Story details
Exa published a study of web reranking on 23 October, describing how its search system orders a shortlist of pages after an initial retrieval stage. The company says ordinary relevance tests no longer clearly separate leading systems, so it compared its production rerankers with Cohere, Voyage AI and Google Cloud on a wider mix of tasks. The post is research reporting by the company that built the system; its reported advantages are not an independent evaluation or a measure of Wellficent search quality.
The distinction between retrieval and reranking is central to the study. Exa says its retrieval stage searches a pool of more than 100 billion web documents using compact representations, favoring recall and speed. A later reranker processes a much shorter list with more context and compute to decide which results should appear first. In the evaluation, models received the same query and candidate set in the same order, with their final ranking as the variable. Exa says it reports normalized discounted cumulative gain at 10, or nDCG@10, and 95% confidence intervals.
The comparison combines three public benchmarks—BEIR, MIRACL and MAIR—with four evaluations the company describes as internal: coding, navigational queries, freshness and authority under adversarial content. The study reports a macro average across seven benchmarks of 0.738 for Exa Reranker Auto and 0.733 for Exa Reranker Fast. Its table lists 0.693 for Cohere Rerank 5 Pro, 0.711 for Voyage Rerank 3 and 0.682 for Google Semantic Ranker 5. Exa Fast’s 0.733 also exceeds those competitors’ default-tier averages in the company’s comparison. These figures summarize the study’s chosen tasks and methods; they do not establish universal superiority.
The task-level results show why the overall average needs context. On the six-dataset BEIR benchmark and 18-language MIRACL relevance benchmark, the company says leading systems are within roughly one or two percentage points. Exa Auto posts 0.840 on BEIR and 0.887 on MIRACL; Voyage and Google score higher on MIRACL, while Cohere is higher on BEIR. On the internal coding set of 317 long-context queries, Exa Auto scores 0.685 and Exa Fast 0.658, below Voyage at 0.771 and Cohere at 0.744 for their default models. A broad average can conceal these differences by task.
Exa reports stronger relative results on its freshness, navigational and adversarial-authority tests. The freshness set contains 1,000 multilingual synthetic queries, each paired with 60 current or stale candidates; Fast scores 0.797 nDCG@10 and Auto 0.738. The navigation set has 639 queries across company pages, executive lookups and general destination searches; results are drawn from Exa’s live index. The authority test uses 1,000 queries and 60 candidates with prompt-injection text, keyword stuffing, Unicode payloads and false authority claims. Exa says its rerankers were more robust in these company-designed settings.
The methodology also sets limits on what the results mean. Exa says coding labels were assigned by GPT-5.4, while navigation labels also used that model; freshness and authority data were synthetic. The blog’s appendix describes a stratified bootstrap with 10,000 resamples for confidence intervals and paired comparisons. Internal task design, model-based labels, candidate selection and the use of Exa’s own index all affect results. Although the company says it compares identical candidate lists for each reranker within a benchmark, this does not make the benchmark independently audited or prove performance on every search workload.
The post says Exa does not currently plan to release a standalone reranking API; it presents the comparisons as an evaluation of the component powering Exa search. The paper is relevant to energy and resource information services because current prices, regulations, technical documentation and primary-source notices can become stale or be manipulated. Yet benchmark scores cannot verify a company claim, identify a real well or mine, or replace editorial source checks. Exa’s announcement supplies no basis here for claims about model weights, licensing, deployment guarantees or pricing.