Context
OpenDS-World introduces 200 independently verified tasks requiring autonomous web search, browsing and evidence integration across heterogeneous online formats; 12 frontier models were evaluated under a unified tool protocol.
What changed
Established a live-web deep-search evaluation setting spanning navigation, multi-source evidence and interactive access, beyond static retrieval-only benchmarks. The best evaluated model achieved 39.0% accuracy on the task set.
Why it matters
AI discovery depends on whether agents can find, navigate and interpret source material before deciding what to cite or use in an answer.
Visibility OS interpretation
Visibility OS interpretation: measure agent access, extraction, evidence integration and final answer quality as distinct stages. Benchmark scores do not reveal production search ranking or citation preferences.
What remains uncertain
How representative are the 200 selected tasks? How do tool constraints and model versions affect results? Can benchmark performance predict real consumer agent referrals or source citations?