The formula, and the four things it depends on
One line:
AI share of voice = your brand's mentions divided by all brand mentions, across a fixed set of prompts, engines, runs and dates.
Those four inputs have to be written down every time you quote the figure:
- Prompt set, and its version.
- Engine set.
- Number of runs.
- Date range.
Prompt coverage is not market share. A share of voice figure says how often you appear in a defined set of answers, and nothing about your share of the market.
Adviora's own panel is an example of a fixed set: 22 prompts, version 1, created 15 October 2026, with the rule "keep these prompts the same between runs".
The prompts matter as much as the formula. Adviora generates them by crawling the site for intent, ideal customer profile and messaging, then using search data: what Google shows and the People Also Ask questions, turned into prompts a buyer would type. Share of voice is only as meaningful as that set.
Three denominators, and when each is right
Buyers distrust the figure because most tools do not say what it divides by. One buyer put it as "smoke and mirrors". Another said, "If I go and do a search and I do another search the next day, it's different." Both are complaints about a missing denominator.
- Answer-level: answers that name you divided by answers checked. Right for "are we in the conversation at all". This is how Adviora's methodology defines the AI Visibility Score: "the share of relevant AI answers that name or cite you".
- Mention-level: your mentions divided by all brand mentions. Right for comparing against named competitors. It shifts when an engine lists 5 brands instead of 15.
- Citation-level: your cited pages divided by all cited sources. Right for "do we get a link back", which one buyer described as "a pathway back to our website".
Three things inflate the figure: branded prompts, duplicate prompts ("40% of the prompts were overlapping", said one buyer of a tool they had tried) and counting a single run.
Worked example: Adviora's own panel, 15 October 2026
15 October 2026 baseline, one run, API answers are directional samples.
| Measure | Denominator | Adviora |
|---|---|---|
| Answer-level | 15 prompt-by-engine cells checked (7 unbranded category prompts: P01, P02, P03, P05, P08, P09, P10) | Named in 0 of 15 answers (0%) |
| Citation-level | About 128 citation links observed | adviora.ai cited 0 times (0%) |
| Mention-level | All brand mentions in the 15 cells | Not yet computed in this run |
The cells cover Google AI Overviews (stored answers) and ChatGPT, Perplexity, Gemini and Claude (API samples). The brands named instead included Profound, Otterly, Peec, Semrush, SE Ranking, Rankscale and AthenaHQ.
The denominator lesson
A different source gives a different order. Ahrefs' all-platform AI citation counts for the same seven domains, dated 15 October 2026, show Profound 50%, Rankscale 16%, Otterly 15%, AthenaHQ 14%, Peec 5%, Scrunch 0.2% and Adviora 0%. Adviora is at 0% either way, but the competitor order differs from what our panel's answers listed. The reason is the denominator: Ahrefs divides by its whole index, and we divide by our own prompt set.
Why branded prompts stay out
The branded prompt P20, "What is Adviora?", came back 4 of 4. Fold that in and the share jumps for no real reason. Keep branded prompts out of the figure.
Run counts and variance: the same prompt names different brands
Ask the same prompt in the same engine ten times and the list of brands will not be identical. A brand can be named in six runs and absent from four. That is variance, and it is why one run is an observation and not a share. So the number of runs belongs next to every share of voice figure, the same way a sample size belongs next to a survey result.
- Report the run count. "Named in 6 of 10 runs of this prompt in this engine" says something a single percentage does not.
- Report it per engine. Engines move differently, and an average hides a brand that is strong in one and absent in another.
- Report it per market. A US result and an India result are two numbers, not one.
- Never blend them into one number. A single blended share of voice across engines, markets and answer types cannot be audited, and a reader cannot tell what moved it.
- Show the zeros. A prompt where nobody names you is a result.
Adviora's 15 October 2026 baseline above was one run per cell, which is why we call it a sample. Where a repeat run is available, a share that holds across runs is stronger evidence than one that appears in a single run.
How often to re-run, and how to read a change
Same panel version, same engines, re-run monthly. Adviora spot-checks at Day 30 and does a full re-run at Day 45. Repeat observations matter because cited sources and brands change from run to run ("Don't Measure Once", arXiv 2604.07585). A one-off appearance is an observation, not a ranking.
Publish your confidence rule, too. On the methodology page, High confidence needs at least 100 observations, 10 prompts and 3 runs, and Low is 29 or fewer observations. A published threshold is what turns a number into something you can argue with.
What a change in share of voice does and does not mean
A change in share of voice in AI answers means you were named in more of the same answers. It does not mean more traffic, more revenue, or that a particular fix caused it, because engines change models without notice. For the referral side, see track AI referrals in GA4.
How Adviora reports it
Share of voice is live in Adviora, per product and per market, alongside the competitor and citation views. The free audit checks ChatGPT, Gemini, DeepSeek and Perplexity; a free trial or paid plan covers six AI engines (ChatGPT, Claude, Gemini, Perplexity, DeepSeek and Mistral Vibe) plus Google AI Overviews. Read how Adviora works out each score and see the share of voice report. For how share of voice sits beside mention, citation, recommendation and click, read how to measure AI visibility. Next, find out your own starting point with the free AI visibility checker.