How to measure your brand's share of voice across AI engines

As buyers move their research into ChatGPT, Perplexity and Google's AI answers, marketers want the one number they have always wanted: share of voice. It carries over, but with a catch. AI answers are non-deterministic, so you cannot count them like impressions. What you can do is build an honest, sampled index of how often your brand appears across a fixed set of questions, measured against your competitors and tracked over time. Here is how to do it, and how to read the result without fooling yourself.

What AI share of voice actually is

AI share of voice is the percentage of a tracked prompt set where your brand appears in the answer, or, put the other way, your brand's mention count as a share of all brand mentions, measured against your competitors across answer engines. It is the AI-answer cousin of the classic marketing metric: instead of your slice of ad spend or search impressions, it is your slice of the brands the engines actually name when someone asks a question in your category. If you sell project management software and someone asks an assistant to recommend project management tools, your share of voice is how often you are one of the names that comes back.

The important word is sampled. You are not observing every real conversation people have with these engines, because you cannot. You are asking a fixed, representative set of questions and measuring who appears. That makes AI share of voice a directional index, a thermometer for your presence in AI answers, rather than a precise census of your standing in the market. Used with that understanding, it is genuinely useful. Read as literal market share, it will mislead you.

Share of voice versus share of answer

There is a distinction worth learning early, because it changes what you measure and what you conclude. Share of voice is about presence: does your brand appear at all in the answer? Share of answer is about ownership: are you the brand the engine recommends or cites as the answer, rather than one name in a list? The two are not the same, and conflating them is where a lot of AEO reporting goes wrong.

You can have high share of voice and low share of answer at the same time. Imagine your brand is mentioned in nine out of ten answers to a category question, which sounds like a win, but in each one it appears fourth on a list of five while a competitor is the one the engine actually recommends. Your presence is high and your ownership is low. Both numbers matter, and they matter differently: presence tells you the engine knows you exist, ownership tells you the engine trusts you enough to put you forward. Track both, and do not let a strong presence number hide a weak recommendation one.

"Share of voice asks whether the engine names you. Share of answer asks whether it picks you. You can win the first and lose the second."

Want to see how often AI engines name your brand right now? Get a free visibility audit and see where you stand across ChatGPT, Gemini and Perplexity.

The method: prompt-panel testing

There is no impression log to pull from these engines, so the only honest way to measure share of voice is to sample the answers yourself. The technique is prompt-panel testing, and the shape of it is straightforward even if doing it well takes discipline. You define a representative prompt set for your category, run those prompts across the engines on a schedule, capture the answers, parse each one for brand mentions, position and cited sources, and compute your share from what you find.

The part that trips people up is that the answers are non-deterministic. Ask the same engine the same question twice and you can get two different lists, in a different order, citing different sources. That is not a bug in your process, it is how these systems behave. It means a single run tells you very little. Share of voice only becomes a real number when you run the panel repeatedly, date-stamp every run, and look at the aggregate. One reading is noise; a series of readings is a signal.

• Define a representative prompt set. Write the real questions a buyer in your category would ask, in the language and phrasing they would use. Cover the range from broad ("best tools for X") to specific ("X that integrates with Y"). This set is the backbone of the whole measurement, so keep it fixed so runs stay comparable.
• Run across engines on a schedule. Put the same prompt set to each engine on a regular cadence, weekly or monthly, rather than whenever you happen to remember. A fixed schedule is what turns snapshots into a trend.
• Capture and parse the answers. Record each answer and pull out which brands are named, in what position, and which sources are cited. Position matters because being named first is not the same as being named last.
• Compute the share. Express your presence as the percentage of prompts where you appear, or your mentions as a share of all brand mentions, alongside the same figures for your competitors.

What you can and cannot measure

Being honest about the limits is what makes the number trustworthy. A prompt panel measures a defined slice of reality well, and it is silent about everything outside that slice. Knowing which is which keeps you from overclaiming.

What you can measure: your presence in a defined prompt set, your relative ranking against named competitors, which sources the engines cite when they build the answer, the trend in all of that over time, and a per-engine breakdown so you can see where you are strong and where you are absent. Those are real, actionable outputs.

What you cannot measure: your true market share across all the real prompts users actually type, the exact impression volume behind any answer, personalised or logged-in variation at any scale, and deterministic reproducibility, because the same question does not reliably return the same answer. If a tool or a report implies it has captured any of those with precision, be sceptical. Nobody can, because the underlying system does not expose it.

