Can AI Geolocate a Screenshot Without Exif?

Press two buttons and your phone writes a brand new image file. It holds the pixels that were on the display a moment earlier and nothing else: no camera model, no capture timestamp, no coordinates. That single fact governs everything about screenshots and location guessing, and it catches people out in both directions. Some assume the metadata comes along for the ride. Others assume a file with no metadata is a dead end.

Neither is true. Raven never looks at metadata in the first place, so a screenshot arrives on exactly the same footing as a photograph straight off a sensor. What genuinely shifts the odds is everything else a screenshot does to an image: the tighter crop, the second round of compression, the strip of interface furniture along the top. A little of that helps. Most of it hurts.

A vision model has no notion of provenance. It receives a grid of pixels and looks for the same evidence it always looks for: building materials, window proportions, the script on a shopfront, the species of tree at the kerb, the direction and hardness of the light. Whether that grid came from a camera, a scanner or a frame buffer makes no difference to the reasoning, which is laid out at greater length in how AI guesses where a photo was taken.

So the honest answer is that screenshots are ordinary inputs with a slightly worse starting hand. A screenshot of a wide street scene often does perfectly well. A screenshot of somebody's face against a white wall does not, and would not have done as a photograph either.

The distinction is worth being precise about. Copying a photograph duplicates the whole file, headers included. A screenshot samples the composited display output — the flattened result of whatever windows, overlays and scaling the system happened to be drawing — and hands that to an encoder. The output is typically PNG on iOS and on most Android builds, a format published in 1996 and written here with no camera fields filled in. The Exif block standardised in 1995, with its GPS tags and lens data, is assembled by a camera at the moment of exposure. No exposure happened, so there is nothing to record.

This is the same end state people work hard to reach on purpose when they strip tags before posting, covered step by step in how to remove location data from photos. A screenshot gets there by accident. It also means the privacy question people ask about screenshots is the wrong way round: the file is already stripped, and Raven ignores metadata regardless, as described in what Raven actually sees when you upload.

The scene survives, which is the part that matters. But a screenshot also brings a small amount of incidental furniture that a photograph would not have, and people tend to overrate it. Here is roughly what a model can pick up from the edges of the frame.

• —The status bar. A clock in 24-hour form, a battery glyph, a signal indicator. The time of day can sit alongside the light in the scene as a weak consistency check, nothing more.
• —A carrier name. Where the network operator is shown, it names a country outright — but it names the country of the SIM, not necessarily the country the picture was taken in.
• —Interface language and script. Menus, buttons and system labels in Japanese, Arabic or Finnish say something about the device owner. They say very little about the location in the photograph.
• —Date and unit conventions. A day-month-year date, a comma as the decimal separator, a temperature in Celsius. Broad regional hints that narrow almost nothing on their own.
• —App chrome. The rounded corners of a particular messaging app, a video player's scrub bar, a browser address bar. Mostly useful for working out what kind of image this originally was.

None of these are strong signals, and it is worth saying so plainly. A phone can be bought in one country, carried to another and set to a third language. The status bar tells you about a device; the buildings tell you about a place. When the two disagree, the buildings win.

Most screenshots are crops, whether or not anyone thought of them that way. Someone screenshots the interesting half of a picture, or the part with a face in it, or the section that fitted on screen without scrolling. The discarded portion is usually the top and the sides, and those are the regions carrying the geography. Roof lines vary enormously between countries. Sky colour and shadow length hint at latitude and season. Signage clusters at the edges of a frame, because that is where the shopfronts are.

There is a second, quieter loss: scale. A wide frame lets a model reason about how far apart the buildings sit, how wide the road is, whether the pavement is raised. Crop to a doorway and all of that context goes, leaving a door. This is the same failure pattern described in what makes a photo hard for AI to geolocate, just arrived at by a different route.

Two categories behave unlike anything else. The first is a screenshot of a map. If place names, road numbers or a route line are legible, the answer is written on the image in words, and the model will read it. That is transcription rather than deduction, and it proves nothing about visual reasoning — a map screenshot with the labels cropped off is often harder to place than a photograph of an ordinary street.

The second is a frame grabbed from video. Streaming and broadcast video is compressed far more aggressively than stills, and a single frame may carry motion blur, interlacing artefacts or a rolling-shutter skew. A cinematic aspect ratio also leaves less sky and less ground in frame than a still camera would have caught. Broadcast overlays — scoreboards, channel logos, lower thirds — cover a slice of the frame while sometimes betraying the broadcaster, which is a hint about audience rather than about place. A still photograph of the same scene would be an easier problem by a wide margin.

A clean PNG screen capture is lossless and adds no damage. The trouble starts afterwards, when the file is sent through a messaging app, pasted into a chat, uploaded and downloaded again, or screenshotted a second time from someone else's screen. Each hop typically re-encodes to JPEG, a standard published in 1992 that works by discarding the high-frequency detail the eye is least sensitive to. Unfortunately, small lettering is high-frequency detail. The 4:2:0 chroma subsampling most encoders default to throws away roughly 75 percent of the colour samples on top of that, which is why the red of a stop sign or the blue of a route shield can bleed several pixels past its own edge.

The practical consequence is specific rather than general. The model can still see that there is a sign; it can no longer read which alphabet is on it. Given how much a single legible word narrows a guess — the point of how script and language on signs narrow down a location — losing four or five characters of shopfront text can be the difference between a city and a continent. A screenshot that has been round a group chat three times has usually lost that battle already.

• —Capture wide, then crop later if you must. The sky, the horizon and the top of the buildings are worth more than the subject you were actually pointing at.
• —Keep any text in frame, even text you cannot read yourself. A half-legible shop name is often the single most valuable object in the picture.
• —Grab the original rather than a forward. A screenshot of a screenshot has been re-encoded twice, and the second pass is where sign lettering usually dies.
• —Do not zoom in before capturing. Digital zoom enlarges pixels without adding information, and the model reads the enlarged blur as blur.
• —Send the PNG if you are given the choice. A 1,080 px wide lossless capture beats a heavily compressed copy at any resolution.

None of that turns a hopeless image into a good one. A screenshot of a hotel ceiling has nothing in it, and being lossless will not help. But between two captures of the same scene, the wider and cleaner one is reliably the better upload, which is the argument made at length in best and worst photos to upload.

The short version is that losing the metadata costs a screenshot nothing here, because none of it was ever being read. What costs a screenshot is the crop and the re-encode, and those are choices you make at capture time. Grab the whole scene, keep the text, avoid the second-hand copy, and a screenshot performs about as well as the photograph it came from — here, or in the Geospy AI app if the picture is already on your phone.

Screenshots are one case of a wider question, and the three routes worth trying — metadata, image matching, visual reasoning — are compared side by side in find a location from an image.