AI visibility earns a budget line only when it reports like one - and most of what gets reported today is noise: single screenshots, one blended score, unverifiable "share of AI" claims. Four metrics survive board scrutiny: recommendation rate on real buyer questions, description accuracy, citation presence in the sources AI engines use, and the before/after delta on identical questions. Each is measurable, repeatable and honest about its own confidence.
The four numbers worth reporting
- Recommendation rate.Of your priority buyer questions, in what share of repeated runs across providers are you absent, mentioned or recommended - reported per funnel stage, because a strong shortlist rate with weak validation answers is a different problem than the reverse.
- Description accuracy.When AI systems describe you, how often are the category, claims and facts right? Wrong descriptions convert worse than absence, and this number moves fastest after entity and markup fixes - an early proof point for the programme.
- Citation presence.Of the sources AI engines actually cite for your category questions, how many mention you accurately? This is the leading indicator for live-search visibility.
- Before/after delta.The same questions, same providers, remeasured after implementation. This is the only number that proves work happened - everything else is weather.
The numbers to refuse
Refuse one blended "AI score" with an opaque formula - it hides which system is failing and cannot survive a "what changed and why" question. Refuse single-run evidence - model answers vary between runs, so a screenshot proves availability of luck, not visibility. Refuse metrics without disclosed confidence - a serious report states how many runs completed, which provider tests failed and how much agreement supports each finding. If a vendor's reporting cannot answer these, the reporting is the product.
The $49 audit gives you the baseline for these numbers, with a confidence level on every finding →
Reporting cadence and ownership
Quarterly reporting fits most B2B boards: long enough for implemented changes to reach live answers, short enough to catch drift - with monthly measurement underneath for the team running the work, because answers shift as models and sources change. Ownership belongs with whoever owns pipeline, not with a side-of-desk volunteer: the metrics above translate directly into shortlist presence, which translates into deals - which is the sentence that gets the budget approved.
Frequently asked questions
Can AI visibility be tied to revenue?
Directionally yes: shortlist presence on high-intent questions correlates with inbound quality, and validation-stage accuracy protects late-stage deals. Attribution is honest at the correlation level - anyone promising per-deal attribution from AI answers is overclaiming.
What is a good recommendation rate?
It depends on category competitiveness and question stage - which is why the meaningful benchmark is your own baseline, remeasured. Cross-company league tables mostly compare measurement setups, not companies.
How many questions and runs make a credible sample?
Enough repeated runs across several providers to see stability - as a reference point, an EntityRise audit records up to 100 observations across ChatGPT, Claude, Gemini, Grok and Perplexity, with confidence labelled per finding.
Kamil Kępiński · Founder of EntityRise and Smophy Labs Inc.
Kamil Kępiński is the founder of EntityRise and Smophy Labs Inc. He writes about how models choose the companies they recommend and how to measure a company's visibility in AI answers.
About the authorXLinkedInThe methodologycontact@entityrise.ai
Boards fund what reports honestly.
EntityRise measures the baseline with confidence labels, not weather, so your own team or agency can fix what matters and repeat the audit to show the delta.