People Search Bench

Leaderboard • Methodology • Case Studies • Reproducibility

No existing benchmark measures how well AI can find real people from natural language queries. We built one: 119 queries, 4 scenarios, 3 scoring dimensions, all graded against web evidence — not LLM opinion. [Paper](https://arxiv.org/abs/2603.27476).

Leaderboard

Platform	Relevance Precision	Effective Coverage	Information Utility	Overall
Lessie	70.2	69.1	56.4	65.2
Exa	53.8	58.1	53.1	55.0
Claude Code	54.3	41.1	42.7	46.0
Juicebox (PeopleGPT)	44.7	41.8	50.9	45.8

How scores are computed

We extract checkable criteria from each query, verify each returned person against those criteria via web search (Tavily API), and produce a relevance grade (0–1) per person.

• Relevance Precision — padded nDCG@10. The ideal DCG assumes 10 perfect results exist, so returning 3 perfect results out of 3 still scores below a platform that returns 10.
• Effective Coverage — TCR × mean(min(qualified / K, 1.0)) × 100. Qualified = relevance grade >= 0.5.
• Information Utility — average of profile completeness, query-specific evidence, and actionability.
• Overall — equal-weight mean of the three.

**Platform notes:** Juicebox lea