Which models are tested?

Current frontier models from the major providers, answering the same prompts under the same conditions. Model versions move quickly, so each run records what was used.

What does the Citation Integrity™ Intervention Layer do in the test?

It supplies structured Citation Integrity™ evidence alongside the question, so the model can see which sources are corroborated before it answers.

Does a better benchmark result mean my business will be cited?

No. The benchmark measures model behaviour in a controlled test. It does not predict or guarantee how any external AI system will treat a particular organisation.

CiteAbility™ Labs

Citation Integrity™ Benchmark

Testing whether AI-generated claims are actually supported by their evidence and citations: a Baseline AI Answer compared with an AI Answer with Citation Integrity™ Intervention.

See the framework Back to Labs

By CiteAbility™ · Last updated 02 September 2026

Key takeaways.

The benchmark asks identical questions of the same models, once as a Baseline AI Answer and once as an AI Answer with Citation Integrity™ Intervention.
It measures whether citations actually support the claims made, not whether an answer sounds confident.
Results are published as observed model behaviour on the date of the run.
The benchmark is Labs work in progress and its method may change.
CiteAbility™ cannot guarantee citation, recommendation, ranking or inclusion by any external AI system.

Method

Same question, same model, one variable.

Does applying the Citation Integrity™ Framework improve the integrity of claims, evidence and citations in AI-generated answers? Each model answers the identical question twice: once as a Baseline AI Answer, and once as an AI Answer with Citation Integrity™ Intervention, where the Citation Integrity™ Intervention Layer is applied during source selection. Wording, model version, generation settings and answer constraints are held the same.

Sources are selected, frozen and assessed from one shared evidence pool before either answer is generated. Both answers then work from that same frozen evidence: the Baseline AI Answer receives no Citation Integrity™ assessment of it, and the AI Answer with Citation Integrity™ Intervention receives the Framework's assessment of exactly the same sources, so both columns are compared on the same footing.

The Citation Integrity™ Benchmark measures whether claims in AI-generated answers are supported by the evidence and citations attached to them. It compares a baseline AI answer with the same model operating under a Citation Integrity™ intervention. It does not measure whether an answer is factually correct. The Citation Integrity™ Framework is developed by CiteAbility™.

These experimental results do not guarantee behaviour in other models, questions or live search environments.

Results

What the runs show.

Loading results…

FAQ

Frequently asked questions

The questions organisations ask us most often.

Which models are tested?: Current frontier models from the major providers, answering the same prompts under the same conditions. Model versions move quickly, so each run records what was used.
What does the Citation Integrity™ Intervention Layer do in the test?: It supplies structured Citation Integrity™ evidence alongside the question, so the model can see which sources are corroborated before it answers.
Does a better benchmark result mean my business will be cited?: No. The benchmark measures model behaviour in a controlled test. It does not predict or guarantee how any external AI system will treat a particular organisation.