Answer Engine Optimization
AI Share of Voice: How to Measure It Honestly
AI share of voice tells you roughly whether AI answers put you in the category conversation. It is a useful compass and a misleading scoreboard. Here is exactly how the metric is built, why a handful of prompts cannot carry the weight tools put on it, and what to measure instead when a deal is on the line.
Key takeaways
- AI share of voice is the percentage of AI answers, across a chosen prompt set, in which your brand appears or is recommended versus the competitors named alongside you. It is a relative visibility metric, not a count of pipeline.
- It is calculated by running a prompt set (usually 10 to 50 questions) through one or more answer engines and dividing your brand's appearances by total brand appearances. The number is entirely a function of which prompts, engines, and window you picked.
- First limit: "mentioned is not selected." Most share-of-voice counts are mention-based, and a mention next to a competitor's recommendation sends no one anywhere.
- Second limit: it extrapolates a broad conclusion from a small sample. Answers vary by session and personalization, and cited sources turn over fast, so a single decimal reads far more precise than the data underneath it.
- Use share of voice as a compass, not a target. The work that changes outcomes is measuring specific buying questions, reading the cited evidence, publishing a source-backed fix, and re-measuring that same question.
Every AI-visibility tool sells a version of the same headline number. Some call it AI share of voice, some call it share of AI voice, some fold it into a single visibility grade. The pitch is intuitive: a competitor is named in AI answers more often than you, so here is a percentage that captures the gap. It is a genuinely useful compass. It is also, in the way most dashboards present it, a scoreboard that reads far more precise than the data it rests on. This guide defines the metric properly, shows exactly how it is calculated, and is honest about where it breaks, so you can use it without being misled by it.
1What is AI share of voice?
AI share of voice is the percentage of AI answers, across a defined set of prompts, in which your brand appears or is recommended, measured against the competitors that appear in the same set. It is a relative visibility metric. A 15% share of voice does not mean 15% of buyers chose you; it means that across the questions a tool sampled, your brand accounted for roughly one in seven of the brand appearances the answers produced.
The idea is borrowed from traditional marketing, where share of voice measured your slice of category advertising or search presence. Applied to answer engines, it tries to answer a reasonable question: when buyers ask AI about your category, how often are you in the room versus your rivals? That is a fair thing to want to know. The trouble starts with how confidently the number is reported, and with what it quietly leaves out.
2How is AI share of voice calculated?
The mechanics are simpler than the marketing implies. A tool assembles a prompt set, runs it through one or more answer engines, records which brands each response names, and divides. In plain terms:
Four inputs decide the result, and none of them is universal:
| Input | Why it moves the number |
|---|---|
| The prompt set | Usually 10 to 50 category questions. Broad head-term prompts favor big incumbents; specific use-case prompts surface niche players. Swap the set and the ranking of brands changes. |
| The engines | ChatGPT, Perplexity, and Gemini retrieve and cite differently, so a brand strong on one can be weak on another. A blended score averages away those differences. |
| What counts as an appearance | Mention, recommendation, or citation are three different bars. A mention-based score and a recommendation-based score for the same brand can diverge sharply. |
| The time window and run count | Non-deterministic answers mean a single run is one draw from a noisy distribution. Sampling once versus averaging many runs produces different numbers for the same week. |
This is the first thing to internalize: a share-of-voice figure is not a property of your brand. It is a property of the measurement setup. Two reputable tools can hand you two different percentages for the same company in the same week, and both can be correct for the prompts and engines they chose. That is not a bug to be fixed with a better tool; it is inherent to sampling a non-deterministic system.
3Why is being mentioned not the same as being chosen?
The first limit is what the metric silently blends. Most share-of-voice counts are mention-based, because mentions are the easiest thing to count at scale. But a mention is the weakest possible outcome, and the sharpest statement of why came from a practitioner in a generative-search community:
“Mentioned is not selected. Plenty of businesses get mentioned somewhere. Mentioned doesn't send a customer anywhere.”
A share-of-voice number treats every appearance as equivalent, so a response that lists your brand once, at the bottom, while recommending a competitor at the top, scores as a point in your favor. Aggregate enough of those and you can watch your share of voice climb while your actual recommendation rate falls. The number goes up; the pipeline does not.
For a B2B SaaS buyer, the outcome that moves a deal is being recommended, the brand the answer puts forward as the choice worth considering. Mention, recommendation, and citation are three distinct outcomes and should never collapse into one percentage. The distinction, and how to measure the three separately, is the spine of the companion guide to measuring AI search visibility. Share of voice is one number cut from that fuller picture; do not let it stand in for the whole thing.
4Why is an AI share-of-voice number a sampled proxy?
The second limit is statistical. An AI share-of-voice figure extrapolates a broad conclusion about your visibility from a small number of prompts, usually 10 to 50, run against a system that gives different answers on different runs. The practitioners building these tools say the quiet part plainly:
“As of now there are no consistencies in AI Search. The same prompt could give you completely different recommendations each time.”
