.
