A cited source can be authoritative and still fail to support a particular claim. It may discuss a different product, cover only one region, describe a preview, or supply a narrower result than the answer suggests.
Audit the relationship between the claim and the source, not just whether a link exists. Save the exact sentence, its material conditions, the cited destination, and the passage that appears to support it. Then classify the support and decide what needs correction.
The miniature sources and answers below are invented teaching examples. They are not quotations from a real company, model output, or research paper.
Break a compound claim into checkable parts
Consider a fictional answer: “The product is available worldwide, includes automated exports on every plan, and cuts review time in half.” That sentence contains at least three different factual claims. One citation at the end may support only one of them.
Split the sentence into availability, plan coverage, and performance. Preserve the qualifiers for each: available where, on which plans, for whom, measured how, and during what period? The audit row should be small enough that a reviewer can make a clear support decision.
Do not add missing conditions from memory. The claim being audited is what the answer actually says. A knowledgeable reviewer may know a more accurate version, but that should appear as a proposed correction with its own evidence.
Use four support categories
These categories are an editorial method, not a universal standard. Use them consistently and preserve the evidence behind the code.
| Invented source | Invented answer claim | Classification |
|---|---|---|
| The export feature is available on the Team plan. | The Team plan includes export. | Supported for this narrow claim |
| A beta is available to selected US accounts. | The feature is available to customers. | Partial: selection, region, and beta limits are missing |
| The product does not support offline mode. | The product supports offline mode. | Contradicted |
| A company announces its new office. | The product reduces processing time by 50%. | Irrelevant to the performance claim |
Add an unavailable or unreadable state when the source cannot be inspected. That state describes missing evidence, not proof that the claim is false. Record the failure and avoid promoting the citation into verified support.
Check the source's scope and date
For an announcement, distinguish its publication date from the event it describes. For documentation, record when it was checked and any stated update date. A page reviewed today can still describe an older launch or an unfinished preview.
Read the source's surrounding conditions. A result from a small internal test should not become an industry-wide performance claim. A vendor's own evaluation should remain attributed to the vendor unless independent evidence supports a broader conclusion.
Follow the cited destination far enough to identify the actual supporting material. A publisher summary may point to an original report with different definitions. Preserve both roles when they matter: what the answer cited and what the underlying evidence establishes.
Write a correction that preserves the useful answer
For the partial beta example, a supported correction would retain the selected-account, US, and beta conditions. The reader can then decide whether the feature is relevant without assuming general availability.
For the contradicted offline example, replace or remove the claim. Adding “may” to a statement contradicted by the source is not a sufficient repair. If product behavior has changed, obtain current primary evidence and explain the change rather than smoothing over the conflict.
For the irrelevant performance citation, search for actual measurement evidence if the claim is important. If none is available, remove the numerical promise. The source's prestige does not transfer support to a sentence about something it never measured.
Keep a compact audit record
Use fields for claim, answer context, source URL, inspected passage, source date, check date, support category, proposed correction, and reviewer. If the claim combines several ideas, create several rows linked to the same answer.
Report counts with their scope. “Eight of ten inspected claims were supported” describes that inspected set. It does not mean 80% of all statements an engine produces are correct. If the team deliberately selected suspicious citations, the result should be described as a diagnostic review rather than a representative accuracy estimate.
Review disagreements using the reviewer calibration method. A partial-support boundary can be subjective, especially when omitted conditions materially change a buying decision.
Connect the audit to the next action
If the problem is on your own page, correct the underlying explanation and preserve a useful revision note. If another answer misstates your content, document the evidence and use any appropriate correction channel without promising that a future answer will change.
Use the evidence editing guide when preparing your own articles. The objective is a claim whose source, wording, and conditions agree, so a reader can inspect the support without reconstructing the whole research process.
For the next implementation step, use Investigate a brand mention with no visible citation.