The essentials
- Principle: counting, across a fixed set of questions, how many times your brand and each of your competitors are cited by AI assistants.
- Protocol: 10 to 20 questions, 3 assistants, 3 repetitions per question, one reading per quarter; the variability of answers rules out any conclusion from a single test.
- How to read it: the gap with competitors counts more than the absolute value, and the source cited tells you which page to work on.
- 2026 benchmark: in a small-business market, a citation share of 20 to 40% puts a company among its sector's reference names.
Knowing whether AI assistants cite your company isn't enough: what counts is the place you occupy against your competitors on the questions that precede a purchase. Share of voice in AI measures that. Its general definition and its calculation by channel are in share of voice: definition ; this article describes the competitive protocol, the dashboard to keep and how to use the results. It complements the tracking of your own brand described in monitoring your brand's presence in generative AI.
Building the question set
The reading is only as good as the questions chosen. Three families to cover, ten to twenty questions in total, fixed once and for all.
| Family | What it measures | The full destination URL |
|---|---|---|
| Shortlisting | Do you appear in the list of solutions considered? | "Which marketing tracking solution for a small manufacturer?" |
| Comparison | Are you cited against a named competitor? | "What alternatives to this tool for a small team?" |
| Expertise | Are you cited as a source on your subjects? | "How do you measure the return on an advertising budget in B2B?" |
The questions are written the way a customer would ask them, with their constraints (company size, budget, sector, place), not as a search engine query. A question that is too generic produces answers citing the big players and teaches you nothing.
The measurement protocol
- Three assistants at minimum: ChatGPT, Gemini and Perplexity cover most usage; add Copilot if your customers are on Microsoft.
- Three repetitions per question: the same question produces different answers; without repetition, the reading measures chance.
- A neutral session: no conversation history and no account logged in, to avoid personalisation.
- Note four things: the brands cited, the order they appear in, the source used for each, and the justification given.
- One reading per quarter: the models change slowly and the variability makes monthly measurements hard to read.
- Automate if the volume justifies it: GEO tracking tools take these readings at scale; the manual approach remains sufficient for a small business, as tracking visibility on ChatGPT and Perplexity.
The competitive dashboard
| Column | Example | Use |
|---|---|---|
| Brand | Your company and 4 to 6 competitors | A fixed scope |
| Citations | The number of questions where the brand appears | The numerator of the share of voice |
| Share of voice | Citations / total questions tested | The main metric |
| Average position | Average rank in the answers where the brand appears | Being cited first or last is not worth the same |
| Dominant source | The page or site most often cited for the brand | Points to the content doing the work |
| Argument selected | The justification the assistant gave | Reveals what the models retain about the brand |
In a small-business market, a citation share of 20 to 40% places a company among its sector's reference names; above 50%, it is the perceived leader. These benchmarks vary a great deal with how concentrated the market is and with the presence of comparison sites and specialist media, often cited ahead of the companies themselves.
Reading the gaps and acting
- Absent everywhere: the problem lies upstream, in indexing or in blocked AI crawlers; see indexing by AI.
- Cited on expertise, absent from shortlisting: your content is working, but the company isn't identified as a supplier; work on the offer pages and the entity description.
- Cited last: you exist but the argument is missing; look at the justification given to those ahead of you and check your strengths are written down in plain terms.
- A competitor dominates with a single page: analyse that page — it answers a sub-question you don't cover.
- Comparison sites cited ahead of everyone: the battle is fought on those third-party sites as much as on yours; a complete listing and reviews count there. For a local business, see local GEO.
- A false argument: the assistant attributes an inaccurate characteristic to you; the source to correct is identifiable in the reading.
Connecting it to the rest of your measurement
Share of voice is read alongside two other metrics: referral traffic from assistants, isolable in GA4 by the AI domains, and the trend in brand searches in Search Console, which captures the indirect effect of recommendations. A share of voice that grows with no effect on those two series signals citations on questions too far from the purchase. The full set of metrics is presented in GEO KPIs.
How GreenRed helps
Rather than juggling several tools, GreenRed's GEO tracking brings these metrics together in a single dashboard, compares them over time and tells you which actions come first. You can try it free, with no card, from the Pricing.
Frequently asked questions
How do you measure your share of voice in AI?
By fixing ten to twenty shortlisting, comparison and expertise questions, putting each of them three times to three assistants in a neutral session, and counting for each brand the number of questions where it is cited, against the total. One reading per quarter is enough.
What citation share should you aim for?
In a small-business market, appearing in 20 to 40% of questions places the company among its sector's reference names; above 50% it is the perceived leader. These benchmarks vary with market concentration and with comparison sites being cited ahead of companies.
Why repeat the same question several times?
Because assistants don't produce the same answer twice: variability from one run to the next is significant. Without at least three repetitions per question per assistant, the reading measures chance rather than a real position.
What do you do if a competitor is cited and we aren't?
Look at the source cited for that competitor and the justification given. The source points to the page working for them, the justification points to the argument the model retained. Those two things identify precisely the content to produce or the information to make explicit.