Two independent findings show why a single decimal cannot be trusted as a stable rank. Search Engine Land's citation research found that 40 to 60% of the sources cited in AI answers change from one month to the next. The evidence base an answer is built from is churning underneath you, so a share-of-voice trend line partly tracks that churn rather than anything you did. And citations concentrate hard: roughly 30 domains capture about 67% of all AI citations, so whether your category's answers happen to lean on a domain you are present on can swing your number more than your own effort does.
Sources: Search Engine Land on citation turnover, Search Engine Land on citation concentration, and Semrush's answer engine optimization research.
The last figure is instructive. Semrush reported growing its own AI share of voice from 13% to 32% in a single month. Read one way, that is a success story. Read honestly, a metric that can nearly triple in a month is a metric with enormous variance, and variance that large means small samples cannot tell a real gain from noise. None of this makes the number useless. It makes it a proxy, an estimate of a tendency under stated conditions, that must be read as a direction of travel rather than a precise, reproducible rank.
5When is an AI share-of-voice number actually useful?
Used within its limits, share of voice earns its place. It is a fair coarse compass for three jobs: a first read on whether you are in the category conversation at all, a rough sense of which competitors keep recurring in your space, and a trend to watch over long horizons if, and only if, the prompt set, engines, and method are held constant every time you measure. Change the setup between measurements and the trend is meaningless.
Where it fails is as a target. The moment a team sets out to raise a share-of-voice percentage, it optimizes for the thing easiest to move, more mentions across more prompts, rather than for winning the questions that decide deals. The audience has already grown tired of tools that stop at this kind of number:
“A lot of tools blur those together. I'd pay for measurement plus source and citation analysis. I would not pay much for another generic tracker with a GEO label on it.”
6What should you measure instead of a share-of-voice score?
Measure the specific buying questions that decide deals, one at a time, and let share of voice be a background compass. A brand-wide percentage cannot tell you which question you are losing, to which competitor, or why, and those are the only facts you can act on. The method that changes outcomes is a closed loop run per question:
| Step | What you do |
|---|---|
| 1. Measure the question | Take a real buying question ("best [category] tool for [use case]," "[competitor] alternatives") and put it to the engine under labeled conditions. Record who is recommended, who is merely mentioned, and which sources are cited. |
| 2. Read the cited evidence | For a question a competitor wins, the sources the answer cited are your instruction list. They show exactly what is shaping the outcome, which is where a fix has to land. |
| 3. Publish a source-backed fix | Strengthen the surfaces that are actually cited, grounded only in the evidence you observed, never invented: a use-case page, a comparison, a corrected third-party fact. |
| 4. Re-measure the same question | Re-ask it on the same engine after the sources update. A changed answer on one question is worth more than any aggregate percentage, because it is attributable to what you did. |
This is measurement in service of a fix, not a scoreboard. The full method, and how answer engines decide who to name in the first place, is laid out in the pillar guide to getting recommended by AI, and choosing the right questions to track, rather than the most prompts, is covered in which prompts to track for AI visibility.
Track the questions that decide deals, not a vanity percentage
Frequently asked questions
What is AI share of voice?+
AI share of voice is the percentage of AI answers, across a chosen set of prompts, in which your brand appears or is recommended, measured against the competitors that appear alongside you. If a tool runs 30 category prompts through an answer engine and your brand shows up in 9 of the responses while all named brands together account for 60 appearances, your share of voice is roughly 15%. It is a relative visibility metric: a compass for whether you are in the conversation at all, not a count of customers won.
How is AI share of voice calculated?+
A tool defines a prompt set (usually 10 to 50 category questions), sends each one to one or more answer engines, and records which brands are named in the responses. Your share of voice is your brand's appearances divided by total brand appearances across the same sample, expressed as a percentage. Variants count mentions, recommendations, or citations, and some weight results by rank position or sentiment. The number depends entirely on which prompts, engines, and time window were chosen, so two tools measuring the same brand can report very different figures.
Is AI share of voice accurate?+
It is directionally useful and unreliable as a precise score. Answer engines are non-deterministic and personalized, so the same prompt can return different brands on different runs, and cited sources turn over fast. A share-of-voice figure extrapolates a broad conclusion from a small sample, so treat it as an estimate of a tendency under stated conditions, not a fixed rank. The trustworthy signal is whether a specific answer changed after you changed the sources behind it.
What is the difference between AI share of voice and being recommended?+
Share of voice usually counts mentions, and being mentioned is not the same as being recommended. A response can name your brand once in passing while putting a competitor forward as the choice the buyer should make. A rising share-of-voice number that hides a falling recommendation rate is a metric working against you. For a B2B SaaS buyer, the outcome that moves a deal is being the brand the answer recommends, not merely one of the brands it lists.
Should I optimize for AI share of voice?+
Optimize for the specific buying questions that decide deals, and let share of voice be a secondary compass. A single aggregate percentage tells you roughly whether you are in the category conversation, but it cannot tell you which question you are losing, to whom, or why. The work that changes outcomes is measuring the individual prompts where a competitor wins, reading the sources those answers cite, publishing a source-backed fix, and re-measuring that same question.