The fidelity score quantifies how accurately an AI model's grounded response reflects the specific web pages it cited:

\[\text{Fidelity Score} = \left( \frac{|\text{survived}|}{|\text{survived}| + |\text{lost}| + |\text{distortions}|} \right) \times 100\]

During assessment, a dedicated model call evaluates the retrieved page content against the specific response segments mapped to that URL, categorizing extracted facts into survived, lost, and distorted claims. If no claims are identified (totaling zero across all three categories), the formula returns a default score of 100.0%. Once every individual source URL for a search query has been evaluated, their individual percentages are averaged to generate the run's final aggregate fidelity score.