How Does AI Geolocation Work?

The single most useful thing to understand is what these models are not doing. There is no database of photographs being consulted, and no match being found. Nothing about your image is looked up anywhere.

• Text and script — a single sign plus a writing system often narrows to one country.
• Road furniture — line colours, bollards, signposts and kerb markings are standardised nationally.
• Utilities — pole shape, wire arrangement and postbox colour change at borders where landscape does not.
• Vehicles — licence plate proportions and colour, and which models are common locally.
• Architecture — roof pitch, window proportions and materials track climate and building tradition.
• Vegetation and light — species mix gives a climate band; shadow angle constrains latitude.

None of these is decisive alone, and a human reading the same clues in order works exactly this way. A GPS tag would place the camera within about 5 m; visual inference works at a resolution thousands of times coarser, and that gap is the whole story. The estimate is an intersection: each signal removes possibilities, and the answer is whatever survives the most of them. That is also why a photo with few visible signals produces a weak answer rather than a confidently wrong one.

They complement each other rather than compete. The sensible order is search first — exact beats approximate — and infer only when nothing matches. The engine comparison covers which to try for which subject.

This is a limit of the method, not of model size. A larger model gets the city right more often; it does not turn a photograph of a generic residential street into a house number. Any tool presenting street-level precision from pixels alone is presenting a guess as a measurement.

The failure mode worth watching for is a confident answer on a featureless photograph. A model that names a specific city from a picture of a blank wall has not found something you missed — it has produced noise with a number attached.