AI models do not connect to the internet or open a web connection to perform a search.
Instead, the model writes a small, structured note asking the surrounding application to run a search on its behalf.
That structured note is essentially a formatted request, similar to filling in a form rather than typing a query into a browser.
The application, sitting between the model and the internet, receives this request and passes it to a search application programming interface (API), a system that lets separate pieces of software exchange information.
It never touches the API
Once the search runs, the results come back to the application, and the model reads the text it is handed.
The model never touches the API or the web itself; the application handles that part entirely.
This same pattern, write a request, wait for a result, read the output, is not unique to search.
It is the basic structure behind every "tool call" an AI makes, whether that is searching the web, running a calculation, or checking a database.
Understanding this distinction is important for anyone trying to make sense of how AI agents work.
Terms like "function calling" and "tool use" describe exactly this handoff between a model's request and an external system's response.
Separation by design
Large language models (LLMs), the systems behind chatbots such as ChatGPT and Claude, are built on this separation by design.
The model provides the reasoning and the request; the infrastructure around it provides the hands.
That division is easy to miss, because from the outside, an AI reciting search results looks a lot like it went and found them itself.
