AI visibility metrics are useful when a reader can tell exactly what was observed, how it was classified, and which denominator produced the percentage. A chart labeled “visibility” can hide several unrelated events: a brand mention, a favorable recommendation, a link to a website, or an accurate description. Those events answer different business questions.
The definitions below form a proposed reporting protocol. They are not universal platform standards. Use them consistently within an AI share of voice program, document any changes, and keep the underlying answers available for review.
Start with collection health
Before interpreting brand performance, show the planned attempts, successful captured answers, failed requests, incomplete captures, and unresolved classifications. A valid answer with no tracked brands is different from a response you could not collect.
Define success for your study. A sensible rule is a complete, inspectable response to the scheduled prompt with sufficient context to apply the coding rules. A refusal may be a successfully captured response if the protocol says so; label the response type rather than silently removing it. Unknown brand identity or missing citation capture may affect particular metrics even when the answer text is available.
Keep product surfaces, locales, languages, dates, and prompt versions visible. Consumer interfaces and APIs should be separate observation groups. They are not interchangeable replicas. If the question sample changes, explain the break or calculate a comparable subset instead of presenting an unqualified trend.
Define mention rate and comparative share
Mention rate is successful sampled answers mentioning the brand divided by successful sampled answers. Count a brand once per answer, regardless of how many times its name appears. Document aliases and ambiguous identities.
Share of voice uses a different denominator: the total distinct brand-answer appearances for all brands in a fixed tracked competitor set, including your own brand. Divide your brand's appearances by that total. Several brands may appear in one answer, so the denominator may exceed the number of answers. An answer with no tracked brands adds no appearances. If the appearance total is zero, share of voice is undefined.
The calculation guide works through an example. Both metrics describe the selected sample. Neither establishes market share, the size of the exposed audience, or the likelihood that every user will see the same answer.
Specify the citation event
A citation rate needs a type and a denominator. For an own-domain citation rate, a practical definition is successful answers with at least one visible link to an approved brand-owned domain divided by successful sampled answers. Count each qualifying answer once, even if it contains several links.
A third-party brand-reference citation rate is different: successful answers with a linked third-party source that actually discusses the brand divided by successful sampled answers. Inspect the destination and relevance; a link to a publication's homepage does not necessarily substantiate the brand statement. One answer can qualify for both measures.
You may also report “own-domain citation rate among brand-mentioned answers,” but label that conditional denominator explicitly. Do not compare it directly with a rate using all successful answers.
OpenAI's API web-search documentation describes URL citation annotations for web-search responses. That is a documented API output structure, not proof that every consumer answer exposes identical data or every linked source supports the adjacent claim. Preserve collection-specific limitations.
Classify recommendations and accuracy explicitly
Define a recommendation as a concrete behavior in the answer, such as proposing a brand as suitable for the requested conditions. A neutral list, a warning against using the product, and a historical reference should not automatically count as favorable recommendations. Keep the evidence passage and a manual-review option for borderline cases.
Recommendation rate can use all successful answers, or an explicitly defined eligible subset. Name the choice. If your question set contains many factual prompts that do not ask for a choice, an eligible subset may be more informative, but it must be selected before reviewing outcomes.
Accuracy requires a checked statement and an authoritative reference. Distinguish correct, partly correct, incorrect, and unverified claims. Use a denominator of reviewed eligible claims or reviewed eligible answers, and say which. Do not mark an uncited statement incorrect solely because it lacks a link; missing support and factual error are different findings.
Use an inline metric dictionary
Fill the final column before starting a recurring report. The examples are proposed definitions, not observations.
| Metric | Proposed numerator | Proposed denominator | Your documented rule |
|---|---|---|---|
| Mention rate | Answers mentioning the brand | Successful sampled answers | Approved aliases and ambiguity handling |
| Share of voice | Brand's distinct answer appearances | All appearances across fixed tracked brands | Competitor set and version |
| Own-domain citation rate | Answers linking to an approved owned domain | Successful sampled answers | Domain list and link-capture limits |
| Third-party brand-reference rate | Answers citing a third-party page about the brand | Successful sampled answers | Destination relevance check |
| Recommendation rate | Answers proposing the brand as suitable | Successful answers or predeclared eligible subset | Suitability rule and reviewer |
| Accuracy rate | Reviewed answers meeting the accuracy rule | Reviewed eligible answers | Source, scope, and partial-credit policy |
| Capture success rate | Fully captured scheduled answers | Scheduled attempts | Failure and retry policy |
If an unresolved classification affects the numerator, report the unknown count and any resulting range. Avoid forcing uncertain results into “absent” just to produce a clean percentage.
Read a hypothetical dashboard correctly
Imagine a fictional sample of 20 successful answers plus two failed attempts. Eight answers mention Brand A, four link to its domain, and three explicitly recommend it under the study rule. Mention rate is 40%, own-domain citation rate is 20%, and recommendation rate is 15% if all 20 answers are the declared denominator. The failures remain visible and are not coded as brand absences.
If the fixed tracked set produces 25 brand-answer appearances and Brand A accounts for eight, its share of voice is 32%. That is different from its 40% mention rate. It says nothing about actual category revenue. These invented numbers illustrate reporting logic only.
Connect metrics to decisions
Use metrics to choose where to inspect evidence. A low own-domain citation rate with accurate mentions might lead to a source review. Frequent mentions with incorrect capability claims calls for factual correction. Rising share of voice alongside flat appearances suggests inspecting competitors and the sample before declaring progress.
Use the AI visibility benchmark report guide to present counts and comparable segments, and the content-update measurement guide before attributing a movement to an edit. For a practical discussion of which definitions suit your audience and collection method, book a demo with a few saved answers and the decisions your report needs to support.