Agentic search

How does agentic search work?

Mistral, which announced a product with this name on August 20, 2026, describes a loop of searching, opening, navigating and reading sources, in contrast to answering from a single list of top results. Firecrawl defines it as retrieval in which the agent formulates queries from context, reads the actual page content of each result and then decides whether to answer or search again. Anthropic describes a research system with a lead agent that splits a question into subtasks and lets several subagents search in parallel and report back.

Classic retrieval-augmented generation (RAG) usually fetches matching passages once, often from a prepared index. Agentic search adapts its queries to the intermediate results and can work on the live web. Google describes a related technique for AI Overviews and AI Mode, query fan-out, in which several related searches run at the same time.

What does it mean for SEO?

An agent opens pages while it answers, often several in sequence. Pages it cannot reach or read are unlikely to feed into the answer (assessment). For Google’s AI features, the documentation still requires that a page is indexed and eligible for snippets.

Example

A shopping agent looks for project management software for 50 people with a DATEV interface. It searches for providers, opens price and integration pages, finds that two providers lack the interface and searches specifically for alternatives. The result is a shortlist with source links.

Common questions

How does agentic search differ from RAG? RAG typically retrieves passages once. Agentic search repeats searching and reading and refines its queries.

Is agentic search the same as an AI agent? An AI agent can do many kinds of tasks. Agentic search is its use for research.

Last updated: October 8, 2026