How to Verify Where a Photo Was Taken
Finding a plausible location is easy. Proving one is a different discipline. The checks reporters use to test a photograph's claimed origin — and why an AI guess is a lead, never evidence.
Short answer
To verify where a photo was taken, corroborate rather than guess. Establish a candidate location, then test it against independent evidence: sun angle and shadow direction against the claimed date, weather records, signage and street furniture, and matching imagery of the same spot. One agreeing source is a coincidence; several is a finding.
There is a gap between two sentences that look almost identical. "This photo was probably taken in northern Italy" and "this photo was taken in northern Italy" differ by one word and by an entire standard of proof. Getting from the first to the second is verification, and it is a different activity from guessing.
The distinction matters more than it used to. Photographs circulate stripped of context, relabelled, and increasingly generated outright. A tool that returns a confident-sounding place name is useful for curiosity and actively misleading if you treat its output as a conclusion. Here is how the gap is actually closed.
What is the difference between guessing and verifying?
A guess reads one image and proposes a candidate. Verification tests that candidate against sources that had no knowledge of the photograph. The value lies in independence: agreement between things that could not have influenced each other is what makes a coincidence implausible.
A vision model looking at a street scene notices script, architecture, plate format and vegetation, and proposes somewhere consistent with all of them. That is genuinely useful work, and it is where the process starts rather than ends. The proposal comes from a single source of evidence — this one frame — interpreted by a single reasoner. Nothing has been tested against anything.
Verification inverts the direction. You take the candidate and go looking for ways it might be wrong. Every check you apply either survives or fails, and a candidate that survives several independent attempts to break it starts to deserve confidence. This is also why an honest confidence score matters so much, a subject covered in how much to trust an AI confidence score.
Which checks actually carry weight?
Four do most of the work: matching the exact spot in independent street-level imagery, testing shadow direction against the claimed date and time, checking weather records for the claimed day, and reading signage, plates and street furniture for regional consistency.
- Match the spot. Line up fixed geometry — a building corner, a kerb line, a lamp post spacing — against street-level imagery of the candidate location. Three independent alignments is close to conclusive.
- Check the sun. Shadow direction is set by the solar azimuth, which follows from latitude, date and time. A shadow pointing a direction impossible at that place on that day is a real contradiction, not a quibble.
- Check the weather. Historical records are widely available. Bright sun in a picture from a day of recorded heavy rain is a straightforward problem for the claim.
- Read the furniture. Bollards, kerb paint, road markings, plate proportions and signage typefaces are regionally consistent and hard to fake convincingly at scale.
- Trace the earliest copy. Find the oldest version of the file you can. A photograph presented as breaking news that has been circulating for four years is settled by provenance alone.
The ordering is deliberate. Matching the spot is the most powerful check and the most laborious, so it is worth doing only once cheaper checks have failed to eliminate the candidate. The sun and weather tests are fast and frequently decisive, and both have the useful property of being able to disprove rather than merely support.
Can I trust the GPS data in the file?
Treat it as a claim rather than a fact. Exif coordinates are trivially editable with free tools, and a file that has been forwarded, re-saved or exported may have lost or acquired metadata along the way. Corroborate coordinates the same way you would corroborate a caption.
Exif is convenient and it is not evidence. Any metadata editor can write arbitrary coordinates into a file in about 30 seconds, and there is nothing in an ordinary JPEG that binds the metadata to the pixels. This cuts both ways: absent coordinates mean very little, since most platforms strip them on upload, and present coordinates mean only that somebody, at some point, wrote them.
The useful move is to test the metadata against the image. If the file claims a coastal town in December and the picture shows a full-canopy deciduous tree and short shadows, something is wrong regardless of which part is lying. Metadata and content disagreeing is itself a finding.
How do I handle a photo that might be generated?
Verification changes shape entirely, because there is no true location to find. Look for internal inconsistency instead: lighting from two directions, text that dissolves under magnification, architecture that does not resolve, and reflections that do not correspond to the scene.
This is now a routine part of the work rather than an exotic case. A generated image has no provenance, no matching street-level view and no weather record to check, so the classic corroboration toolkit returns nothing — which can be mistaken for a hard-to-place real photograph. The tells are different in kind, and the practical ones are gathered in how to spot AI-generated travel photos.
Where do automated tools fit into all this? Near the start, as a lead generator. A model that says "probably Portugal" has turned an unbounded search into a bounded one, which is worth a great deal when the alternative is staring at a photo with no starting point. The technique differences between matching and reasoning are covered in Google Lens vs AI geolocation, and the index-based approach in reverse image search vs AI geolocation.
Need a starting point rather than an answer? Upload the photo and use the guess as your first candidate to test.
Upload a photo →The discipline is mostly a habit of mind. Ask what would have to be true if the claim held, then go and check whether it is. A candidate that survives the sun, the weather, the street furniture and a geometric match has earned the sentence without the hedge — and one that fails any of them has told you something more useful than a confident guess ever could.
Frequently asked questions
- Is an AI guess enough to verify a location?
- No, and it is not meant to be. A vision model produces a plausible candidate from visual clues. Verification means testing that candidate against independent sources until agreement becomes hard to explain by chance.
- What is the single strongest verification check?
- Matching the actual spot in independent imagery. If you can line up a building edge, a kerb line and a lamp post between your photo and a street-level view of the claimed place, you have something close to proof.
- Can shadows really disprove a claimed date?
- Yes. Shadow direction and length are governed by the sun's position, which is fixed by latitude, date and time. A shadow that could not occur at the claimed place on the claimed day is a genuine contradiction.
- What if the photo has GPS metadata — is that proof?
- It is strong but not conclusive. Exif coordinates are easy to edit and easy to fabricate. Treat them as a claim to be corroborated like any other, especially when the file has passed through several hands.
Sources
- Exif — WikipediaThe metadata block that may record coordinates and timestamps — and which ordinary editing tools can rewrite.
- Solar azimuth angle — WikipediaThe sun's compass bearing, determined by latitude, date and time — the basis of shadow-direction checks.
- Google Street View — WikipediaStreet-level imagery covering more than 100 countries, the usual reference layer for matching a specific spot.
Reminder
Raven is built for entertainment and curiosity. Its guesses are AI estimates that can be wrong, and it must never be used to track or identify real people. Uploaded photos are processed in memory and immediately discarded — never stored.
