Researchers introduce Mango, an AI web agent that finds information 7% more reliably by exploring websites smarter instead of starting from the homepage

A research team published Mango, an AI agent (software that decides actions on its own) designed to navigate complex websites more effectively. In tests on real-world web tasks, Mango achieved a 63.6% success rate using GPT-4-mini, beating the previous best result by 7.3 percentage points.

Instead of always starting from a website's homepage and exploring randomly, Mango maps out the site's structure first, then strategically picks the best starting points for each search task—like knowing which section of a store to enter based on what you're looking for. It also remembers what it tried before, learning from failures to avoid repeating the same mistakes.

For anyone relying on AI to research information online—from students fact-checking for essays to business analysts gathering market data—this means fewer failed searches and faster, more reliable answers. AI assistants built on this approach would waste less time going down dead ends and actually find the information you asked for.