Have LLMs Finally Mastered Geolocation?

Bellingcat benchmarked 20 LLMs and Google Lens across 500 geolocation trials on 25 unpublished images, finding ChatGPT o3 and o4-mini variants marginally best overall, with most others trailing or hallucinating. LLMs outperformed Lens in urban scenes by exploiting small textual and stylistic cues and multilingual search, while Lens was stronger on well-photographed touristic landscapes. Deep or extended reasoning modes did not reliably improve accuracy and sometimes made models more cautious or slower. The authors conclude LLMs are useful aides for OSINT geolocation but remain prone to hallucinations and should be combined with other methods.