Which engines to cover, and how to weight them

A share-of-voice programme should span the engines your buyers actually use, not every engine that exists. The practical set to cover is ChatGPT at chatgpt.com, Google's AI Overviews and AI Mode, Perplexity, Gemini, Claude and Copilot. Perplexity is often the easiest to parse because it shows explicit citations, which makes capturing sources far less painful than on engines that bury or omit them.

Coverage is only half of it, though. A raw average across all six or seven engines flatters or punishes you for audiences you do not have. The fix is to weight the total by where your buyers actually are. For a B2B brand, that usually means leaning the weighting toward ChatGPT and Perplexity, where a lot of professional research now happens. For a consumer brand, Google's AI Overviews may deserve much more of the weight, because that is where general-audience queries surface. A weighted number that reflects your real audience is worth more than an unweighted one that treats a niche engine and your buyers' main engine as equal.

What to demand from any measurement

You can run a prompt panel by hand to start, and doing so once is the best way to understand the data before you automate anything. Beyond that, a category of tools has grown up to run the prompts and parse the answers for you, and this is the work Stellarcast does. Whether you build it, buy it or use us, the honest test of any share-of-voice measurement is the same, so judge every option against these five demands rather than a feature list.

• It shows a range, not a single number. Because answers are non-deterministic, a credible tool reports a band across repeated runs, not one falsely precise figure. If it gives you a clean number with no spread, it is hiding the noise, not removing it.
• It reads each engine separately. A blended score across engines you do not care about is misleading. You need the per-engine breakdown so you can weight it to where your buyers actually are.
• It is measured per locale. A German-language answer and an English one are different questions. Anything that rolls all markets into one figure is answering a question you did not ask.
• It names the sources, and whose they are. The point is not just whether you appear, but which pages the engine cited and whether they are yours, earned, or a rival's. Presence without the cited sources is half the picture.
• It date-stamps everything. Citation behaviour shifts in weeks. A number without a date is not a measurement, it is a screenshot.

Named tools in this category come and go and change their coverage and pricing monthly, so we will not hand you a shortlist to memorise. Hold whatever you use, ours included, to the five demands above, and you will not be fooled by a confident-looking dashboard that quietly fails three of them. If you want to see how we do it, our measurement methodology is written up in full.

Set realistic expectations for the numbers

Before you read your first result, calibrate what "good" looks like, because AI citation is still a low-base, volatile phenomenon. Similarweb's 2026 analysis found that ChatGPT citation presence rose from about 1.6 percent in June 2025 to about 6.8 percent in May 2026, and that the figure varies enormously by vertical, from around 23 percent in Travel to under 4 percent in Professional Services. In other words, the whole surface is still young, the base rates are low, and they swing depending on your category.

That matters for how you interpret your own share of voice. A modest-looking presence number may be entirely normal for your vertical, and month-to-month wobble is expected rather than alarming. It also matters for the bigger picture: Gartner has projected that traditional search volume could drop by around 25 percent by 2026 as queries shift toward AI, which is a projection rather than a measured fact, but it points to why building a share-of-voice habit now is worth the effort. The surface is small today and growing, so the brands that start measuring early learn to read it before it becomes the main event.

How to read your share of voice over time

Once the panel is running, the discipline is to treat the trend as the truth and any single reading as suggestive at best. Because the answers are non-deterministic, a good month and a bad month sitting next to each other tell you less than the slope across six of them. Judge progress by direction of travel: is your presence climbing across engines, are you appearing higher in the lists, are the sources the engines cite starting to include the ones you can influence?

• Aggregate before you conclude. Never draw a conclusion from one run. Average across repeated, date-stamped runs so the non-determinism averages out into a stable estimate.
• Watch the per-engine breakdown. A rise on ChatGPT and a fall on Google's AI answers net out to "no change" only if you ignore where your buyers are. Read the engines separately, then weight.
• Follow the cited sources. The sources an answer leans on are your fastest levers. If the same few sites keep shaping the answer, earning a clear, consistent mention there moves your share more than anything on your own site.
• Expect volatility, judge the slope. On a low base, single months bounce. The signal is the direction over a quarter or two, not the reading on any one day.

The takeaway

AI share of voice is a real, useful metric as long as you hold it honestly. It is a sampled, directional index built from prompt-panel testing: define a representative prompt set, run it across the engines your buyers use on a schedule, parse the answers for mentions, position and sources, and read the trend rather than any single run. Separate presence from ownership by tracking share of answer alongside it, weight your engines by where your audience actually is, and set expectations against a base that is still low and volatile. Do that, and you get an early, trustworthy read on how AI answers see your brand, while most of your competitors are still guessing.