Can AI-written web pages actually be used? Tencent built an evaluation that clicks through web pages on its own.

Beating AI News Flash: Tencent Hunyuan, in collaboration with Tsinghua University and Peking University, has launched IWC-Bench, specifically designed to evaluate how AI-generated webpages actually perform in real-world use.

Many webpage generation benchmarks focus mainly on code or screenshots. A page may look beautiful, and the code may indeed implement the functionality, but buttons cannot be clicked, forms cannot be submitted, and switching pages triggers errors. These problems are difficult to detect through static checks.

IWC-Bench is more like real human acceptance testing. It has an Agent actually open the webpage, click buttons, enter content, switch pages, and then check which functions were actually executed. Finally, it separately evaluates three things: whether the page looks good, whether the operation is smooth, and whether the features requested by the user were implemented.

The benchmark includes 369 real user requirements and 5,088 acceptance criteria, with 17 models generating a total of 6,273 web applications. In another 197 sets of manually reviewed comparisons, IWC-Bench matched human choices in 168 sets, achieving a consistency rate of 85.3%.

The results also show that a webpage "looking good" and "actually being usable" are not the same thing. GPT-5.6-Sol ranked highest in page aesthetics, but only sixth in usability; when it comes to individual webpages, good looks basically cannot indicate whether they are easy to use, with a correlation coefficient of only 0.36 between the two.