WebsiteBench: Browser-Evidence Reconstruction for Website-Cloning Agents

Current evaluations of browser agents and coding agents measure two different capabilities: exploring existing websites and modifying software artifacts. And website reconstruction task requires an agent to coordinate both capabilities by acquiring browser evidence, translating observations into executable code, and validating the resulting artifact against the observed interface. We introduce WebsiteBench, a benchmark for website reconstruction from browser observations. Given a sealed local replay of a public webpage, an agent must produce an offline clone whose appearance, structure, assets, and interaction behavior match the hidden reference. WebsiteBench separates reconstruction from direct imitation through sealed local replays and hidden evaluation. Agents interact with the target through browser observations, and submissions are evaluated offline by comparing their visual appearance, structural organization, and reproduced interactions with hidden references. On 250 sealed website reconstruction tasks across 12 categories, a complete gpt-5.5-xhigh run achieves 0.552 overall fidelity. The agent performs substantially better on visual reconstruction (0.718) than DOM/accessibility preservation (0.380), revealing a gap between reproducing surface appearance and recovering the underlying structure of a webpage. These results show that website reconstruction remains an unsolved capability for current browser-control coding agents and provides a controlled setting for studying how agents translate browser observations into executable software artifacts.