What Is Tool Use in LLMs?

Tool use is a large language model's ability to call external capabilities — web search, page browsing, code execution, APIs — during answer generation, incorporating the results into its response. It is the mechanism that turned static chatbots into live answer engines: without tool use, a model can only recite training data; with it, it can look things up.

The toolbox behind AI answers

Production assistants carry a recognizable kit. Search queries an index (OpenAI's engines lean on Bing-backed retrieval; Gemini grounds against Google Search). Browse/fetch retrieves specific pages, announced by user agents like ChatGPT-User or Perplexity-User in your server logs. Code execution runs calculations and data analysis. Connectors reach third-party services, increasingly via the Model Context Protocol, the open standard Anthropic introduced in November 2024 and other vendors adopted in 2025. The model decides per query which tools to invoke — guided by its system prompt and training.

The decision that shapes your visibility

For GEO, the pivotal moment is whether a search tool fires on a given prompt. The split behaves roughly like this:

The same brand question can route differently across engines and even across runs — one reason tracking mentions across engines beats single-engine testing.

Reading tool use from your side

You cannot see an engine's tool calls directly, but you can infer them. Citations in the answer mean retrieval happened. Fetcher user agents in server logs (ChatGPT-User, Claude-User, Perplexity-User) are direct evidence your pages are being pulled at answer time. And a wrong-but-confident answer about something your site states plainly usually means no tool fired — a parametric-memory problem, not a content problem. The glossary entries on function calling and web browsing mode cover the adjacent mechanics.