One useful thing you can do with ChatGPT is ask it to analyze a webpage or use a particular page as a source. But there’s a catch: Sometimes ChatGPT can’t access the page. And instead of stopping, it may look elsewhere to cobble together a plausible answer—without telling you.
I learned this the hard way
I once asked ChatGPT to evaluate some text using a specific health literacy tool. I gave it the URL for the scoring instructions, and it came back with numerical scores and detailed explanations.
Everything looked believable, but something still felt off.
When I dug in, I discovered that ChatGPT hadn’t been able to access the scoring rubric at all. Instead, it had created its own scoring system based on what it thought the rubric might look like.
Impressive? Yes. Helpful? Not so much. I had to go back to the beginning and redo the whole thing. And if I hadn’t caught it, I would have delivered an analysis based on fiction to my client.
Why can’t ChatGPT reach everything on the web?
ChatGPT can’t always retrieve the content of a webpage, even when you can open it just fine in your browser. There are several possible reasons:
- The site blocks automated access. Some websites deliberately prevent AI tools, bots, and other automated systems from retrieving their content.
- The content requires a login or subscription. ChatGPT may be able to reach the site but not access what’s behind the login or paywall.
- The page relies heavily on JavaScript or interactive elements. Content in tabs, expandable sections, or other dynamically loaded elements may not be available to ChatGPT.
- The content isn’t presented as ordinary webpage text. Text inside certain viewers, databases, embedded tools, or other interfaces may not be retrievable.
You don’t necessarily need to know why ChatGPT can’t access a particular page. You just need to know that it can happen so you can look out for it.
If left to its own devices, ChatGPT will fill in the gaps
If ChatGPT can’t access a specific webpage, it may still have enough contextual clues to produce a convincing answer. Depending on the situation, it could draw on:
- the URL, page title, or search-result snippets
- other webpages with similar information
- information it already knows about the topic
- patterns that help it predict what the page probably says
If the output were obviously nonsense, this wouldn’t be much of a problem. The risk is an answer that sounds sensible and confident but isn’t actually grounded in the source you asked ChatGPT to analyze. Plausibility is the danger here.
Tell ChatGPT: If you can’t access the page, STOP
If you want to prevent ChatGPT from barreling ahead, give it explicit instructions to stop if it can’t access your source. Put wording like this at the beginning of a longer prompt or use it as a separate prompt to check before you start:
“First, confirm that you can access and read the webpage. If you can’t, stop and tell me why. Do not answer the rest of the prompt.”
That last bit is useful here. If you just ask ChatGPT to confirm, you leave open the possibility that it will say I couldn’t access it, but here’s what I can tell you… and merrily proceed into the kind of speculation you’re trying to avoid.
What you want is for it to simply stop so you can decide what you want to do next.
No access isn’t a dead end
If ChatGPT can’t retrieve the webpage, you don’t have to give up! Often you can give it information about the webpage in another way. For example:
- Copy and paste the page text into ChatGPT. This works especially well for shorter pages or when you only need part of the content.
- Save or print the page as a PDF and upload it. This can help preserve headings, tables, and other page structure.
- Save the webpage as an HTML file and upload it. In most browsers, press Ctrl+S and choose the option to save the HTML.
- ChatGPT may also suggest a workaround itself if it can’t access a certain page.
And depending on your goals, you might ultimately decide to have it find other accessible sources instead.
A simple checkpoint can prevent a lot of trouble
There’s nothing inherently wrong with asking ChatGPT to extrapolate or to work from alternate sources. The important thing is to know when that’s happening, rather than discovering later that the answer wasn’t based on the source you thought it was.
That’s why I now build this checkpoint into any prompt where I’m asking ChatGPT to retrieve material from a webpage. It takes only a few extra words, and it can save you from doing a whole lot of work based on speculation.