TL;DR
A useful AI visibility report lets you trace a conclusion to dated evidence and decide what to do next. Require the questions, platforms, collection method, denominator and limitations. Keep factual accuracy, mentions, citations, visits and appointments separate.
Put the scope above the score
A report should first state which services and locations it covers. Then it should identify the questions tested, platform and mode, observation dates, and whether previous conversations influenced the session. Record collection failures separately from answers that did not mention the practice.
Without that context, a percentage can sound more precise than the evidence supports. To express a sample mention rate, divide completed observations that mention the practice by all completed observations in the defined sample. If no observations completed, report the rate as unavailable rather than dividing by an empty denominator. If a later report changes to easier branded questions, a higher rate does not establish improved unprompted discovery.
Ask for the original question set and any revisions. Keep a stable comparison group and report exploratory questions separately. The goal is a defensible comparison, not a larger number.
Use a report structure that supports action
| Section | Minimum useful content | Decision it supports |
|---|---|---|
| Scope and method | Services, locations, platforms, questions and collection dates | Is this relevant and comparable? |
| Observations | Saved responses, mentions, citations and factual errors | What actually happened? |
| Website evidence | Relevant search, page and booking-path findings | Is there a public-source problem to fix? |
| Completed work | Specific URLs and changes, with review status | What did the team deliver? |
| Next decisions | Priority, owner, evidence needed and dependency | What should happen next? |
Treat this as a reporting specification to agree with your partner. It does not require a particular dashboard or a proprietary scoring formula.
Read platform metrics within their limits
Bing’s AI Performance documentation describes citation activity across supported experiences. Its page-level citation counts do not establish ranking or placement, and grounding queries are a sample. Keep those labels attached to the data instead of renaming citations “patients reached.” Bing’s product announcement explains the measures.
A manually saved answer and a platform’s aggregate citation report are different evidence sources. They can complement each other, but adding them into one unexplained score obscures what each measures. Missing platform coverage should appear as a limitation, not a zero.
Make factual errors visible
A practice can appear frequently while being described incorrectly. Give wrong addresses, unavailable services and incorrect booking destinations their own issue list. Record whether the likely source is controlled by the practice, a third party or unknown.
For each priority, connect the observation to a proposed response. Use a fillable action record: “Observed answer cited [URL]; [fact owner] verified [correction]; update [controlled page] and recheck [saved question].” Complete it only with observed and verified facts. “Visibility is down; publish more” does not explain the cause.
Ask the provider to specify how reports will be delivered and whether software access is included. The sample format above is an editorial specification, not evidence of client outcomes.
Frequently asked questions
Should we reject every composite score?
No. A score can summarize a defined method. Ask for its components, weights, denominator and underlying evidence before using it to judge progress.
Can a report prove that AI caused new appointments?
Only if the claim has appropriate supporting evidence and a sound attribution method. Mentions and citations alone do not establish that connection.