Relevance Not Scaling Drives Exa
Will Bryk, CEO of Exa, on building search for AI agents
This reveals that Exa sees its real moat as relevance, not servers. In practice, adding more machines can let a search API answer more queries, but it does not solve the harder problem of finding the right pages for vague, long, agent-style prompts, extracting the useful parts, and keeping results fresh enough that downstream agents can trust them. That is why Exa customers keep focusing on coverage, recall, extraction quality, and precision, not raw throughput.
-
In one production workflow, Exa powered 5,000 daily prompts and returned 50,000 to 100,000 results per day. The bottleneck was not rate limits, which were manageable, but whether the system could keep finding new, relevant pages, return full text, and avoid stale or duplicate results.
-
The product challenge is easiest to see in edge cases. Customers call out vague semantic queries, very deep pagination, paywalled pages, JavaScript-heavy sites, and wrong published dates as the hard parts. Those are ranking, extraction, and freshness problems, which means they are search quality problems, not infrastructure procurement problems.
-
Competitors show the same split. Parallel is stronger for multi-step research summaries, while Tavily is lighter weight and asset-light. Exa is differentiated when a customer needs a large independent index, many raw results, and full content for downstream processing. That is closer to building a better retrieval engine than simply wrapping search with an agent.
Going forward, the winners in AI search infrastructure will be the companies that make retrieval reliable enough to disappear into agent workflows. As model providers bundle basic web search, independent players like Exa will keep moving up on the dimensions that are hardest to commoditize, better recall on messy real-world queries, cleaner extraction, fresher indexes, and outputs that agents can act on without human cleanup.