On this page
“Does HarborBookings support group reservations?” and “Which booking tools suit small tour operators?” ask different questions about a fictional product.
The first explicitly supplies the brand. It can reveal whether an answer describes the product correctly. The second leaves the choice of products open. It can reveal whether the brand appears during a particular discovery task.
Combining those answers into one visibility percentage can conceal both a factual problem and a discovery gap. Give each question group its own evidence, denominator and next action.
Decide which question each report answers
A branded accuracy report asks: when the product is explicitly named, does the recorded answer describe the relevant facts correctly?
A non-branded discovery report asks: when a relevant customer need is described without naming the target brand, does the recorded answer mention or recommend it, and in what context?
Neither answers how many customers saw the product, how much traffic the answer generated or whether a purchase followed. Those require separate evidence.
| Reporting group | Example question for the fictional product | Primary review |
|---|---|---|
| Branded accuracy | “Does HarborBookings support group reservations?” | Compare the stated capability with dated product facts |
| Non-branded discovery | “Which booking tools suit small tour operators?” | Inspect target appearances, recommendations and cited evidence |
| Named comparison | “How does HarborBookings compare with TourTool?” | Review comparative claims and fit; keep separate from open discovery |
Classify questions before collecting or summarizing answers. A comparison that supplies the target name is not an independent discovery of that brand, even if the answer recommends it.
Define factual accuracy before scoring answers
Write down the fact the question is meant to test and the authoritative evidence for it. Include the relevant version, date and qualifications. A capability available only for a particular workflow is not equivalent to an unrestricted capability.
Your internal fact record might contain the public product name, the verified statement, its supporting documentation, any scope limitation and a review date. It can remain an internal working reference. The report needs enough supporting evidence for a reviewer to understand the judgment; it does not require publishing the whole ledger.
Use a simple manual rubric at the answer level:
| Review label | Meaning |
|---|---|
| Supported | The answer addresses the question and its material factual claims agree with the dated reference evidence |
| Error present | At least one material factual claim contradicts the dated reference evidence |
| Incomplete | No identified material contradiction, but the answer omits information needed to answer the question |
| Unresolved | Available evidence is insufficient to judge a material claim reliably |
Apply one label per answer under a declared precedence rule: a confirmed material error takes precedence; without one, an unresolved material claim takes precedence over incompleteness. Keep claim-level notes so a mixed answer is understandable.
This is a proposed editorial review method. It is not a claim that an AI visibility dashboard automatically computes these labels or an accuracy score.
Show the two denominators side by side
Illustrative example only. All counts and product names in this article are fictional.
A collection contains ten successful branded answers and twenty successful non-branded answers. All thirty planned responses complete.
All ten branded answers name the product, which the questions already supplied. Manual review finds six supported answers, two with a material error, one incomplete answer and one unresolved answer.
Among the twenty discovery answers, four contain the target brand. Two of those four recommend it. The other two mention it without recommending it.
| Group | Result to report | Interpretation |
|---|---|---|
| Branded accuracy | 6 supported, 2 error present, 1 incomplete, 1 unresolved out of 10 reviewed answers | Two confirmed errors require investigation; uncertainty remains visible |
| Non-branded discovery | 4 target appearances out of 20 successful answers: 20% | The target appeared in this defined discovery sample |
| Non-branded recommendations | 2 recommendations out of the same 20 successful answers: 10% | Recommendations are a subset of the four target appearances |
A blended appearance count would be 14/30, about 46.7%. That figure hides that ten appearances came from questions that named the product. It also hides the two factual errors: an inaccurate answer can still contain the brand name.
If a percentage is useful for the accuracy review, “6 of 10 supported under our rubric” is clearer than an unexplained “60% accuracy.” Do not drop the incomplete and unresolved cases to make the result look stronger. Show failed or unavailable requests separately from these successfully collected answers.
Keep discovery evidence specific
Define whether the discovery measure counts mentions, recommendations or explicit citations. Count a target appearance once per successful answer, even if the name occurs repeatedly. Verify ambiguous aliases before assigning an appearance.
A recommendation is not an extra appearance to add on top of the same answer's mention. Similarly, a cited product page and a brand recommendation are different observations. See the guide to mentions, citations and recommendations when defining the report.
Retain the exact questions, provider context, market, language and collection dates. If the question set or provider coverage changes, explain the break before comparing rates. Twenty carefully chosen questions still represent a selected monitoring scope, not the full population of customer conversations.
Give each finding an appropriate owner
For a branded factual error, preserve the answer and the dated fact it contradicts. Then inspect the public explanation and any relevant cited source. A wrong statement on an owned page creates a different task from an inaccurate answer whose cited page already states the correct fact.
Use the existing brand-information correction workflow for the repair process. Keep the accuracy report focused on the finding, its evidence, the owner and the condition for closing it.
For a discovery gap, first confirm that the product actually fits the question. If it does, review how well the relevant pages explain that use case and whether the answer's source pattern suggests a specific issue to investigate. An absent brand name alone does not establish that another article is needed.
A product lead may resolve a capability question. A content owner may clarify a public explanation. A developer may repair access to the relevant page. Assign the work based on the evidence, rather than routing every finding to content production.
Track corrections without rewriting history
Product truth can change. An answer collected before a capability launched may have been accurate then, even if the same statement would be wrong today. Preserve the reference date and distinguish an answer error from a later product change.
When a page is corrected, record the publication and verification dates. Later, record whether comparable answers changed. “Page corrected” and “later answer supported” are separate milestones; the first does not guarantee the second.
Keep exploratory questions outside the established comparison until you deliberately add a new reference version. Adding several easy branded questions must not make the discovery report appear to improve.
Put both reports in one meeting, with separate conclusions
The two groups can share a reporting document and review meeting. They need separate conclusions because they support different decisions.
In Prerender Buddy, use the recorded question-level answers as evidence where AI Visibility is available. Prepare separate summaries and a manual factual review when the available views do not provide that separation. A brand-appearance percentage does not establish factual correctness.
For the wider client document, use the AI visibility reporting framework. Carry forward one conclusion about factual reliability, one about discovery in the defined sample and the next action supported by each. Readers should be able to see whether the product was described correctly, whether it appeared without being named in the question, and what the team will investigate next.