Trace a numerical AI claim to the original measurement before repeating it. A citation to an article that quotes another article is only the beginning of verification.
Break the statement into checkable parts
Record the number, unit, population, geography, time window, numerator, denominator and claimed relationship. “Adoption doubled” is incomplete until the baseline, end value and measured group are known.
The FActScore paper describes evaluating atomic facts against reliable evidence. Our numerical review checklist is an editorial application of that principle, not a new automated factuality score.
Follow the source chain
Open the cited page and locate the exact passage. If it attributes the figure to a study, follow that study. Look for its methods, dataset definition and publication date rather than stopping at a headline.
Illustrative example: an answer says “60% of retailers use the feature,” but the cited survey concerns respondents who already use a particular platform. The arithmetic may be correct within that survey while the broader retailer claim is unsupported.
Do not substitute a similar number from another study without correcting the original statement. A plausible-looking percentage is not a valid replacement source.
Verify scope and calculation
- Confirm that the cited number actually appears in the original source.
- Check whether it describes people, companies, responses or events.
- Distinguish percentages from percentage-point changes.
- Recalculate simple ratios from available inputs.
- Record selection, missingness and weighting limitations.
- Check whether the conclusion is descriptive, correlational or causal.
If the dataset or method is unavailable, say what cannot be verified. Do not present a precise benchmark as independently validated merely because several websites repeat it.
Write the supported version
A defensible report may retain a narrower claim with its population and date. It may also remove the number and describe the unresolved question. Choose based on what the evidence establishes.
Keep the observed AI statement separate from your corrected wording. Record the verification date and correction owner so later readers can understand the change.
Historical research can inform a method without representing current provider performance. Avoid converting an older experimental result into a present-day market statistic.
Use the citation quality guide for the broader evidence review. Next, select one numerical claim from the citation tracking workflow, trace it to its original method and withhold it from public copy if that chain cannot be inspected.
Fact-checked: 17 October 2026. Examples are illustrative; no customer results are reported.