Relevance isn’t enough: what it takes to rerank the web

Exa presents an evaluation of web rerankers that argues topical relevance alone is insufficient for AI-agent search, which also requires handling navigational intent, content freshness, source authority, adversarial manipulation, coding queries, and user instructions. It compares Exa’s production rerankers with Cohere, Voyage AI, and Google Cloud models across three public benchmarks—BEIR, MIRACL, and MAIR—and four internal benchmarks covering coding, navigation, freshness, and authority, using common candidate lists and nDCG@10 scores with confidence intervals. The results show closely clustered performance on conventional relevance tasks, while Voyage and Cohere lead on Exa’s coding set and Voyage leads instruction following on MAIR. Exa reports stronger results on its internal navigation, freshness, and adversarial-authority evaluations, with its Fast model outperforming competing default-tier systems in the reported seven-benchmark macro average. The company concludes that ranking systems should optimize for multiple search objectives and notes that it does not currently plan to provide its reranker as a standalone API.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.