Benchmark cohort · 12 May 2026

Which search API finds official docs best for AI agents?

We ran Brave Search API, SerpAPI, and Tavily against 30 documentation-retrieval tasks an agent might perform while building or debugging software. This cohort reports objective retrieval metrics, not a custom black-box score.

Short verdict: SerpAPI had the strongest relevance metrics in this 30-task official-docs cohort, with 100% Success@3 and 0.933 MRR. Brave was close on relevance, faster, and also reached every expected source within the top 10. Tavily returned useful results but more often ranked third-party pages above official docs for this specific task family.

Results summary

1

Serpapi

100% Success@3 across 30 official-docs tasks. Success@1: 90%. Median latency: 2016 ms.

2

Brave

97% Success@3 across 30 official-docs tasks. Success@1: 83%. Median latency: 1074 ms.

3

Tavily

80% Success@3 across 30 official-docs tasks. Success@1: 47%. Median latency: 1623 ms.

ProviderSuccess@1Success@3Success@10MRRMedian latency
Serpapi90%100%100%0.9332016 ms
Brave83%97%100%0.9031074 ms
Tavily47%80%100%0.6351623 ms

Success@k means at least one expected official source appeared in the top k results. MRR is mean reciprocal rank for the first expected source. Relevance labels were URL-pattern and official-domain based, then reviewed for moved official docs domains before publishing.

What this means for agent workflows

Method in brief

The task set contains 30 official-documentation queries across AI APIs, browser automation, infrastructure, data stores, and workflow tools. Each provider was called once per task on 12 May 2026. We saved response status, latency, result count, top result URLs, and rank observations.

The primary relevance signal was whether an expected official URL pattern or accepted official domain appeared in the top 10. This is intentionally simple and auditable. It does not judge snippet quality, generated answer quality, pricing, rate limits, or long-term stability.

Download the evidence tables

Limits and next steps

Need this for your stack?

AgentFirstTools can inspect a tool shortlist or agent workflow and produce a narrow evidence-backed audit before you depend on it in production.