AI share of voice: what it is and how to measure it

AI share of voice is the percentage of AI-generated answers in your category that mention or cite your brand, measured against the competitors named alongside you. Where classic share of voice counts ad impressions or search rankings, AI share of voice counts how often engines like ChatGPT, Perplexity and Gemini put you in the answer itself, and how prominently.

This guide defines the metric, shows how it differs from the share of voice marketers already know, walks through the measurement method, and works a full example on a hypothetical ten-prompt set.

What is AI share of voice?

Share of voice has always measured how much of a conversation belongs to you. AI share of voice narrows that to one specific conversation: the answers AI engines generate when someone asks a buying question in your category. Every time ChatGPT, Perplexity, Gemini, Claude, Grok or Google AI Overviews composes a response and names a handful of brands, there is a finite amount of attention on offer. Your share of voice is the slice of those mentions that goes to you.

It has two moving parts. The first is presence: does the answer name you at all? The second is relative prominence: of the brands it does name, how big is your slice versus each competitor? A brand can score reasonably on presence and still lose on share of voice if every answer that mentions it also mentions three louder rivals. That comparison against a named competitive set is what makes share of voice more useful than a raw count of your own mentions, it is a metric of competitive position, not just activity. It is also the sharpest single number inside the broader discipline of AI visibility, which is the parent metric this sits under.

AI share of voice vs classic share of voice

The word is the same; the surface is not. Classic share of voice, whether the advertising version (your spend or impressions as a fraction of the category) or the SEO version (your ranked positions and clicks for a keyword set), assumes a shared, visible shelf. On a search results page, ten links appear and you compete for a higher slot; in an ad auction, everyone who pays shows up somewhere. Being present is close to guaranteed; the game is position.

An AI answer inverts that. Instead of a ranked list where everyone appears, you get a single synthesized paragraph that names maybe one to five brands, and often names none in your category, or none you would want. The shelf is far shorter, so absence is the default state, not the exception. Two further differences matter for measurement:

• There is no fixed list to rank within. The model composes a fresh answer each time, so you cannot read your position off a stable SERP, you have to sample answers and count.
• Answers vary run to run. Ask the same question twice and the brands named can change. A single check is a snapshot with real noise in it, which is why cadence matters more than it does in classic SOV.

One more distinction is worth naming because it changes how you count: a mention (the model names your brand in its prose) is not the same as a citation (the model links or attributes a source, often your page). Perplexity and Google AI Overviews tend to show explicit citations; a model answering from memory may mention you with no link at all. Both count toward share of voice, but they mean different things, the AI visibility glossary pins down these terms if you want the precise definitions.

How to measure AI share of voice

The measurement reduces to three inputs and two numbers. The inputs are your prompt set, the engines you run it through, and the cadence you run it on. The numbers are your citation rate and your share of voice.

1. Build a representative prompt set

Write the questions your buyers actually ask, twenty to fifty is a workable range for most brands, spanning category questions (“best CRM for a small agency”), problem questions (“how do I stop churn”) and comparison questions (“X vs Y”). This set is your definition of “your category,” so keep it fixed once you settle it; changing the questions every month makes the trend meaningless. Add prompts deliberately, and note when you do.

2. Run it across the engines your buyers use

Presence on one engine tells you little about another, so run every prompt through each engine and score them separately before you aggregate. The set worth covering today is ChatGPT, Perplexity, Gemini, Claude, Grok and Google AI Overviews, the six Cituna checks daily. For each answer, record which brands are named and where each falls in the response.

3. Run it on a schedule

Because answers drift run to run, a one-off scan cannot tell you whether a change is a real move or just variance. Run the set on a regular cadence, daily if your category moves fast, at least weekly if it is slower, and read the trend, not any single day. The point of measurement is the direction over time and the specific gaps, not a one-off score.

4. Compute the two numbers, and weight by position

Keep citation rate and share of voicedistinct. Citation rate is the share of answers that name you, your baseline presence. Share of voice is your slice of all the brand mentions in those answers, your standing versus rivals. They use different denominators and can diverge sharply.

Then, if you can, weight by position. A brand named first, or cited at the top of the source list, earns more attention than one mentioned in passing at the end. Applying a simple weight (first mention worth more than the fourth) turns a flat mention count into a truer picture of prominence, and often reveals that a rival dominates a category far more than a raw count suggests.

A worked example: a ten-prompt scan

Imagine a hypothetical CRM, Tessera, with three main rivals: Northwind, Quorum and Bento. You build a ten-prompt set and run it through four engines, 10 prompts × 4 engines = 40 answers. (Four engines here just to keep the arithmetic readable; in practice you would run all six.) You read every answer and tally which brands it names and in what order. The numbers below are illustrative, not measured, and exist only to show the method.

Say Tessera is named in 12 of the 40 answers. That is a citation rate of 30% (12 ÷ 40), you are in the room roughly a third of the time. Encouraging on its own. But those 40 answers contain 120 brand mentions in total once you count every rival and the long tail of other tools, so Tessera’s 12 mentions are a simple share of voice of 10% (12 ÷ 120). Same brand, same scan, two very different-sounding numbers, because they answer different questions.

The position-weighted column is where the real story surfaces. Tessera and Bento have identical raw presence, both named in 12 answers, both at 10% simple share of voice. But Tessera is usually named last, while Bento lands mid-answer, so once you weight by position Tessera slips to 7% and Bento holds 10%. And Northwind, which looked like a 23% leader on raw mentions, is actually being named first most of the time, so its weighted share jumps to 41%. Raw counts said Northwind was ahead; position weighting shows it is dominating.

Benchmarks, honestly

The most common question about any new metric is “what is a good score?” For AI share of voice, the honest answer is that there is no universal benchmark, and you should be wary of anyone who hands you one. A “good” number depends entirely on your category (a two-brand niche and a fifty-brand market are not comparable), on which engines you measure, and on the exact prompt set you chose. An “average AI share of voice is 14%” statistic, quoted without your category and your prompts behind it, is noise dressed as a benchmark.

Two comparisons are legitimate, and they are the only two you need. The first is your own competitive set: are you gaining on the specific rivals the engines actually name in your answers? The second is your own trend: is the number moving in the right direction as you ship fixes? Both are grounded in your reality rather than a borrowed average.

Measuring by hand is a fine way to see reality once, but running a fixed prompt set across six engines day after day, recording every brand, every position, every run, is a job nobody keeps up manually. That is the work Cituna automates: it runs your prompt set on a schedule, computes your share of voice per engine and against the competitors it sees cited, and pairs the answer-side numbers with a native Google Search Console connection so you can tie the metric back to real clicks instead of trusting a floating score. For the wider field of tools that measure this, see our rundown of the best AI visibility tools.

Share of voice is only the scoreboard, though, the next question is how to move it. That is answer engine and generative engine optimization: make sure the AI crawlers can reach you, structure content answer-first so a clean passage is easy to lift, build your brand into a consistent entity, and earn mentions on the sources the engines already trust. Start with the AEO guide and the GEO guide, then re-measure to prove the lift. Measure, fix, prove, repeat, and watch the share of voice move.