AI Geolocation Accuracy: An Honest Answer

Two photographs go into the same model within a minute of each other. The first is a wide street scene: tiled facades, overhead tram wires, a shopfront lettered in Portuguese, kerbstones cut in a pattern used in very few places. It comes back correct to the city. The second is a beige hotel corridor with a fire door and a patterned carpet. It comes back with a hedge and a shrug. Nothing about the software changed between the two uploads.

That is the shape of the whole subject in one example. Accuracy here is not a fixed property of a tool, the way a kitchen scale is accurate to within a gram. It is a property of the pairing between a picture and a guesser, and the picture carries most of the weight.

Ask how accurate one of these tools is and the only honest reply is a question back: accurate to within what? Every serious claim needs a radius attached to it. The convention borrowed from mapping is great-circle distance, the shortest path over the surface of the globe between the guessed point and the real one. Once that distance exists, a guess can be scored against a threshold. Inside 25 km. Inside 200 km. Right country. Right continent. Those are four different tests, and a system can pass one comfortably while failing the next.

This is why casual comparisons fall apart so quickly. One person means it named Italy and the answer was Italy. Another means it put the pin on the right piazza. Both are describing the same output and reaching opposite verdicts. Before any figure is worth arguing about, the radius has to be on the table.

Country level is what visual reasoning is genuinely built for. Scripts on signage, plate proportions, driving side, socket shapes visible indoors, plant species, road markings, the profile of a bollard: these vary by nation because nations legislate them. A frame holding two or three such clues narrows the field hard, and a model that reads them carefully lands the country a decent share of the time.

Region level is harder, and how much harder depends on the country. Separating the Scottish Highlands from Kent is not difficult. Separating one prairie province from the one beside it is close to impossible, because the buildings, the road furniture and the vegetation genuinely do not differ. The same argument runs across borders as well as within them, which is the ground covered in telling neighbouring countries apart.

• —Continent or hemisphere. Reachable from light, vegetation and general building style alone. The weakest useful result, and the one a thin photograph usually gets.
• —Country. The natural resolution of visual evidence, because so much of what appears in a frame is set by national standards.
• —Region or city. Available when a place has a distinctive local vernacular — a roof type, a tram, a stone colour — and unavailable when it does not.
• —Street or address. Not a description problem at all. Reaching this level needs a matching index, not a better description.

The gap between this looks like a Lisbon street and this is the corner of that particular street cannot be crossed by description. Crossing it means comparing the frame against a catalogue of pictures whose locations are already recorded — a lookup rather than an act of reading, with a ceiling set by whatever happens to sit in that catalogue. The trade-off is weighed up in reverse image search and AI geolocation. Description narrows. Matching pins. A vision model does the first and does not attempt the second.

It helps to see how fine street level really is. Coordinates written in decimal degrees resolve to roughly 11 metres at the fourth decimal place near the equator, while one degree of latitude covers about 111 km. A phone works at the fine end of that scale because it is timing signals from GPS satellites, and has done so at metre scale since the deliberate degradation of civilian precision was switched off in 2000. A model looking at a photograph is working at the coarse end, from evidence that was never designed to be unique to one spot.

A figure such as eighty-five per cent accurate tells you almost nothing without two further facts: the radius it was scored against, and the set of photographs it was scored on. Change either and the number moves enormously. Score a system on well-lit outdoor street scenes and it looks superb. Score the same system on a random slice of a real camera roll — meals, faces, ceilings, screenshots, a close-up of a dog — and the figure falls through the floor, because most photographs people take contain no geographic evidence at all.

There is a second trap hidden in the benchmarks themselves. A test set built from photographs that already carry coordinates is skewed towards the kind of picture people bother to geotag, which is disproportionately travel, landmarks and outdoor scenes. Accuracy measured on that set is real, but it describes a curated slice of the world rather than the folder sitting on your phone. Any number you read should be assumed to come from the flattering half.

The useful mental model is cross-checking rather than recognition. One clue suggests; several clues that point the same way convince. Cyrillic lettering narrows the field to a set of countries. Add right-hand traffic, a particular utility pole design and a birch treeline, and the set shrinks quickly. Take those away one at a time and the answer widens back out. That is why a featureless interior is so hard, and why the difference between a good upload and a poor one is usually larger than the difference between two competing tools.

It follows that the biggest lever an ordinary person has is not the choice of tool but the choice of frame. A wider shot, a second photograph from the same spot, or simply a picture taken outdoors will move the result more than any amount of switching between services. Which pictures work and which do not is set out in the best and worst photos to upload.

An honest low-confidence answer carries real information. It says the frame is thin, or the clues disagree, or the scene type occurs in a hundred places. That is a true finding about the photograph. A confidently wrong answer carries the opposite, because it invites you to stop looking. Reading the reasoning matters far more than reading the figure attached to it, and how to read that figure is a separate subject in itself, covered in how much to trust a confidence score.

Ambiguity is frequently the correct answer as well. Plenty of photographs genuinely could have been taken in four places, and a system that names all four is more accurate, in the sense that matters, than one that quietly picks a favourite. What happens when a photo has several plausible answers is worth reading alongside this, because it explains why a spread of options is a feature rather than a fudge.

Published figures are worth very little. A personal one takes an afternoon and is worth a great deal more, because it is measured on the photographs you actually own rather than on somebody else's showcase. Gather 20 images whose true location you already know, deliberately spread across scene types instead of picked from your best travel shots, and run them through whatever you are judging.

1. 1.Decide the radius before you start, and write it down. Right country is a different experiment from within 50 km.
2. 2.Include the boring photographs. Kitchens, car parks, corridors and close-ups are most of a real camera roll and belong in a fair test.
3. 3.Record the scene type beside each result, so the pattern of failures is visible rather than buried in an average.
4. 4.Count hedged answers in a separate column instead of marking them wrong. A refusal to guess is a different outcome from a bad guess.
5. 5.Run the same 20 images through a second tool. Any comparison made on different photographs tells you nothing at all.

The short version is that accuracy questions are really photograph questions. Given a rich outdoor scene, a good model is impressive and worth taking seriously at the level of a country. Given a plain interior, it is guessing, and the fair thing for it to do is say so plainly. Raven is built to answer that second case honestly rather than to protect a statistic, which is the trade any entertainment tool ought to make.

To see the confidence figure in context on a real photograph, try the AI photo location finder and read the result alongside this article.