Search Infrastructure Over Models

This reveals that Exa is positioning retrieval as the scarce infrastructure layer, while models are becoming interchangeable. In practice, the hard part is not generating a nice answer at the end, it is finding the right pages, extracting usable text, ranking results for vague queries, and doing it fast enough and cheaply enough for production agents. That is why customers often plug Exa into their own model stack instead of buying a bundled answer engine.

• For teams running real workflows, retrieval quality shows up as recall, depth, and raw text access. One Exa power user runs 5,000 queries a day, pulls 50,000 to 100,000 results, and values Exa mainly for the number of results, full content access, and precision on vague searches, not for final summarization.
  3 sacra
• The market is splitting between search infrastructure and answer products. Tavily is built to return pre ranked text chunks and research outputs without owning a full index, while Exa invests in its own crawling, embeddings, vector search, and high volume retrieval. That makes Exa heavier infrastructure, but also more useful when customers need raw materials instead of a finished narrative.
  1 sacra 4 sacra
• This also explains why Exa can coexist with model labs for now. Enterprise builders like Cohere still pay external vendors for web retrieval because replacing that stack is not core to their product, even when they have strong models themselves. The value is in outsourcing the messy work of crawling, extraction, ranking, and freshness.
  2 sacra

The next step is retrieval vendors moving up one layer without abandoning their core. Exa is already adding research and answer endpoints, but the durable advantage will come from owning the freshest index, the best extraction pipeline, and the fastest path from a messy web query to structured evidence that any model can use.