Image similarity models exploit features which a human eye do not directly model. The outcome is the same, but the path and information model isn't. (At least, I do not believe I perform a jpeg like DCT and hash on reduced bits from a domain and phase shift, but I dont know how the brain works)

Why should computer driven geoloc be any different? Clearly there is an information model which delivers geospatial likeness, on what basis I do not know, but it's not likely to be how humans do it. Image similarity may be part of it, but other features may also denote likeness. Time of day, shadow direction, features, scale, I have no idea.

Since the scan carries other metadata it's more likely from statistical models of information inherent in the image, not associations outside the image but I don't discount other metastate like accurately geoloc images known to be in the same dataset, or similar refinement of location to a circle of error around the real location which can then be refined from known good data.

If the image source is predominantly British imagery, then matches at Angkor Wat are less likely. You could narrow down the search space significantly.