On this page
A teammate sends a screenshot: “The AI is describing our product incorrectly.” The image shows one sentence, but not the question, date or source behind it.
The concern may be valid. The screenshot alone leaves the next person guessing about what was asked, what the answer actually said and which fact needs checking.
An evidence packet is a small record of one observation. It preserves enough context for another person to inspect the same answer and decide what to investigate. It can support a support request, an editorial task or one example in a client report.
Write the question the investigation should answer
Start with one concrete concern. “Our AI visibility is wrong” is difficult to investigate. “This saved answer names our product, but the result says it was absent” points to a specific classification question.
Other useful investigation questions include whether an answer describes an unsupported capability, whether a citation supports the sentence beside it, or whether a similarly named business was counted as the target brand.
Keep the concern separate from the proposed remedy. An inaccurate statement does not yet establish that you need a new article. It might reflect an outdated page, an ambiguous name or an unsupported interpretation of the answer.
Preserve the original question and context
Copy the exact question, including meaningful constraints such as market, language and business type. Record when the answer was collected and where it came from: a monitored provider API call, a consumer application or another tool.
Include the provider/model/search information actually exposed by the record. Leave unavailable settings marked unknown. Do not fill them in from another conversation or assume a familiar provider name identifies the complete request configuration.
For a continuing conversation, preserve the relevant earlier context through an approved record or a clearly labeled summary. A standalone question and a question following several brand-specific messages are different observations.
When the answer comes from PB, include the selected site, collection and run reference where available. Also record whether the run succeeded. A failed or incomplete operation is a collection issue, not usable evidence that the brand was absent.
Keep the answer attached to its source
Preserve the original answer in an approved location, then quote only the passage needed for the investigation. Include enough surrounding text to retain qualifications, negation and the purpose of the list or recommendation.
For example, “HarborDesk supports this workflow” and “HarborDesk does not support this workflow” cannot be distinguished by a cropped brand-name highlight. An answer may also name a product specifically to exclude it from a recommendation.
Screenshots are useful when a displayed state or a source marker matters. Keep the question and date visible where practical, and pair the image with text that can be reviewed. If you redact private material, identify the redaction without rewriting the remaining answer.
A repeat request can produce another observation. Save it separately with its own date; do not substitute it for the original answer that prompted the concern.
Record the relationship of each source
A source beside an answer can play different roles. In your packet, distinguish an explicit citation from a source returned alongside the response or recorded as search evidence. Then record whether it belongs to your approved domain or a third party.
These are separate questions. A third-party source can discuss your brand, and an owned source can be returned without being explicitly cited in the final answer.
When you inspect a linked page, record the time of that inspection and what the page currently says. A page fetched today does not necessarily show the content that was available when an older answer was generated. Keep any approved historical evidence separate from the current page check.
The guide to mentions, citations and recommendations explains the underlying distinctions. In this packet, the purpose is to attach each interpretation to the specific record that supports it.
Separate the observation, interpretation and next action
Illustrative example only. HarborDesk and the records below are fictional; they demonstrate an investigation format, not a real provider answer or customer issue.
Suppose a recorded answer says the fictional desk-booking product supports recurring reservations. A later inspection of its approved documentation says recurring reservations are not supported.
| Part of the packet | What the reviewer records |
|---|---|
| Observation | The saved answer attributes recurring reservations to HarborDesk |
| Corroborating record | The approved product documentation inspected later says the feature is unsupported |
| Current interpretation | There is a discrepancy between the answer and the current documented capability |
| Uncertainty | The packet does not yet establish what the documentation said when the answer was collected, or which evidence led to the statement |
| Next action | Check the relevant source history and clarify any inaccurate public product explanation |
That is enough to begin useful work. It is not enough to claim that a particular page caused the answer or that publishing a correction will change future answers.
A copyable investigation record
Use the following as a manual worksheet. The bracketed fields are instructions for the person preparing the record, not PB-generated output.
# AI answer investigation
Concern: [One specific question to resolve]
Owner: [Person responsible for the next check]
## Original observation
- Site/brand: [Correct context]
- Exact question: [Submitted wording]
- Answer collected at: [Date, time and timezone]
- Surface/provider: [API, consumer app or other]
- Model/search/market/language: [Known settings; unknowns stated]
- Relevant prior context: [Approved reference or labeled summary]
- Run status and record reference: [Available evidence]
- Original answer location: [Approved access-controlled location]
- Relevant excerpt: [Exact passage with necessary context]
## Sources and comparison
- Recorded URL and relationship: [Explicit, returned or searched]
- Ownership: [Approved owned domain, third party or unresolved]
- Source inspected at: [Separate inspection time]
- Current supporting/conflicting evidence: [Reviewed record]
- Historical source evidence: [Available record or unavailable]
## Assessment
- Observed fact:
- Interpretation:
- Unresolved question:
- Next action and acceptance condition:
- Sharing/redaction limits:Keep the record proportionate. One answer with one disputed claim rarely needs a long report. Link the underlying material and state the next decision clearly.
Use PB's evidence without assuming a full export
Prerender Buddy's recorded answer and source views can support this investigation where AI Visibility is available to the account. Open the underlying result when the judgment requires its wording or source context.
Scoped workspace API/MCP summaries do not necessarily contain the full answer body. A summary is useful for locating the relevant evidence, but it should not be expanded into an invented quotation or treated as an exact copy of the original answer.
Likewise, a shared technical visibility-check report describes a public-page check. It does not automatically export this answer investigation or the full client-reporting method.
Hand over a decision another person can verify
Choose the next action based on the concern: investigate collection or classification, verify a product claim, inspect the cited source, or gather missing context. Include an owner and a clear completion condition.
For example: “Confirm whether the approved documentation supported this capability on the collection date, and attach the evidence.” That creates a useful next step without promising a future citation or immediately assigning a website rewrite.
When the investigation feeds a broader client update, use the AI visibility reporting framework. Keep the packet behind the summary so the next reviewer can inspect the same observation instead of trying to recreate it.