What AI Share of Voice Measures
AI share of voice is the share of AI answers that name or cite your brand, next to the same number for each competitor. You pick a fixed set of questions your buyers ask, ask them on each AI engine, and count which brands each answer includes.
A brand can appear in an answer in two ways, and a tracker should record both:
- Mentioned: the brand is named in the answer text.
- Cited: one of the brand's pages is listed among the sources the answer links to.
An answer can do one without the other. Engines that answer from their training data, without searching the web, can name a brand but have no sources to cite.
Two Formulas, and Why They Give Different Numbers
Tools and articles use two formulas under the same name.
1. Share of answers. The percentage of answers that include the brand.
> AI share of voice = answers that mention or cite the brand ÷ answers checked × 100
2. Share of mentions. The brand's mentions as a percentage of all mentions of the brands you track.
> AI share of voice = mentions of the brand ÷ mentions of all tracked brands × 100
Take 20 answers. Your brand appears in 6, Rival A in 12, Rival B in 4.
| Brand | Answers it appears in | Share of answers | Share of mentions |
|---|---|---|---|
| Your brand | 6 | 30% | 27% |
| Rival A | 12 | 60% | 55% |
| Rival B | 4 | 20% | 18% |
Share of mentions adds up to 100%; share of answers does not, because one answer can name several brands. Now add Rival C to your tracking list, which appears in 10 of the same 20 answers. Your share of mentions falls from 27% to 19%, although no answer changed. Your share of answers stays at 30%.
That is why share of answers is the more stable choice: it only moves when the answers move. Share of mentions is useful when you want one pie chart for a fixed list of brands, as long as the list never changes between the periods you compare.
How to Calculate It, Step by Step
- Write the question set. Use the questions buyers ask before they know which brand they want: "best payroll software for a 20-person company", not "is [your brand] good". Keep the set fixed between periods.
- Pick the engines. Calculate the share per engine as well as overall. ChatGPT, Perplexity, Gemini and Google's AI surfaces often name different brands for the same question.
- Ask each question on each engine and record, for every answer, whether your brand is named, whether your domain is among the sources, and which competitors appear.
- Drop failed checks. A request that timed out or returned an error is missing data, not an answer that left you out. Counting it as "not mentioned" pulls your share down.
- Count one answer per question per engine per period. If you re-ran a question during the period, use the latest answer, or a question you happened to re-run counts twice.
- Count every brand the same way. If your domain in the sources counts as an appearance for you, a competitor's domain in the sources counts for them. In each answer, a brand counts once, whether it was named, cited, or both.
- Compare periods over the same questions. If you added 20 questions this month, compare last month with this month on the questions both periods share.
Mistakes That Inflate or Distort the Number
Branded questions in the set. A question that names your brand returns an answer that names your brand. Mixing them in raises your share without telling you anything about being found. Questions that name a competitor have the same effect for that competitor. Keep both in separate panels.
Counting yourself two ways and competitors one way. Many trackers find your brand by name and by domain, but find competitors only by name in the text. Your share then rises above competitors who are cited as sources more often than you. Count both sides with the same rules.
Failed checks counted as misses, or re-runs counted twice. The first deflates your share; the second inflates whatever brand the re-run question favors.
Reading noise as change. The same question asked twice can produce a different answer and cite different sources. With 20 answers, a 30% share has a 95% range of roughly 15% to 52%. With 200 answers it is roughly 24% to 37%. A move of a few points between two weeks is usually within that range.
One number across engines. Averaging all engines together hides an engine where you are absent. Report each engine next to the total.
What Share of Voice Does Not Tell You
Whether a change you made worked. Share of voice moves for many reasons: engine updates, a competitor's new page, a new review on a site the engines cite. To tell whether your change caused a move, compare the prompts it affects against prompts it did not touch, over repeated runs. Prominara's methodology describes how it does this.
What the answers say about you. An answer can name your brand with outdated pricing or a wrong feature list. Accuracy and sentiment are separate checks.
Whether anyone clicked. AI referral traffic is a separate measure, read from your analytics.
How Prominara Calculates AI Share of Voice
Prominara uses share of answers. For each period:
- It takes the latest successful answer for each prompt on each engine. Failed checks are left out.
- It counts only prompts that name neither your brand nor a tracked competitor. Branded and competitor prompts are tracked in their own panels.
- Your brand counts when it is named in the answer text or your domain appears in the answer's sources. Each competitor is counted the same way, once per answer.
- The share is reported overall and per engine, with the change against the previous period of the same length, in percentage points.
- Sentiment is calculated separately, over every answer that names or cites you, including answers to branded prompts.
See it on your own prompts with AI citation tracking. If you are still sorting out the terms, AEO vs GEO covers what each one means.