Verify citations in an AI research draft
Check AI citations for URL availability, original provenance, passage support and current relevance, using a claim-by-claim acceptance worksheet.
What must a citation prove?
A citation needs more than a working link. It should lead to evidence that supports the precise sentence, including population, period and limitations. An AI draft can attach a plausible official page to a claim that the page never makes. Check the passage before deciding that the source is valid.
OpenAI notes that structured outputs can still contain factual mistakes and recommends review with access to original information. A valid citation field therefore remains a review task. Neither JSON validity nor a source-looking title establishes passage support. [Structured model outputs] [Safety best practices]
A completed citation acceptance worksheet
| Check | Observed evidence | Decision |
|---|---|---|
| URL exists | Official documentation page opens | Availability passes only |
| Primary provenance | Publisher owns the product documentation | Suitable for a product behavior claim |
| Passage support | Page describes event collection, not customer qualification | Reject “event delivery proves a qualified lead” |
| Date and scope | Documentation applies to the named product and collection method | Retain only the supported technical statement |
| Publishable wording | Delivered event confirms collection in that test | Keep business outcome verification separate |
The completed worksheet can narrow the draft to a useful statement: a delivered event shows that the tested measurement reached the collection system. It does not establish that a real person became a qualified lead or customer. That distinction is operationally important even when the same event name appears in a conversion report.
If a source is a summary, follow its citation to the original evidence where practical. Several articles linking to one report do not supply several independent studies. Keep the chain so a later reviewer can tell where the claim originated and which retelling introduced a stronger interpretation.
How should numerical claims be checked?
- Locate the exact table, passage or dataset supporting the number.
- Record the denominator, population, collection period and unit.
- Reproduce arithmetic where the draft derives a percentage or ratio.
- Check whether relative change was confused with percentage points.
- Inspect the title, chart and summary for stronger wording than the evidence supports.
A fictional example shows the distinction: a report says 18 of 60 respondents selected an option. That is 30% of those respondents. A draft that says 30% of all customers has changed the population. Correcting the URL does not repair that claim; the sentence itself must change.
Do not substitute a newer-looking page automatically. A recent overview may omit the method needed to support the original statistic. Prefer the source that actually substantiates the statement, and explain historical scope when it matters. If current relevance is unknown, say so or remove the time-sensitive assertion.
What happens when support cannot be found?
Classify the failure: unavailable URL, secondary provenance, missing passage, wrong scope or obsolete evidence. That classification helps decide whether to replace the source, narrow the claim or remove the sentence. Asking the model to “try another citation” without preserving the claim can produce another plausible mismatch.
I would publish fewer supported claims rather than keep an unsupported percentage for visual authority. An original worked example can demonstrate a method without borrowing a questionable benchmark. The final article should let a reader reproduce the reasoning from the evidence that is actually available.
Sources
AI-assisted writing and an AI-generated editorial cover. Source links and checked dates are listed below; illustrative examples are labeled in the text.
- Structured model outputs — checked 2026-10-18
- Safety best practices — checked 2026-10-18