AI Share-of-Voice.
We ask ChatGPT, Perplexity, Gemini and Claude the questions your buyers really ask, and count who gets named. More prompts widen the sample; more runs tighten the confidence band — and runs are free.
Share-of-voice is the closest thing there is to a ranking for AI assistants. We ask ChatGPT, Perplexity, Gemini and Claude the questions your buyers actually ask, and count how often they name you versus how often they name your rivals. It's 45% of your AI Visibility score — the largest single component, and the one people most often misread.
The four engines
We sample four answer engines — the surfaces people actually ask:
- ChatGPT · Gemini · Claude — these answer from what they know and what their tools fetch.
- Perplexity — web-grounded, and it returns real source citations. When you want to know which page got you cited, this is the engine that tells you.
How many of the four you sample depends on your plan, and you can narrow the set yourself in the setup popup — useful if you only care about, say, ChatGPT and Perplexity. You can't go above your plan's allowance. The per-plan matrix →
The prompt library
The prompts are built for your brand from the keywords in your market, ranked so the commercial and transactional ones come first. They fall into three groups, and the proportions are deliberate:
- Unbranded category questions — about two-thirds. “What are the best {category} providers?” · “Top companies for {category}” · “How to choose a {category} provider” · “Who offers the best {category}?” · pricing-and-options questions, with your real keywords in place of {category}. These are the ones that matter, because this is how a stranger asks. Nobody who hasn't heard of you types your name.
- Competitor comparisons — about a fifth. “You vs them — which is better?” and “Best alternatives to them”, across up to eight rivals. This is where you find out whether you're the alternative or the afterthought.
- Branded and defensive — the remainder, last. “Is [your brand] any good?” and “[your brand] reviews”. These check what an assistant says about you when your name is already on the table.
The keywords behind these questions come from your brand's market intelligence, and that leans hard on your services list. A thin services list produces generic, industry-level prompts — and a generic prompt measures a category you don't really compete in. Fill in your services first →
How we detect a mention
You count as appearing in an answer if either of these is true:
- Your brand name appears in the answer text. Matched as a whole word, so a short brand name can't be caught inside a longer one.
- Your domain appears in the citations. Matched on the registrable part of the domain, so a subdomain or a different country TLD still counts as you — and “box” never counts as “dropbox.com”.
That second rule is why Perplexity is so informative: being cited without being named still counts, and it's the earliest sign that your content is being read.
Prompts × engines × runs
A check is a matrix. Every prompt is asked to every engine, and every prompt-and-engine pair is asked more than once. The three dials do different jobs, and confusing them is the single most common mistake:
| Dial | What it changes |
|---|---|
| Prompts | How much of your market you cover. More prompts = a broader sample of the questions people ask. |
| Engines | How many assistants you're measured on. More engines = a broader sample of where you're asked about. |
| Runs | How confident the number is. More runs = a narrower ± band. Nothing else. |
Prompts and engines determine what a check draws on your balance. Runs don't. Ask the same engine the same question twice and you can get two different answers — that's how these systems work — so we ask repeatedly and average. Running the maximum your plan allows costs nothing extra and gives you a number you can actually trust.
The two numbers, and the ± band
Two different things get measured, and the headline blends them:
- Presence rate — the share of AI answers you appear in at all. “Mentioned in 34% of AI answers.” This is the stable one, and the one to watch month to month.
- Competitive share — your appearances as a proportion of all appearances, yours and your rivals' together. This is the one that tells you whether you're winning the category or just present in it.
The share-of-voice figure weights presence more heavily than share, because presence is the more reliable measurement.
Why there's a ± on it
AI answers vary from one run to the next. A single answer is an anecdote, not a measurement, so we sample and we publish the uncertainty alongside the number — you'll see something like “±9 pts, 30 prompts × 4 engines × 5 runs”.
One subtlety worth knowing, because it's why our band is wider than a naive calculation would give: repeated runs of the same question on the same engine tend to agree with each other, so they aren't independent evidence. We compute the confidence band over the number of distinct prompt-and-engine cells, not over the raw run count. Runs still tighten the band — a cell with five agreeing runs is a firmer data point than one with a single answer — but we won't let a big run count pretend to a precision the sample doesn't have.
The practical rule: if the band is wider than the change you're looking at, you haven't measured a change. Broaden the sample with more prompts before you conclude anything.
Not cited · behind · even · ahead
The result lands in one of four buckets, and the bucket is what to act on:
| Bucket | What it means | What to do |
|---|---|---|
| Not cited | You appeared in none of the answers we tested, but your rivals did. | Check the crawlers first, before you write a word. Step one → |
| Behind | You appear, but well below the leader. | Publish answer-first content on the specific prompts you're losing. |
| Even | You're roughly level with the top rival. | Push on the prompts where they're named and you aren't. |
| Ahead | You're the most-named name in your category. | Keep the content fresh. This position decays. |
Beneath the headline, the check also lists the individual questions where a rival was recommended and you weren't — with the engine that recommended them. Those are the most actionable rows in the whole product: each one is a specific question, a known asker, and a competitor already answering it. They open straight into BlogCraft.
What it costs, and what it needs
- The first check on each brand is free. After that, a check draws on your credit balance — the size of the draw depends on how many prompts and engines you choose (runs are free). See what things cost →
- One check per brand per day. Try to run a second and you'll be told to come back tomorrow. Share-of-voice doesn't move hour to hour, and re-running it wouldn't tell you anything new.
- At least one competitor is required. Without one there's nothing to take share from, and the component can't run.
- It's manual. Nothing schedules itself. If you want a monthly reading, you have to run it monthly.
A large check has a wall-clock limit. If an engine is slow and some answers don't come back in time, those answers simply don't count toward the sample — which shows up as a wider ± band. We'd rather widen the band than quietly invent the missing answers.