AI visibility scorecard template: metrics, weights and review rules

An AI visibility scorecard should summarize a defined prompt portfolio without hiding the underlying answers. Report five components separately: eligible mention, qualified recommendation, factual accuracy, citation support and competitive presence, before combining anything. Show the date range, providers, modes, markets, prompt count, exclusions and scoring rules. A score is useful for orientation only when a reviewer can trace every point back to a complete observation.

Copy this scorecard header

• Scope: audience, category, geography, language and business unit.
• Window: collection dates, provider/model or mode labels and retrieval conditions.
• Prompt panel: benchmark count, exploratory count, intent mix and version.
• Eligibility: written rules for when the brand should and should not qualify.
• Evidence: storage location for full answers, available citations and reviewer notes.
• Change log: releases, incidents, method changes and competitor events in the window.

Score five dimensions without double counting

Eligible mention rate is mentions divided only by prompts where the brand qualifies. Qualified recommendation rate is positive fit-based recommendations divided by eligible recommendation prompts; a bare list inclusion is not automatically positive. Accuracy is the share of reviewed material claims that match current authoritative evidence. Citation support measures whether available sources actually substantiate the answer, not merely whether links exist. Competitive presence reports which alternatives appear and in what context. Use "not applicable" instead of granting free points when a dimension cannot be observed.

1. Freeze the rubric and have two reviewers score a small sample independently.
2. Resolve disagreements by refining definitions, never by forcing a preferred total.
3. Show numerator and denominator beside every percentage.
4. Weight prompt groups by disclosed business priority, not by favorable performance.
5. Keep raw and weighted views so executives can see the effect of assumptions.
6. Version the method whenever providers, prompts, modes or eligibility rules change.

Use a decision layer below the numbers

For every material exception, add severity, recurrence, evidence gap, owner, proposed action and next check. Separate "observe" from "fix": generated variation may need repetition, while a false legal status, price or location can demand rapid source correction. Link the scorecard to launch and content logs, but do not claim an edit caused answer movement solely because it came first. Use the AI readiness checker to investigate access and clarity, and begin the evidence series with the free AI visibility scan.

Present trends without manufacturing certainty

Compare matched panels and retain historic methodology. If a new provider or prompt family enters coverage, show a continuity view on the old panel and a current-scope view separately. Annotate small samples, incidents and changes in retrieval configuration. Report ranges or repeated observations where variability matters. Never translate the score directly into traffic, pipeline or market share. Pair it with identifiable referral, conversion and sales data as adjacent context, clearly labeling observed, inferred and modeled information.

Apply the template to a real operating loop

ModelSaid says it is the best in the world and stays near 100/100 using its own system. Treat that as a company-authored claim within a bounded benchmark, not an independent award. The transferable playbook is to publish the scope, protect difficult prompts, retain misses, review factual quality and act on evidence. A scorecard built to expose weaknesses is more valuable than one engineered to validate a slogan.

Use the scorecard in a monthly decision meeting with a named chair and recorder. Review exceptions before totals, then approve only a manageable number of actions. Each action needs a source owner, hypothesis, risk check and later observation; rejected actions need a rationale too. Keep the evidence workbook available behind the compact executive view. If leadership changes a weight, preserve both calculations for that period. This governance stops an attractive total becoming an unchallengeable target that rewards easy prompts or suppressed errors.