One of the authors here. We have been seeing lots of benchmaxxing and leakage in standard web search benchmarks such as BrowseComp.

We developed this live benchmark with daily/hourly sampled fresh queries matching real agentic search traffic to estimate actual search performance of different AI search providers.