Web search for agents is here: Why ranking still decides the answer

AI agents have struggled with finding current information for a while. When they need up-to-date details but lack a search tool, they often assume the correct web address and end up with "page not found" errors. This guessing habit leads to outdated or incorrect information being used. To address this issue, Cloudflare introduced a beta Web Search API in its AI Gateway.

This allows agents to access live search results from multiple providers through a single interface, complete with source links, logging, and access controls. However, even with live search, another important challenge remains: determining which of the fresh pages and internal documents the agent should rely on. The agent must prioritize this information based on relevance, quality, and authority.

This decision-making process occurs in the retrieval layer, where two distinct pools of evidence are combined - the web and the company's internal data. Each of these pools has its own characteristics, such as freshness, quality, structure, and authority. For example, web results may be hours or days old, while internal data can be months out of date without anyone noticing.

The agent must reconcile these differences and make ranking and filtering decisions based on company-specific data and business rules.