When a vision model guesses where a photograph was taken without web search, it relies on a mixture of botanical cues, road markings, utility poles, and architectural vernacular.1

Median error across models

Across the 400-frame synthetic split, median great-circle distance varies by more than a factor of three between the strongest and weakest models.

Show data
ModelMedian geolocation error (km)95% CIn
Gemini 3 Pro 118 km 94 km – 146 km 400
Claude Opus 4.5 142 km 116 km – 171 km 400
GPT-5 164 km 132 km – 198 km 400
Qwen3 Max 248 km 205 km – 294 km 400
Grok 4 312 km 261 km – 368 km 400
Llama 4 Maverick 425 km 370 km – 488 km 400
Fig. 1 — Median great-circle error across 400 closed-book frames (placeholder data). Lower is better.

Geographic error vectors

Plotting predicted coordinates against ground truth reveals whether errors stay regional—confusing neighbouring coastal towns—or jump continents entirely.

● Truth  ● GuessNorth Atlantic harbor (Gemini 3 Pro) — error 8.5 km
Show data
SampleModelTruth (lon, lat)Guess (lon, lat)Error (km)
North Atlantic harbor Gemini 3 Pro -21.94, 64.15 -21.79, 64.11 8.5
Andean coastal terrace Claude Opus 4.5 -71.61, -33.05 -70.66, -33.45 98.4
Steppe railway siding GPT-5 71.45, 51.16 69.18, 53.29 284.1
Inland river delta Qwen3 Max 105.78, 10.04 106.63, 10.82 127.6
Mediterranean limestone ridge Grok 4 15.29, 37.07 12.49, 41.90 589.2
Boreal logging road Llama 4 Maverick 25.47, 65.01 18.06, 59.33 741.8
Fig. 2 — Representative guess-to-truth vectors rendered on self-hosted Natural Earth outlines (placeholder data). Lower is better.

What separates coastal from inland recall

Models consistently perform better near distinctive coastlines and major transit corridors than on unlabelled agricultural or boreal roads. Full per-track breakdowns will accompany the first non-dummy snapshot run.