Can ChatGPT scrape a website? What it can and cannot pull from a page

Short answer: partly, and not dependably. A chatbot with browsing can sometimes open a public page and put a few of its items into a table. It often cannot open the page you need, and when it can, a list of 80 items tends to come back as ten. You get a table in a chat window, not a spreadsheet file.

What a chatbot does well

If the data is already in front of it, a chatbot is good at giving it shape. Copy a messy list from a page, paste it into ChatGPT, Claude or Gemini and ask for a table with name, price and rating, and you will usually get a clean one. For twenty rows you can paste by hand, that is all you need.

Chatbots are also good at the work around scraping: explaining how a page is built, or writing a script for you if you are willing to run and maintain code yourself.

Where it goes wrong

The trouble starts when you give it an address and ask for everything on the page.

• It cannot open the page. Many sites turn away requests from AI assistants. Some assistants say so. Others answer from memory or from a search result, which is not the page you asked about.
• The page arrives empty. Many shops and listing sites build their content with JavaScript after the page loads. An assistant that fetches the raw page can see a shell with no products in it.
• You get a sample. Asked for all the listings, a chatbot tends to give the first few and then "and so on", or a summary of the rest.
• Rows are invented. To make the table look complete, it may add items or fill in prices that are plausible and not on the page. Nothing marks them as guesses.
• No file. The result is text in a chat. Getting it into Excel is copy and paste, and long tables are often cut off partway.

The fourth one is the dangerous one. A table that is short is annoying. A table with a few made-up rows is worse, because you cannot tell which rows to trust.

Why this happens

A chat assistant is built to answer a question in a reasonable amount of text. Reading a website is a side step it takes on the way, with a simple fetcher, not a full browser, and with a limit on how much of the page it keeps.

Writing out 80 rows exactly is also not what it is tuned for. Its habit is to be helpful and brief, so it compresses. And when part of the page is missing, a fluent answer is easier to produce than an honest gap.

How to test it yourself

Pick a page where you can count the items, such as a category page with a visible "48 results".

1. Give the chatbot the address and ask for every item as a table, with name, price and link.
2. Count the rows it returns.
3. Check five rows against the page: the name, the price and whether the link opens that item.

If the count matches and the five rows are right, the chatbot handled that page. If rows are missing, prices differ or links lead nowhere, it did not, and the rest of the table is in doubt too.

What to use instead

Use a tool that opens the page in a real browser and is set up to return every item. Our web scraper takes a page address, renders it in Chrome so JavaScript content is there, and gives you the repeated items as a table with one row each, including each item's URL. You can add a plain-English instruction such as "name, price and rating of each product". There is no code to write and no extension to install.

The first rows are free and need no account, so you can run the same counting test on your own page. The full table as Excel or CSV is $5, paid once.

It has limits worth knowing. It reads the page you give it and does not click through to the next page, so for several result pages you paste each page's address. Pages behind a login cannot be opened, some sites block automated visits, and you are responsible for respecting the terms of the site you collect from.

If what you want is a copy of the page itself, not its data, see webpage to PDF.