AI share of voice is a comparative measure of brand presence within a defined sample of answers. In the protocol used here, count each tracked brand once per successful answer. Divide your brand's total appearances by the total brand-answer appearances across the fixed tracked competitor set, including your brand.
This definition is a proposed reporting convention, not an official metric issued by an AI platform. It describes sampled presence. It is not market share, audience reach, recommendation strength, or revenue attribution. Read the AI share of voice guide for the strategic context; this article concentrates on calculation and quality checks.
Define the unit before counting
The observation unit is one complete answer to one recorded prompt on one recorded surface at one time. If you repeat a question three times according to your protocol, those are three observations. A failed request is an attempt with no usable answer and belongs in a separate status field.
The counted unit is a brand-answer appearance. If Brand A occurs five times in one answer, it contributes one appearance. If that answer also mentions Brand B, it contributes another appearance for Brand B. A brand in a neutral comparison counts as a mention even if the answer does not recommend it.
Freeze the tracked set and aliases before collecting the baseline. For example, the set could be Brand A, Brand B, and Brand C. Untracked brands may be recorded in a separate discovery field, but adding them to the denominator halfway through a time series would change what the metric means. Restate history or start a newly labeled series when the tracked set changes.
Write the formula explicitly
Use this notation in your reporting worksheet:
brand appearances = successful answers mentioning that brand
all tracked appearances = sum of appearances for every fixed tracked brand
AI share of voice = brand appearances / all tracked appearances * 100
mention rate = brand appearances / successful sampled answers * 100
For this formula, “mentioning” means a distinct identity match under your documented rules. An ambiguous acronym is not automatically a match. Save the answer for review and record unknown classification separately. If unresolved cases could materially change the result, show the uncertainty and withhold a precise headline until reviewed.
When no tracked brands appear in any successful answer, all tracked appearances equals zero. Report share of voice as undefined or not applicable. Do not manufacture a zero percent distribution that appears to be a valid comparative result.
Calculate a hypothetical sample
The following fictional answers demonstrate the method. They are not observed platform results.
| Answer ID | Status | Distinct tracked brands mentioned | Notes |
|---|---|---|---|
| A1 | Successful | A, B | A appears three times; count it once |
| A2 | Successful | B | Neutral mention only |
| A3 | Successful | A, C | No own-domain link for A |
| A4 | Successful | None | Valid answer, zero tracked mentions |
| A5 | Successful | C | C is explicitly recommended |
| A6 | Failed | Unknown | No usable response captured |
There are five successful answers and one failed attempt. Brand A has two appearances, Brand B has two, and Brand C has two. The denominator is six brand-answer appearances. Each brand's share of voice is 2/6, or approximately 33.3%. Small rounding differences can mean displayed shares do not sum to exactly 100%.
Each brand's mention rate is 2/5, or 40%. This is not a contradiction: share of voice and mention rate use different denominators. A4 contributes to the successful-answer denominator for mention rate but adds no appearances to the share-of-voice denominator. A6 is reported as a failure and contributes to neither denominator.
The absence of an own-domain link in A3 does not remove Brand A's mention. Citation and recommendation metrics are separate. A2's neutral reference still counts toward presence; A5's recommendation additionally qualifies under a stated recommendation rule.
Build the calculation worksheet
Use one row per answer-brand pair or a wide table with one binary column per tracked brand. In either format, preserve enough evidence to recalculate the report.
| Required field | Purpose |
|---|---|
| Prompt ID and version | Prevent silent changes in the question sample |
| Surface, locale, language, time | Identify the conditions observed |
| Attempt status and answer reference | Separate failures from valid zeroes |
| Brand identity and alias rule | Support repeatable entity matching |
| Mention flag | One or zero per brand per answer |
| Recommendation classification | Record suitability separately from presence |
| Citation URL and type | Distinguish own domain from third-party brand reference |
| Reviewer and unresolved issue | Make judgment calls inspectable |
Do not infer consumer behavior from API outputs. Keep consumer interfaces and API collection as separate segments with their own labels. The LLM brand tracking guide discusses the collection boundary.
Compare periods without changing the question
Calculate both periods with the same competitor set, prompt mix, coding rules, and planned observation conditions. Report the number of successful answers and failures alongside each result. Show the topic and platform breakdown before combining segments, because a shift toward easier branded prompts can inflate apparent visibility.
A rise in Brand A's share can occur while its own appearances remain flat if competitors appear less frequently. Therefore, display absolute appearance counts and mention rate beside share of voice. The AI visibility benchmark report guide explains how to make those comparisons reviewable.
Keep source evidence separate from calculated presence
A linked source is not automatic proof of a brand claim. OpenAI advises readers to inspect citations because search results and citations may be incomplete or incorrect in its ChatGPT search guidance. Retain the source URL and verify its relationship to the statement before labeling the answer well-supported.
Your final report should say “Brand A held 33.3% of tracked brand-answer appearances in this hypothetical sample,” not “Brand A owns one-third of the market.” Use the calculation to choose a topic for investigation, then inspect the actual answers. To discuss applying the worksheet to your own tracked set, book a demo with the brands, questions, and surfaces you intend to compare.