Why can't it just check the URL to see if it's valid and actually contains the data it thinks it contains before recommending it to the user.

Don't modern LLMs like GPT4 have access to load web pages?

The downside is it will slow down the output I guess.