Quite good, per Bellingcat [0] — Google Lens and ChatGPT could localize the majority of their test photos pretty specifically.

[0] https://www.bellingcat.com/resources/2025/08/14/llms-vs-geol...

Pretty good. I played a bit with gpt-4 a year or so ago by feeding it random screenshots from Google street view. It will pick up a lot of subtle hints from what otherwise looks like generic streets. I imagine more recent models might be better at this now.

Pretty good. I test it every now and then from random photos. Sometimes spot on, sometimes gets very close, unless it's really ambiguous.

Basically all up to the training data, as things often are.