Researchers built a benchmark that treats AI-generated web apps the way a QA team would: by actually clicking through them, not just reading the code.
The new system, called IWC-Bench, instruments each generated app and uses code coverage to guide an AI agent through simulated user interactions, rather than following a fixed checklist. It turns that exploration into a state-transition graph and scores every app on three axes: visual aesthetics, usability, and how well it matches the original requirements. The benchmark draws on 369 real-world requirements and 5,088 acceptance criteria, and its creators ran 17 frontier models through it. No single model led on all three dimensions.
Most prior benchmarks either credit code for features nobody ever sees rendered, or run a scripted set of test cases that miss functionality outside the script. IWC-Bench's coverage-guided exploration is built to close that gap, and its authors report 85.3 percent agreement with human judges across 197 sampled sessions, with model rankings holding steady even when the judge model changes.
Benchmarks that claim to match human taste have a habit of aging badly once developers start building specifically to pass them.