You upload a photo, the map flies somewhere, and it is not where the photo was taken. Or it lands near the right place but draws a circle the size of a city when you wanted a street. Either way it feels like the tool got it wrong.
You are not imagining it, and we owe you the real reason.
The honest answer: it came down to control
On our consumer product, we deliberately limited how precise results could be. A tool that can place a photograph on a street can be pointed at a person, and we did not want stalkers or other bad actors using Oceanir to find someone. So we held the consumer model back, accepting less accuracy for everyone to keep that capability out of the wrong hands.
It was the wrong trade. How people actually use Oceanir has shown us that the limit cost the many people with good reasons far more than it ever cost the few without them. Journalists checking a source, people verifying a listing, someone trying to find where an old family photo was taken: they all got a weaker answer than we could have given them.
So we are overhauling that system. Instead of capping accuracy for everyone, we are building a smarter way to protect the people using Oceanir and the people who appear in their photos, aimed at misuse itself rather than at every analysis. Our newest Orca model is built for exactly this: full accuracy for the work people actually do, with protection where it belongs.
Until it lands, some of what you see is the old limit. The rest of this post is about the other reasons a result can look off, which apply to any geolocation system, and what you can do about them today.
The photo sets the ceiling
Oceanir works from what is visible in the image. It does not read GPS metadata, so a result is only as specific as the evidence in the frame: the language on a sign, road markings, a licence plate format, the style of a building, the vegetation, the angle of the sun.
Some photos carry plenty of that. A street corner with a shop sign and a bus stop can be placed to the block. Others carry almost none: an indoor shot, a close-up of food, a stretch of beach, a forest trail, a car interior. Those photos look much the same in a lot of countries, so the honest answer is a region rather than a point. Heavy cropping, filters and repeated screenshotting make it worse, because they destroy exactly the small details that separate one place from another.
An area is an answer, not a miss
When a result comes back as a wide area, that circle is the claim. It means the evidence supports a region and no more, and the system is declining to pretend otherwise. A pin dropped confidently on a random street inside that region would look more impressive and be wrong far more often.
The ranked list of other candidates is part of the answer too. If the top result is off but the right place is second or third, the photo genuinely fit more than one location, and the list shows you which ones were close.
Some places are harder than others
Geolocation is easiest where the world has been photographed the most: big cities, well-known streets, places with dense street-level imagery. It gets much harder in places that rarely appear in any photo collection.
The research field measures this directly. GWS15k, a public benchmark of street photos taken near random city centres around the world, was built so that 92 percent of its locations never appear in the data the field trains on. On it, no published system places even one photo in ten within 25 km of where it was taken (PIGEON, CVPR 2024). If your photo is from somewhere unfamiliar to the rest of the internet, expect a wider answer, and treat a very confident one with care.
Confidence is a judgment, not a promise
The confidence score is the system’s estimate of how well the evidence supports its answer. It is not a guarantee, and it should never be read as one.
We will be direct about this: we have seen results that were confident and still wrong. That is the failure we care about most, because a confident wrong answer is worse than an honest uncertain one. The work we are doing now is aimed at exactly that, so that a high number is something a result has to earn and a specific location only appears when the evidence reaches it.
What to try when a result looks off
- Use the original photo. The full resolution file beats a screenshot, and an uncropped frame beats a cropped one. Every sign and edge you keep is evidence.
- Crop to what matters, after the first run. If one part of the frame holds the clue, cropping to it lets the analysis spend its attention there.
- Run a deeper analysis. A quick result is a fast first read. Standard and Forensic look harder at the details and check more alternatives before answering.
- Look down the candidate list. The right place is often in the list even when it is not first.
- Check it against the world. Open Street View or satellite at a candidate and compare it with your photo. If the buildings and road match, you have your answer; if they do not, move on to the next one.
- Ask for a recheck. Some low-confidence results offer one. It takes another pass at the same photo.
Tell us when it is wrong
The results panel lets you tell us whether a result was right. Please use it, especially when it was not. A wrong result you report is the most useful thing you can send us: it shows exactly where the evidence and the answer came apart, and it is how the next version gets better.
If a photo matters and the result still is not good enough, write to us at [email protected] with what you expected. We read every one.
Do you want to make your analysis stronger? Take our survey.
Six quick questions about what you upload, where it is from, how precise you need it, and how results have let you down. Your answers decide which photos, places and failures the new model is tuned for first.
