Web Fetch
Web fetch refers to the ability of an AI assistant to access a specified internet address itself and read the content of the page. This allows the program to obtain current information that it cannot know from its training material.
A chat program like ChatGPT has learned its knowledge from huge amounts of text collected months or years ago. What was published yesterday is not included there. Web fetch closes this gap: the program accesses a specific internet address, downloads the text of the page, and reads it before responding. You can imagine this like an employee who, on request, opens a page in the browser and reads it aloud. The name comes from the English word “to fetch,” meaning “to retrieve.” Unlike a search on the internet, web fetch retrieves exactly one address that is already known.
Why models without page retrieval often become outdated
The biggest weakness of language models is their fixed cutoff date. Everything that happened afterward is missing. If you ask about a company’s current stock price, the program can only guess or state an old figure. With a page fetch, it instead retrieves the figure directly from the stock exchange website.
A second advantage is verifiability. If the program has actually read a source, it can cite the address. You can then check for yourself whether the statement is correct. Without retrieval, fabricated information — so-called hallucinations — arise more easily: plausible-sounding sentences without any real basis.
Third, fetching saves effort for the operators. Retraining a model costs millions and takes weeks. Fetching a page costs a fraction of a cent and takes a second. For current information, fetching is therefore almost always the better approach.
From link to readable text
Technically, the fetch process runs in several steps. First, the model recognizes that it needs a page and states the desired address. Then a helper program outside the model takes over and sends a request to the server hosting the page. What comes back is the page’s source code, i.e., the text along with all the instructions for its display.
This source code is poorly usable for a language model. It contains ads, menus, cookie notices, and a lot of technical clutter. It is therefore filtered and converted into plain text, usually in a lean format called Markdown. Only this cleaned-up text enters the conversation, and the model responds based on it.
There are plenty of limitations. Very long pages don’t fit entirely into the conversation’s working memory and get shortened. Some content sits behind a paywall or a login and remains invisible. And pages that only load their content within the browser often deliver just an empty shell on a simple fetch. Also important is the distinction from web search: search finds addresses, fetch reads them. In practice, the two are usually combined.
Where you notice this process in products
You see it most clearly when you paste a link into a chat window and ask for a summary. That’s exactly when the fetch kicks in. The same thing happens when a response contains numbered source citations that you can click on. Browser extensions that shorten news articles work on the same principle.
In technical texts and job postings, fetching appears as part of “tool use,” i.e., a model’s ability to operate external tools. For so-called AI agents that carry out multiple steps independently, it is part of the standard toolkit. An agent tasked with comparing prices will access several shop pages one after another for this purpose.
A risk-related aspect is also frequently discussed. A fetched page may contain hidden instructions that the model mistakes for a command. Experts call this an indirect prompt injection. Those who build such systems therefore treat fetched content as data and never as instructions.