A map pin communicates certainty even when the evidence supports only a district. Geolocation systems often must return coordinates because maps require them, but users should distinguish the coordinate used to display a place from a verified camera position.
A real 61-metre precision trap
In our Siem Reap investigation, Nominatim correctly returned a Pub Street attraction node at 13.3548263, 103.8547837. Wikimedia's independently published camera position was 13.354639, 103.854250—about 61 metres away. The place name and district were correct; the exact photo pin was not.
Nearby mapped objects clarified the viewpoint: Red Piano was about 11 metres from the camera coordinate, World Lounge Pizza 22 metres, Street 8's feature centre 53 metres and Cambodian BBQ 57 metres. Their order explained the photograph. No single POI coordinate equalled the camera position.
Four claim levels
| Claim | Evidence required | Honest output |
|---|---|---|
| Country | Independent national clues | Country with alternatives |
| Region | Environment or infrastructure narrows area | Boundary or broad area |
| City/district | Named and visual clues converge | City or neighbourhood centroid as display-only |
| Camera coordinates | Viewpoint and fixed geometry match | Point with verification record |
Measure distance correctly
Use great-circle distance between prediction and ground truth. For a dataset, report median, 90th percentile and percentages within thresholds such as 1, 25, 200 and 750 kilometres. The mean alone can be dominated by a few continent-scale misses.
Choose thresholds before viewing results and connect them to use. A 25-kilometre result may be useful for regional media verification but useless for matching a storefront. A country hit should not be described as coordinate accuracy.
Confidence needs calibration
Group held-out predictions by stated confidence and measure the success rate in each band. A 90% label should succeed around nine times in ten under the defined criterion. Persuasive explanations and decimal coordinates are not calibration.
When numeric confidence is absent, candidate spread offers a practical signal. Five candidates in one neighbourhood indicate a different uncertainty from five candidates across continents. Do not average unrelated coordinates: the midpoint may be somewhere no evidence supports.
Separate semantic and geometric correctness
- Semantic correctness: the system named the right attraction, road or district.
- Geometric correctness: the coordinate lies within a measured distance of the camera.
- Viewpoint correctness: the coordinate and orientation reproduce fixed object relationships.
The Pub Street node was semantically correct and geometrically close, but it did not represent the viewpoint. This distinction should appear in product evaluation and user-facing results.
A responsible result format
Return the supported area, claim level, candidate point used for display, whether that point is a centroid or verified camera position, alternatives, decisive clues, contradictions and next verification step. If exact coordinates are unverified, say so beside the map rather than hiding uncertainty in a tooltip.
Sources
Put the method into practice
When you have a photo to investigate, start with our AI photo location finder and treat its result as a lead to verify, not automatic proof.