NEEDLE benchmark exposes web search blind spots

The result is a sharp reminder that search infrastructure—not just model intelligence—sets an agent’s ceiling.

• –NEEDLE covers news, finance, scholarly retrieval, legal search, and rare long-tail queries.
• –Engines receive identical queries and are judged on returned titles, snippets, ranking, and answer recall.
• –The 75.7% figure represents share of an oracle assembled from all providers, not coverage of the entire internet.
• –Google remains strong on known-item lookups but falls behind on fresh, obscure, and agentic queries.
• –Because the benchmark is public, continuously refreshed, and open source, providers have less room to optimize for a fixed leaderboard.