An AI visibility benchmark report is a documented baseline for a defined set of buyer questions and collection conditions. Its purpose is to support decisions and make the next comparison fair. A report that cannot explain its prompts, brand set, failures, and counting rules is difficult to reuse, even if its charts look precise.

This guide belongs to the AI share of voice cluster. Use it after agreeing on your monitoring scope and before presenting an improvement target to a broader team.

Begin with the decision and scope

State who will act on the report and which decision they need to make. A product marketing team may want to understand how products are described for a particular use case. A content team may need to locate missing evidence. A communications team may need to correct repeated factual errors.

Define the prompt cohort by buyer task, language, and market. Record platforms and surfaces separately, including whether the sample comes from an app or API. Note conversation state and visible search conditions. Include collection dates and repetition rules. Keep the exact prompts in a versioned register using the prompt set guide.

Do not describe a hand-selected set as a representative sample of all questions users ask. It is a benchmark of those selected tasks. Repeat observations help show variability within that boundary, but do not eliminate selection bias.

Establish the coding rules

Fix tracked brands and aliases before analysis. Count each tracked brand once per successful answer, even when its name appears repeatedly. Retain neutral and negative mentions. Resolve ambiguous names using context and record uncertain cases for review.

Share of voice equals a brand's appearances divided by all tracked brand-answer appearances, including the brand's own appearances. Mention rate equals successful answers mentioning the brand divided by all successful sampled answers. An answer mentioning no tracked brands still belongs in the latter denominator. Record failed runs separately and leave share undefined if there are no tracked brand appearances.

Recommendations and linked citations need independent fields. A cited source may support a definition rather than a recommendation. OpenAI's search documentation says search answers may include citations and advises reviewing sources for accuracy. Your coding should retain the passage and URL rather than assuming the link's role. OpenAI search documentation

Use this report outline

1. Decision and headline observation
2. Scope: dates, markets, languages, platforms, and surfaces
3. Prompt set version, buyer-task groups, and selection method
4. Fixed brand set, aliases, and counting definitions
5. Collection: planned runs, successful answers, failures, retries
6. Results: counts, denominators, mention rates, SOV
7. Separate recommendation and linked-citation observations
8. Breakdown by platform and buyer-task group
9. Representative answers and verified source evidence
10. Changes, limitations, and alternative explanations
11. Actions: evidence, owner, deadline, and recheck condition
12. Appendix: prompt register and reviewer decisions

Put the sample boundary beside the headline, not only in the appendix. Someone copying a chart into a presentation should carry its dates, cohort, and denominator with it.

Show counts before interpretations

Start with successful answers and failed runs by platform. Then show each brand's answer appearances, mention rate, and share of voice. Include the number of answers containing no tracked brands. This helps readers see whether competitive share changed because the brand appeared more or because competitors appeared less.

In a hypothetical baseline, a brand could receive a higher share of tracked appearances while appearing in fewer answers than before. That result is mathematically possible and should not be presented as uniformly better visibility. The metrics guide explains why one metric cannot replace the others.

Show platform and task-group results before an overall total. When combining groups, disclose whether the total follows raw observations or a predefined weighting rule. If a platform has missing runs, do not silently allow the remaining platform to gain more weight in a period comparison.

Include evidence a reviewer can inspect

Choose examples that illustrate the main finding, including counterexamples. Link each excerpt to its saved answer ID and the prompt register. For citation claims, open the destination and verify the supporting passage using the citation audit.

Use the examples to identify a plausible action. A missing capability explanation, an outdated comparison, and an irrelevant prompt require different responses. The competitor gap workflow and competitor citation analysis turn those observations into evidence-led tasks.

Make the follow-up comparable

Freeze the original series when prompts, brand set, or collection conditions change materially. Start a new series or clearly label a recalculated history. If old answers were not saved, do not reconstruct a comparable past score from memory.

After an intervention, report what changed in the matched cohort and list other changes during the window. Timing alone cannot prove that a content update or PR placement caused a recommendation. Use measuring AI content updates for intervention reviews and keep the same caution in executive summaries.

End with a small owned backlog: one task, its supporting evidence, an owner, a deadline, and the next comparable check. The benchmark has done its job when the team can make a specific decision and later assess it honestly. Book a BIU demo to discuss how this report could fit your wider visibility review.