Can you use AI to read links and web pages?

Short answer: it depends on whether ChatGPT searches the web in that chat. OpenAI3 says ChatGPT does not search the web unless specific tools are enabled. Those tools are on by default, and ChatGPT may search automatically when a question would benefit from current information. If it did not search, the reply is based on what the model learned in training, however confident it sounds.

What ChatGPT actually does when you paste a link

Most of the arguing on forums comes from treating a pasted link like a link in a browser. To ChatGPT, a link is text. If it does not search, it never goes to the page. It sees an address, and an address often gives a lot away: the site name, and a slug such as “best-budget-laptops”. A language model is very good at turning those clues into a fluent summary.

That is the risk. OpenAI’s own help page says ChatGPT can produce incorrect or misleading answers and may sound confident even when it is wrong. A summary of a page it never opened reads just like one of a page it did.

When it does use search, it can visit the page. OpenAI’s crawler documentation says that when a user asks ChatGPT a question, it may visit a web page using an agent called ChatGPT-User.

When it can open a link, and when it can’t

Keep three cases apart:

1. No search. The link is just text. Anything ChatGPT says about it comes from the address and from its training.
2. Search used. ChatGPT can go and look, and answers that use web search may include citations you can open.
3. Links ChatGPT writes itself. A different problem, covered below.

Even with search, it will not always get in. OpenAI says the model may not be able to get information from a website because of technical issues, paywalls or the site’s robots.txt settings. Its help pages say nothing specific about pages behind a login, so do not assume it can read your account pages, intranet or anything you have to sign in to see.

Web search is available on all ChatGPT plans, including Free, and people who are not signed in can use it too. Searches count towards your plan’s usage limits.

How to tell whether it really read the page

OpenAI suggests two checks: see whether the search or deep research icon is active in the input bar, and see whether the response cites sources from the web. No citations is a reason for suspicion, though it proves nothing on its own.

You can also test it yourself in a couple of minutes:

1. Pick a page you know, with a detail nobody could guess from the address, such as an odd sentence halfway down or a line in the footer.
2. Paste the link and ask ChatGPT to quote that exact detail.
3. If it quotes it correctly, it almost certainly read the page. If it gives you something generic, or a plausible line that is not there, it did not.

If it answered without reading the page, OpenAI says you can regenerate the response and choose Search the web, where that option is offered.

Why links ChatGPT gives you often don’t work

This is the third case. You ask for sources, get a tidy list of links, and some go nowhere. OpenAI lists “fabricated quotes, studies, citations or references to non-existent sources” among the mistakes ChatGPT can make. A model that knows what addresses look like can write one with the right domain and a believable path for a page that does not exist.

OpenAI publishes no figure for how often this happens, and we have not found one. Treat a link ChatGPT writes from memory as you would a reference you have not opened: click it. If it is dead or lands somewhere unrelated, search for the title or the organisation instead. Citations from a search can be wrong too. OpenAI says search results and citations can be incomplete, outdated or incorrect, and advises opening the cited source to check it supports the answer.

How to do it well

• Ask for search. In a conversation, select View all tools, then Search, or type / in the message box and choose Search. Then paste the link and say exactly what you want: a summary, the key dates, the main argument in three sentences.
• Check before you rely on it. Look for citations, open them, and run the quote test above on anything that matters.
• If search is missing, OpenAI suggests starting a new conversation, updating the app or trying chatgpt.com, checking you are on the right account and under your usage limit, and, in a work account, asking your administrator whether web search is enabled.
• If it can’t get in, give it the content. Copy the relevant part of the page into the chat, or upload the document. ChatGPT accepts DOCX, PDF and TXT files, spreadsheets and PowerPoint files, within plan limits. The guide to summarising a PDF covers that route.
• When you ask for sources, ask it to point. Ask for the claim and the publisher’s name, then find the page yourself.

For coursework, the same checks apply, along with your institution’s rules. See using AI to help you study and using AI for homework.

Rules and risks

Accuracy. OpenAI tells users to approach ChatGPT critically and to check important information against reliable sources. That matters most for a page it may never have opened.

Privacy. A link gives ChatGPT an address. Pasting text or uploading a file hands over the content itself. On consumer plans, whether that content helps train OpenAI’s models depends on the Improve the model for everyone setting in Data controls, and OpenAI says it does not use content from business offerings such as ChatGPT Enterprise or the API for training. Uploaded files saved in your Library can be kept separately from the chat, so deleting the chat does not delete them. For who can see your chats and how to change the settings, see who can see your ChatGPT conversations. Before you paste anything, think twice about contracts, medical details and other people’s personal information.

If you run a website. OpenAI publishes the names of its crawlers and what each one does:

• OAI-SearchBot finds sites for ChatGPT’s search features. If you block it in robots.txt, your site will not be shown in ChatGPT search answers, though it can still appear as a navigational link. OpenAI says changes can take about 24 hours to take effect.
• GPTBot collects content that may be used to train OpenAI’s models. Blocking it signals that your content should not be used for training. The two settings are independent.
• ChatGPT-User visits a page when a user’s request calls for it. OpenAI says it does not crawl the web automatically and that, because a user starts these visits, robots.txt rules may not apply.

So robots.txt controls search listing and training. It does not reliably stop ChatGPT opening a page when a user asks it to.