Search as Infrastructure for Agents
Will Bryk, CEO of Exa, on building search for AI agents
The real opening is not better answers to common searches, it is owning the long tail of messy research tasks that standard search was never built for. Google works best when the user already knows the right keywords. Exa is built for vague, multi constraint, low volume queries where the job is to find the right pages, pull the full text, and feed raw results into an agent or workflow, not just rank links.
-
In practice, these valuable failures are queries like find great beginner React articles, or broad open ended questions that need deeper digging across many weakly signaled pages. One Exa user runs 5,000 such prompts daily, pulls 50,000 to 100,000 results, and says basic SERP APIs can cover only about half of those jobs.
-
The product split in the market is becoming concrete. Exa is strongest when customers need a large set of raw results plus page content for downstream filtering and enrichment. Parallel and similar tools can be better at producing a finished synthesis, but they are weaker when the workflow depends on exhaustive recall and deep pagination.
-
Search engines already route queries this way internally. Ecosia sends only more complex, open ended queries to Exa, while navigational or simple one word searches stay on regular search. That shows where AI native search creates value, not replacing Google for every query, but taking over the expensive, high judgment queries where users would otherwise need many clicks.
The next step is search becoming infrastructure for agents, not a destination for humans. As more software runs research, lead generation, coding lookups, and buying workflows automatically, the highest value layer will be the API that can turn fuzzy intent into reliable web retrieval at scale, with enough depth and structure for downstream models to act on it.