How It Works — Methodology and Accuracy

Most "AI location finder" sites give you an answer and no way to judge it. This page is the opposite: it explains what actually runs when you upload a photo, when each step is allowed to give an answer, and — more usefully — the specific situations where the result is likely to be wrong.

Read it before you trust any answer this site gives you, including the confident-looking ones.

The pipeline, step by step

When you submit an image, three checks decide whether there is anything to go on. They are tried in order of how trustworthy their answers are, and each one is allowed to answer only when it clears its own bar. If none does, the tool says it could not determine a location. It does not invent one.

1. GPS in the file

The file is parsed for embedded EXIF data. If the camera wrote GPS coordinates into the file and nothing has stripped them, that is a direct measurement rather than a guess, so it takes precedence over everything else. The coordinates are sent to a map service to look up the name of the place.

The catch is how often it is absent. Facebook, Instagram, X, WhatsApp, and most messaging apps strip EXIF on upload, and a screenshot never had it in the first place. In practice, most photos people bring to a tool like this have no GPS tag — which is precisely why the other checks exist. You can inspect this yourself, without uploading anything, using our EXIF location viewer, which reads the file in your own browser.

2. Landmark recognition

The image is sent to Google Cloud Vision, which matches the scene against a large reference set of known places — the Eiffel Tower, Sydney Opera House, Machu Picchu, and tens of thousands of less famous but still catalogued structures. Its coordinates come from that reference set, not from an estimate. A match is only used when Vision's own score is at least 0.5; in our testing every correct match scored well above that, and the one wrong match below it.

This is the step responsible for the results that feel like magic. It is also the one with the sharpest cliff: a landmark five metres outside its reference set returns nothing at all, and an ordinary street almost always does.

3. A vision model reading the picture

If neither of the first two answers, the image goes to a vision-language model running on Cloudflare Workers AI. It is asked for a place name — never coordinates — along with the region and country, what the answer rests on, and how sure it is. Its answer is used only when all of these hold:

• It rests on something specific. Either legible text that names or pins down the place — a station sign, a festival poster, a shop front — or one individually recognisable structure. An answer based only on the general look of a street, its trees or its architecture is discarded: that identifies a region, not a place.
• The model reports high certainty. Medium-certainty answers are discarded.
• The name is a place, not a description. Text such as PUBLIC LIBRARY or CITY HALL is not looked up.
• The map agrees. The name is looked up together with the stated region and country. If the map places it in a different country, or the same name exists far away in the same country and the stated region does not settle which one is meant, the answer is discarded.

Text is where this step earns its place. Place names and signs are often the single most decisive clue in an otherwise anonymous photo, and they are exactly what a landmark recogniser cannot use. Results from this step are labelled on screen as read from the picture, and shown without a percentage, because the model's own sense of certainty is not a measured probability.

4. Enrichment

Whichever step answered, the place is resolved against map data for a canonical name, country and coordinates; a background summary is fetched from Wikipedia; and a static map is generated so you can see the answer in context.

What we measured

Fifty photographs, one taken near an ordinary point in each of fifty cities on six continents — Tartu, Kumasi, Arequipa, Pune, Townsville and so on — drawn from Wikimedia Commons, skipping anything titled as a famous landmark. Every one records the coordinates it was taken at, so an answer can be scored by how many kilometres it misses by. Before uploading we stripped the EXIF data, so the tool could not simply read the GPS tag, and renamed each file to something neutral like IMG_0001.jpg. Each photo went in through the same endpoint the box on our homepage uses.

We ran the whole set three times, because the model does not answer identically every time.

So: on photographs like these the tool says nothing about six times in seven. When it does answer, it lands within 25 km about four times in five, and it did not once put an answer more than 750 km from where the photo was taken. The worst miss in the three runs was 132 km — a pedestrian street in Bendigo given as central Melbourne. Five photographs were answered correctly in all three runs; the rest of the answers came and went between runs.

