The Photo Types People Love Testing AI Geolocation With

There is a specific kind of fun in handing a machine a photo and watching it try. You are not lost, you are not trying to find anyone, and nothing at all depends on the answer being right. The appeal is watching a model pick a scene apart the way a sharp-eyed friend would: the roofline, the plug socket, the shape of the letters on a distant sign. The awkward part is deciding what to upload first, because the photographs that impress a person are rarely the ones that make an interesting test.

A photograph of the Colosseum is not really a test, it is a lookup. The building is one of the most reproduced images on the internet and any competent vision model will name it instantly, which is satisfying for about four seconds. At the other extreme, a close-up of a white wall or a plate of pasta cropped to the rim gives the model nothing but texture, and a guess made from nothing is just a coin toss dressed up in confident language.

The photos worth uploading sit in between: enough visual information to reason from, no single element that gives the game away. That is the same principle behind the browser game GeoGuessr, launched in 2013, which drops players into an anonymous stretch of street view and asks them to work outwards from kerb stones and bollards. If you have not run a photo through Raven before, the step-by-step walkthrough of using Raven covers the mechanics; this piece is about what to feed it once you know how.

Almost everyone has one: a faded print from a shoebox, a house or a street corner from a holiday decades ago, with nobody left to ask exactly where it was. These are a perennial favourite for good reason. Old prints tend to be rich in period detail that nobody was trying to photograph — a car model that dates the frame to within a few years, a specific streetlamp casting, a shopfront sign painted by hand, a postbox in a shape used in only one country. An AI guess can turn into a genuine moment of recognition, filling in a small gap in a family story nobody thought to write down. If the album is still on paper, our guide to digitising old travel albums covers how to scan prints so the background survives being zoomed into.

This makes food photography a surprisingly good game to play with friends, because the answer feels like it should be in the dish and never is. Point the conversation at the edges of the frame instead. Restaurant furniture is bought locally and follows regional taste; napkin folding, condiment bottles, the way cutlery is laid out and even the standard height of a table vary more than most travellers register. Half the entertainment is discovering how much of a meal photograph is actually a photograph of a room.

Snapped through glass, slightly blurred, usually just a stretch of tarmac, some fields or hills, and perhaps a sign too far away to read. These look like they would be impossible to place, which is exactly why they are worth trying. Road markings are standardised nationally rather than globally, so the colour of the centre line, the rhythm of the dashes, the style of the crash barrier and the design of the reflective posts along the verge all narrow the field before a single word of text is visible. Add the colour of the soil and the shape of the roadside vegetation and a surprisingly specific regional guess becomes possible without a landmark anywhere in the frame.

This has become something of a running joke among people who play with these tools. Chain hotels spend real money making a room in one city feel identical to a room in another, so a photo of beige curtains and a nondescript bedspread is a fair test of whether a model will admit uncertainty. When it does find something, it is usually the infrastructure rather than the decor: mains sockets follow one of about fifteen mutually incompatible national designs, and light switches, radiator styles and window fittings cluster the same way. Half the fun is watching how honestly a tool handles a scene engineered to give nothing away.

• —Postcards and old prints. Especially ones with visible film grain or a printed date stamp in the corner.
• —Paused frames from home video. Camcorder footage often carries a burned-in timestamp, and the low resolution forces the model to reason from shapes rather than text.
• —Airports and railway stations. Signage typography, seating design and platform furniture vary far more by country than travellers expect.
• —Balconies and back gardens. No landmark in sight, just plants, roof tiles and sky — a clean test of how much climate and light alone can reveal.
• —Photos you already know the answer to. The only way to build a feel for when a guess is trustworthy is to check it against places you can verify.

Interior shots of plain rooms, macro photographs of objects and anything cropped tightly around a face fall into the same category. So do photographs of other photographs taken at an angle on a phone, where glare wipes out exactly the fine detail worth reading. If you find yourself disappointed by a vague answer, it is usually the input rather than the model: the frame simply did not contain enough of the world. A quick way to build intuition is to try the same scene twice, once cropped tight and once wide, and watch the answer change.

Raven is deliberately simple: drop in whatever you are curious about, and Google's Gemini model reads the vernacular architecture, vegetation, road markings, signage and light to produce a best guess. Nothing is written to disk or to a database — the image exists in memory long enough to generate an answer and is then discarded, which is worth knowing before you upload anything from a private album. The result is an entertainment-only estimate and it can be confidently wrong, which is part of why comparing several photos of the same place is more interesting than trusting any single answer. If most of your candidates live on your phone, rediscovering forgotten trips in your camera roll is a good way to find them, and a photographer's guide to geotagging explains why the GPS tag many of those files already carry is a separate question entirely.

None of this is about proving anything or finding anyone. It is a genuinely enjoyable way to look harder at photographs you already own, and a reminder of how much information sits quietly in the background of an ordinary picture. If you want the argument for why that distinction matters, why entertainment-only matters sets out the line between a curiosity tool and a surveillance one. Otherwise, start with whichever photo you are already curious about. That is almost always the right one to try first.