Citation support rate
Reviewed citation–claim pairs that support the associated claim divided by all reviewed citation–claim pairs. Use a declared rubric for complete and partial support.
Illustrative calculation, not measured data
- Reviewed pairs supporting the claim
- 8
- All reviewed citation–claim pairs
- 10
8 / 10 × 100 = 80%
What it means for GEO
Inspect whether a source actually establishes the stated feature, date or constraint. Report the review sample and ambiguous cases instead of relying on citation volume.
What not to infer
Support rate is not a complete factual-accuracy score. Uncited claims and omitted limitations require separate review.
How it works
Review a specific claim against the source attached to it. For the educational formula here, full support means the cited evidence establishes the complete claim and its material conditions; partial support is recorded separately and excluded from the full-support numerator. Contradicted and unsupported pairs also do not count as full support. Keep inaccessible sources unresolved instead of guessing. Citation support tests attribution, while factual correctness and citation completeness require additional review, including statements with no citation.
GEO example
Illustrative scenario, not observed assistant behavior or measured results.
Fictional worked example: reviewers inspect ten accessible citation–claim pairs. Eight fully support their claims, one supports only the feature but not the claimed availability on every tier, and one does not support the claim at all. Full-support rate is 8/10, or 80%; report the partial and unsupported cases as well. A citation to a vendor homepage does not establish every product detail simply because the vendor is authoritative. If a compound statement combines two facts, apply a consistent rule for splitting it into reviewable claims.
What you can do
- Create a rubric covering full, partial, unsupported, contradicted and unresolved cases.
- Record the exact claim, cited passage, source version or access date and reviewer rationale.
- Have another reviewer check ambiguous cases and document disagreement rather than silently changing the rule.
What to measure, and what it cannot prove
Report eight fully supported pairs out of ten reviewed accessible pairs in this fictional sample. If additional sources could not be opened, show that unresolved count and the exclusion rule. Do not extrapolate the rate to all answers or to uncited claims. Automated support judgments should be checked against human review rather than treated as infallible.
Evidence and authoritative sources
- Gao et al.: Enabling Large Language Models to Generate Text with Citations (EMNLP 2023)
Separates correctness and citation quality and studies evaluation against human judgments. Our manual rubric and 8/10 example are not ALCE results.
- OpenAI: Searching the web with ChatGPT
Advises opening cited sources to verify support and warns that citations can be incomplete or incorrect.
The sources establish the documented concepts and product behavior. The publishing checklists, scenarios and operational measurement rules are GeckoCheck educational interpretations, not provider-endorsed ranking instructions.
Source: Gao et al.: Enabling Large Language Models to Generate Text with Citations (EMNLP 2023) · OpenAI: Searching the web with ChatGPT
Period: Reviewed October 12, 2026. Living provider documentation; cited research published in 2023 and 2024 where applicable.
Sample: Fictional manual-review example: 8 of 10 pairs supported. Not a benchmark for any model.
Limitation: GeckoCheck educational interpretation. Examples are fictional; formulas use declared operational definitions, not universal provider metrics. Source documentation does not guarantee citation or ranking outcomes.