A citation beside an AI answer can be a useful shortcut to a source. It is not the source's endorsement of the surrounding sentence. The document might exist without containing the claimed result, and a real passage might support only one part of a broader conclusion. Verification becomes easier when the answer is broken into small claims and each claim is compared with the actual evidence. The following exercise uses invented policy wording to show the method; it does not describe a real contract or claim to reproduce an experiment.
Split the sentence before evaluating its links
Consider a fictional answer: a service deletes every uploaded file immediately and never keeps operational records. Even if that sentence has one citation, it contains several questions. Does the cited policy cover uploaded files? What does deletion mean? Does the timing apply to all accounts? What does the document say about operational records? A single link cannot be assumed to settle each question.
Copy the important claims into separate rows. Preserve qualifiers such as only, normally, after approval and except. Record whether the answer presents the statement as an explicit source fact, a comparison or an inference. This distinction matters because an inference may be reasonable without appearing verbatim in the source. It still needs explained premises. If an answer cannot be decomposed into checkable statements, ask for a narrower answer before spending time chasing its bibliography.
Establish that you have the intended document
Open the cited destination through a trusted route. Confirm its title, publisher, date and version. A plausible title can lead to an unrelated publication, and a link may resolve to an updated document that no longer matches an old quotation. For a PDF, check printed page numbers against the viewer's page count. A cover and introductory pages can create an offset.
Keep a local note of the URL and document identifier for an important decision. That note should identify the material you actually read, not merely the page an assistant proposed. If access is unavailable, state that the citation remains unverified. A search snippet or another generated summary can help locate a document, but it cannot replace reading the passage. Do not bypass access controls or upload a private source to an unfamiliar verification service merely to complete a checklist.
Test whether the passage supports the exact claim
Suppose the fictional source says that an uploaded file is removed from a user's workspace after a deletion request, while security records follow a separate retention policy. That would not establish immediate removal of every copy. It would support a narrower statement about a specific interface action and leave another question open. The corrective action is to narrow the answer, not to find more confident wording.
Read the paragraph before and after the relevant line. Inspect definitions and referenced exceptions. If a table supplies the evidence, retain its column heading, units and notes. An outcome measured for one group or configuration cannot automatically be extended to a different one. NIST's generative-AI risk work identifies confabulation as a risk; a fluent citation presentation does not eliminate it. The decisive comparison remains between the claim's scope and the material the source actually establishes.
Follow numbers back to their denominator
A percentage needs a population, time period and definition of success. Ask what was counted, what was excluded and whether the comparison uses the same conditions. An assistant can copy a number correctly while changing its meaning by dropping a denominator. A result for completed trials, for example, may not describe all attempts if failures were excluded.
Create a compact evidence record with the claim, source location, relevant measure and missing conditions. Do not invent a confidence interval, benchmark environment or sample size to make the record look scientific. If the source does not provide the necessary information, preserve that limitation. For a consequential claim, locate the original study, specification or dataset rather than relying solely on a press announcement. A primary document can still have weaknesses; reading the original makes those weaknesses inspectable instead of automatically transferring authority to a summary.
Distinguish repeated links from independent support
Three webpages can repeat one announcement. They are three destinations, not necessarily three independent observations. Trace the evidence chain until you know whether the pages report their own work, interpret an original publication or simply copy its description. Keep commentary separate from the underlying observation in your notes.
This also prevents a circular check in which an AI answer cites a page that was itself generated from a similar answer. Publication identity, method and traceable source material matter more than the number of links. When sources disagree, compare their dates, populations and definitions before declaring one false. The disagreement may reflect different questions. Where the conflict cannot be resolved proportionately, present both qualified positions and identify what additional evidence would change the decision. Do not manufacture consensus by silently dropping the inconvenient source.
Return a claim-level verdict
Use a small set of descriptions: supported as written, supported only with qualifications, not established by this source, contradicted, or inaccessible.
For the fictional retention answer, the useful result might be that workspace removal is documented but universal immediate deletion is not. That result can guide a real procurement question without pretending to audit a backend. Verify high-impact claims first, retain unanswered questions and remove unsupported certainty before sharing the answer.
One citation, three different verdicts
This exercise uses invented policy language. It demonstrates support judgments without claiming to audit a real service.
Fictional evidence passage
POLICY SAMPLE, section R2: After an approved removal request, the uploaded file stops appearing in the user workspace. Security records follow a separate retention policy. This section does not state their duration.
Select or copy plain text
POLICY SAMPLE, section R2: After an approved removal request, the uploaded file stops appearing in the user workspace. Security records follow a separate retention policy. This section does not state their duration.
| Claim | Verdict | Reason |
|---|---|---|
| Workspace visibility ends after approved removal | Supported as written | Matches the stated location and condition |
| Every copy disappears immediately | Not established by this passage | All copies and immediate timing are not specified |
| There are no security records | Contradicted | The passage expressly describes a separate security-record policy |
| Security records last exactly thirty days | Not established by this passage | No duration is given |
Select or copy plain text
Claim Verdict Reason Workspace visibility ends after approved removal Supported as written Matches the stated location and condition Every copy disappears immediately Not established by this passage All copies and immediate timing are not specified There are no security records Contradicted The passage expressly describes a separate security-record policy Security records last exactly thirty days Not established by this passage No duration is given
REFERENCES
Sources and further reading
Sources reviewed October 20, 2026. Provider instructions and product terms can change. Examples identified as fictional are editorial illustrations, not reported events or product tests.