Tool use (function calling): definition and why it matters for AI visibility

A language model on its own can only produce text from what it learned in training. Tool use lets it reach outside that. The developer describes a set of functions (search the web, fetch a URL, query a database), and the model can respond with a structured request to call one instead of a final answer.

The model does not run the function itself. The application does, then hands the output back. "Function calling" and "LLM tools" are different names for the same pattern. For the search-specific version, see the entry on the web search tool.

A typical loop has four steps:

1. The model receives the user's question plus a list of available tools.
2. It decides whether a tool would help. If yes, it emits a call with arguments, such as a search query.
3. The application runs the call and returns results, which land in the model's context window.
4. The model writes its answer from that context, sometimes calling more tools first.

The query the model sends is often not the user's words. OpenAI's help page says ChatGPT search typically rewrites a question into one or more targeted queries. Google's Gemini API docs say grounding automatically generates one or more search queries and runs them.

Who decides whether search runs differs by surface, as of October 2026. ChatGPT search is model-discretionary. On Claude, the model decides and the integrator's system prompt steers it, and Claude.ai search is a user toggle. Perplexity's Agent API requires developers to opt web search in, per its migration docs. Anthropic's web search tool documentation shows what a result looks like: a URL, a title, a page age and short cited text spans.

Every AI answer about your brand falls into one of two cases.

No tool call: memory only

The model answers from training data. That knowledge stops at a cutoff. OpenAI lists a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for the chat-latest Instant path. A product launched after the cutoff does not exist in that answer, and old pricing stays old. Nothing you publish now changes it until a future model is trained.

Tool call made: retrieval

Now your site can enter the answer, but only if it passes access checks first. For ChatGPT search, OpenAI says a site must allow OAI-SearchBot in robots.txt and must let traffic from its published IP ranges through the host or CDN. Content that only appears after JavaScript runs may be invisible to most AI crawlers. Even after retrieval, which results get used depends on factors like position bias that no publisher controls.

This split also affects measurement. Any single visibility number averages answers given with and without search. It's worth asking whether a tool tests the same models and settings people use.

• "AI assistants always search." They don't. Search fires per query, and on several surfaces the model chooses. Perplexity code left on Sonar past its September 27, 2026 retirement can return ungrounded answers without any error.
• "A cited source is the source that shaped the answer." OpenAI's API returns retrieved sources separately from citations, and the retrieved set is often larger. Citation and influence are different things, which is one reason citations sometimes don't support the claim.
• "You can write content that makes the model call search." No published evidence shows a publisher lever on that decision.
• "Allowing GPTBot gets you into ChatGPT search." GPTBot governs training only. OAI-SearchBot governs search, and OpenAI says each setting is independent.

• Web search tool: the most common tool for AI answers, covered in depth.
• Context window: where tool results go before the model reads them.
• Position bias: why order inside the retrieved set changes what gets cited.
• Free AI visibility tracking: what a one-off snapshot can and cannot tell you about answers with and without search.