Web Search as a Utility Layer
Product manager at Cohere on enterprise AI search infrastructure and deep research agents
Cohere’s use of Tavily shows that for an enterprise AI product, web search is usually a utility layer, not the product itself. North’s core job is answering questions over a customer’s own internal documents in secure environments, so Cohere picked the vendor that made web grounding easiest to plug in, returned usable page text instead of just links, and let each enterprise customer pay for its own search usage directly.
-
The practical reason for choosing Tavily was workflow fit. Cohere had previously used Brave, but that required extra steps to turn search results into the raw website text an LLM can actually read. Tavily bundled more of that retrieval and grounding work into the API, with similar quality and easier integration.
-
This also shows the split in the market. Tavily is described as a narrower search and grounding API. Parallel is framed as broader deep research infrastructure for agents, the engine underneath long 10 to 15 minute reports in Manus. Exa sits closer to search infrastructure, with a large semantic index and domain specific streams like financial data.
-
The economics favor buying this layer instead of building it. North is deployed inside customer environments, so each customer brings its own Tavily API key and pays the search bill directly. For Cohere, that turns web search into a low drama dependency while the company focuses engineering effort on internal connectors, retrieval, and enterprise deployment.
Going forward, the winning search vendors will be the ones that move from generic web results toward domain specific research infrastructure. Generic web search is good enough to be outsourced. The next wedge is better access to high value sources like filings, medical journals, and proprietary enterprise systems, where better grounding can materially improve agent outputs.