---
type: "post"
id: "relevance-isn-t-enough-what-it-takes-to-b2317f"
slug: "relevance-isn-t-enough-what-it-takes-to-b2317f"
title: "Relevance isn’t enough: what it takes to rerank the web"
description: "Exa published an evaluation of the reranker used in its production web search."
retrieval_nugget: "Exa published an evaluation of the reranker used in its production web search."
published_at: "2026-10-23"
updated_at: "2026-10-23"
record_date: "2026-10-23"
date_kind: "published_at"
topics: ["google"]
entities: ["Exa","Exa Reranker Auto","Exa Reranker Fast","Cohere","Voyage AI","Google Cloud","BEIR","MIRACL"]
related_patterns: ["ai-engineering-evaluations-are-splitting-into-task-specific"]
novelty_label: "new_angle"
what_changed: "The post argues that agent-facing reranking must assess intent, freshness and other dimensions beyond topical relevance, unlike prior Exa search API benchmark coverage."
prior_refs: ["memory:9b45fbce-3f08-8f3e-ac74-345ace2b017b","memory:51316159-a897-8cf8-aff1-f128497e09d5"]
source_url: "https://exa.ai/blog/sota-reranking"
source_urls: ["https://exa.ai/blog/sota-reranking"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:relevance-isn-t-enough-what-it-takes-to-b2317f"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/relevance-isn-t-enough-what-it-takes-to-b2317f/","markdown":"https://newruntime.com/posts/relevance-isn-t-enough-what-it-takes-to-b2317f.md","json":"https://newruntime.com/posts/relevance-isn-t-enough-what-it-takes-to-b2317f.json"}
---

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

## Retrieval answer

Exa published an evaluation of the reranker used in its production web search.

Exa published an evaluation of the reranker used in its production web search. It compares Exa Reranker Auto and Fast with Cohere Rerank 5, Voyage Rerank 3, and Google Semantic Ranker 5. Exa says its retrieval system searches a pool of more than 100 billion web documents, then reranks a shortlist within hundreds of milliseconds.

The company evaluated seven benchmarks: public BEIR, MIRACL, and MAIR, plus internal coding, navigational, freshness, and authority sets. Each system received the same 50 candidates; Exa reports nDCG@10 with bootstrapped 95% confidence intervals. Its reported seven-benchmark macro scores are 0.738 for Auto and 0.733 for Fast, against 0.711 for Voyage, 0.693 for Cohere Pro, and 0.682 for Google Semantic Ranker.

Results vary by task. On Exa's internal set of 317 coding queries, Voyage and Cohere score above Exa. The company also says competing rerankers lowered the ranking quality of Exa Instant on its navigational tests. Exa does not plan to release a standalone reranking API. The source does not establish independent performance leadership: four benchmarks are internal, and it says some internal labels were produced by GPT-5.4.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)