I've given it a screenshot and it got close enough. So, it's not looking at exif data.

I've been playing a bit with just giving it random things from Google's streetview.

It figured out a landmark in Dresden pretty easily. But then I gave it a generic view of some street in Nuremberg and it took a long time. It extracted lots of information from the screenshot. Including the logo of a travel agency and some very pixelated house numbers. Amusingly I actually triggered some AB testing thing where it was using two models in parallel side by side.

I just let it do its thing and it looked like it got stuck in some loop. Eventually both models produced an answer. They both correctly identified the city but then both models came up with two wrong streets. The travel agency has offices across Germany and probably multiple in Nuremberg as well. There was no street name readable and no house numbers. It based this off just the name of the travel agency and the fact it was next to some hairdresser.

It seems to be using python scripts to zoom in on different bits of the image and search engines to query clues it finds. It also looks at architecture and other aspects to deduce things about country, regions, etc. Overall, I'm pretty impressed with how far it got.