pulling frames and sending them to GPT-4V seems like it would already work, as long as you made the prompts reasonable.

It depends on how exact you want it - do you need the exact URL/community/etc? Or do you just need something like {cars: 40%, relationship questions: 30%, etc.}?

Seems very doable. In fact, even the list of web domains you've been browsing probably says quite a lot about who you are (no AI required).