A new AI framework watches how it browses a website, then tells you exactly what to fix.
Researchers describe AMUSER in a paper posted to arXiv, a system that has LLM agents simulate user behavior on live websites using both the page's visual layout and its text, then automatically generates and ranks UX improvement recommendations. Tested on commercial websites, the multimodal version scored 0.758 on a relevance metric called NDCG@3, versus 0.359 for a text-only simulation - roughly twice as accurate. It also ran at 89% lower simulation cost than the text-only baseline. Recommendation quality was checked against human expert annotations and an LLM-as-judge protocol.
UX audits today usually mean paying humans to click through a site and write notes, which is slow, pricey, and inconsistent between reviewers - a real bottleneck for teams that can't afford proper user testing before every redesign. If a cheap simulated audit gets close to human-level findings, that changes who gets access to real UX feedback, not just how fast it arrives. One wrinkle worth noting: the paper found that feeding the agent images while it writes up recommendations can slightly hurt quality. It is seeing while acting, not seeing while summarizing, that seems to help.
This is a preprint tested on unnamed commercial websites chosen by the authors, so treat the twice-as-good number as a lab result until someone points AMUSER at a live site nobody optimized for in advance.