Where the answers came from:

Why the numbers on this page changed

An earlier version of this page reported a single run on a different set of fifty photographs: 14 answered, 10 of them within 25 km. A second set, gathered the same way, did not reproduce it. With the model we were using then, that second set produced 16 answers of which five or six were within 25 km, and six answers landed more than 750 km out — a street in Penang given as Kagoshima, 4,286 km away; Tbilisi given as France; Chengdu given as Taiwan.

One sample of fifty was too small to publish a hit rate from, and we published one. That is the correction.

The failure was not in the checks around the model but in the model itself, so we compared the vision models available to us on the same photographs. The one now in use answered less often than its predecessor and was right more often, and stopped producing answers on the wrong continent. The table above is that model, measured after the change.

How far to trust these numbers

• Fifty photographs is still a small set. Three runs over the same fifty tell you about consistency, not about the world. Treat the figures as the order of magnitude, not a guarantee.
• These are other people's public photographs, not photos visitors uploaded. We have not measured the tool on visitors' own pictures, and we do not keep visitors' pictures to measure.
• Photographs that illustrate famous Wikipedia articles make any tool look better than it is. They are among the most copied images online, so a vision model has likely seen them in training and may be recognising the picture rather than working out the place. We test on rarely reproduced photographs for that reason.
• The models behind this change without notice. These figures describe the tool on 19 September 2026.

Where this is reliably wrong

We would rather you know the failure modes than discover them at a bad moment.

Two structural limitations are worth stating plainly. First, the confidence score is a model output, not a probability of being correct — a 0.9 does not mean nine times out of ten. Second, landmark coverage is uneven by geography: densely photographed parts of Europe, North America, and East Asia are far better represented than most of Africa, Central Asia, and rural South America, so expect worse results there.

How to verify a result

Treat every answer as a hypothesis and spend two minutes testing it:

1. Look at the map. Does the surrounding terrain match the photo — coast where there should be coast, hills where there should be hills?
2. Check street-level imagery. Line up a fixed feature: the angle between two buildings, a road curve, a roofline against a hill. Geometry does not lie; vibes do.
3. Cross-check the text. If a sign in the photo names a business, search it and see whether the address agrees with the answer.
4. Try a reverse image search. If the photo is published anywhere, that often settles it outright. Our guide to reverse image location search covers which engine to use for what.

If steps 1 and 2 disagree with the answer, the answer is wrong. That happens, and we would rather you catch it than cite it.

Editorial policy for our written guides

The articles in our blog are a separate product from the tool, and they follow their own rules:

• Every guide is written by a person, from methods that person has used, and is edited before publication. We do not auto-generate articles at scale, and we do not publish a guide to a technique we have not tried.
• Tool comparisons name competitors and say when they are better. Where a rival tool beats ours at a specific job — for example, a published photo where reverse image search wins outright — we say so, because a comparison that always concludes "use us" is worthless to the reader.
• No affiliate links, and no paid placements. Nothing in our rankings is bought. The site is funded by advertising, which is labelled and served separately from editorial content.
• Claims are checked before they run. Where a guide states what a tool or platform does — that a given service strips EXIF on upload, say — that behaviour was tested rather than assumed.
• Corrections are made in place, with the change noted, rather than by quietly deleting the page.
• We prune. Guides that stop being accurate or useful are rewritten or removed. We would rather run two dozen guides worth reading than a hundred that pad a sitemap.

Privacy in one paragraph

Uploaded images are used only to answer the question you asked. They are not used to train models, not published, and not sold. Original files are deleted after processing. The EXIF viewer runs entirely in your browser and never uploads your photo at all. The full detail, including our advertising and cookie disclosures, is in the Privacy Policy and Cookie Policy.

Questions and corrections

If you think a result is wrong, or you have found a failure mode that is not on this page, tell us at support@whereisthisplace.org. Reports of bad results are how this page gets more accurate.