New protocol brings transparency to AI audits of website redundancy

Assessing whether a website has too many repeated elements sounds straightforward, but context matters enormously. A duplicated navigation menu might feel redundant during routine browsing yet become a lifesaver when part of the interface fails. Researchers have now proposed a more systematic approach to this fuzzy problem.

Introducing CORA#

Scientists have developed CORA (Counterfactual, Observable Redundancy Audit), a framework designed to replace subjective evaluations with a structured, auditable process. Rather than assigning a single "redundancy score," the protocol breaks the question into three measurable components: the burden of repetition during normal use, the baseline demand on the interface, and how well duplicated elements help recovery when failures occur.

Each audit run captures detailed evidence—including screenshots, persistent identifiers for page elements, and logs of user interactions. A versioned vision-language model generates initial annotations, which then pass through a strict validation layer and quality gates before any score is released.

Testing the approach#

When tested on a controlled benchmark with known properties, CORA's three-part measurement outperformed simpler single-number approaches at predicting whether repeated elements would survive when the interface was deliberately altered. However, when researchers deployed two different small AI models as auditors, both produced consistent outputs yet neither cleared all the protocol's release requirements. CORA appropriately withheld final scores in both cases while preserving the raw observations and error logs for inspection.

The team confirmed separately that the validation checks and gatekeeping mechanisms work as designed—though passing these technical tests does not guarantee that the measurements will prove accurate or meaningful in the real world.

What remains unknown#

The researchers present CORA as a candidate procedure suitable for the specific, controlled test environment they studied, not as a universal standard. Critical questions remain unanswered: whether human evaluators would agree with the measurements, how the AI auditing compares to human judgment, and whether the approach generalizes beyond the narrow benchmark to messy production websites. These open questions underscore the gap between building a reproducible auditing framework and establishing one that genuinely captures meaningful aspects of user experience.