Exclude failed requests from the denominator of a brand mention rate, and report those failures separately as missing measurement coverage. A timeout is not evidence that a provider answered without your brand. This distinction matters when marketing teams compare weeks with different levels of service availability.
Classify the observation before counting it
Use four mutually exclusive states. A valid answer with a qualifying mention is present. A valid answer without one is absent. A transport or tool failure is failed. An empty or unusable answer is unresolved until a reviewer decides whether it meets your stated validity rule.
Write the validity rule before inspecting the brands. Otherwise, analysts can accidentally keep friendly answers and discard inconvenient ones. OpenAI's evaluation guidance identifies output variability as a reason to evaluate systematically. The classification and calculations below are our editorial recommendations, not an OpenAI-defined visibility metric.
A worked denominator
Illustrative example: a panel requests 20 observations. Sixteen return usable answers, four fail, and six usable answers mention the brand.
The observed mention rate is 6 / 16 = 37.5%. Valid-answer coverage is 16 / 20 = 80%. Reporting 6 / 20 = 30% as the mention rate silently treats every failure as a negative answer. Reporting only 37.5% hides the incomplete collection. Publish both figures with their counts.
If another week has 20 usable answers, compare the shared valid prompt/provider pairs as a separate view. A larger usable sample can change the mix of questions, even when the brand has not changed.
Decide how retries enter the record
- Keep the original failure and its collection time.
- Link the retry to the same planned observation.
- Declare whether the retry replaces the missing slot or becomes a separate repetition.
- Prevent a successful retry and its original failure from becoming two denominator entries.
- Mark comparisons provisional when missing coverage is concentrated in one provider or intent.
A useful report row contains planned observations, valid answers, failures, unresolved outputs, qualifying mentions and the resulting rate. Do not combine providers until each provider's coverage is visible.
When the answer remains unknown
High coverage does not prove the panel represents the market. Low coverage does not prove brand weakness. If an important provider is unavailable, investigate collection reliability before making a content recommendation.
Read the existing measurement framework for the broader baseline. Next, inspect one incomplete period in the AI visibility workflow and document its valid-answer denominator before interpreting the score.
Fact-checked: 17 October 2026. Examples are illustrative; no customer results are reported.