GPT doesn’t experience the web like humans do. It doesn’t load full pages or interact with HTML and layouts. Instead, it receives structured snippets during web searches, which include a title, URL, and a short text excerpt, along with some optional metadata. This means GPT is limited in its ability to retrieve full content or navigate a site’s structure. Each snippet acts like a compact search result card, giving just a snapshot of the information available. When GPT wants to explore more content from a snippet, it uses the `open()` and `click()` functions. The `open()` function retrieves a larger excerpt from the same page, allowing GPT to "scroll" through content, while `click()` follows an outgoing link to fetch another snippet. There’s no limit to how many times it can call these functions, but practical constraints mean GPT can’t access full pages. Each expansion is capped, and even if it sees multiple windows, it has to summarize rather than reproduce long passages. Developers can adjust context size settings for their assistants, which determines how much content GPT receives with each snippet. These settings range from low, which provides small excerpts, to high, which allows for larger slices of text. Testing showed that even with higher context sizes and repeated expansions, GPT only retrieves parts of a page at a time. It can access sections, headings, and examples, but never the whole article. This windowed approach limits the depth of understanding GPT can gain from a webpage.