Benchmark
Bellingcat Photo Geolocation Evaluation
How close is a model’s location estimate when the photo has never been online?
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
Foeke Postma and Miguel Ramalho describe the October 2026 round: 52 previously unpublished photos and one coordinate estimate per photo. Earlier rounds used different conditions.
Methodology
Distance from the true coordinates is scored two ways. The GeoGuessr measure rewards finding the general area; Bellingcat’s logarithmic measure distinguishes errors in the final kilometres. Getting the region right does not verify an exact location.