Listen: How GPT Sees the Web

Many people assume that a Generative Pre-trained Transformer, or GPT, browses the web just like humans do, loading full pages, layouts, and images. In reality, the process is heavily restricted. GPT does not actually browse or ingest entire documents. Instead, it relies on a controlled, windowed system.

When GPT first performs a web search, it receives only a tiny snippet of text, usually just one to three sentences. To see more, the model must request specific expansions. It can use an open command to fetch a larger slice of text from a specific line number, which acts like scrolling down the page. It can also use a click command to follow links.

This creates a sliding window pattern. While GPT can request multiple expansions to read further down a page, it can never reconstruct the entire document. There are strict limits on how much text each window can hold and how much information the model can output.

Developers using these search tools in their assistants can choose low, medium, or high context settings. A higher setting provides larger slices of text, but the hard limits remain. Ultimately, GPT never sees raw HTML or full web pages. It only reads sequential, plain-text fragments, piece by piece.