GeoLocator v4.0: Eyes, Judgment, and a Map — Our New Three-Stage Engine
Every photo carries evidence. A licence plate. The wording on a stop sign. Which side of the road the cars are on. The shape of a bollard, the wiring on a utility pole, the roofline of a house. v3.0 taught us that reasoning over that evidence matters. v4.0 rebuilds the entire pipeline around that lesson: we split the job into three specialists — one that sees, one that decides, and one that turns a decision into a pin on the map.
Today, v4.0 replaces the v3.0 engine behind every GeoLocator photo geolocation analysis. On our own held-out benchmark of 69 photographs of ordinary towns — no famous landmarks, no easy wins — it names the right country 84% of the time and lands within 2 km of the true location on the median photo. It runs in about 15 seconds, at roughly one US cent of our own model cost per analysis — the demo and accounts are free.
Key facts: GeoLocator v4.0 at a glance
- Released: September 21, 2026. Replaces the v3.0 engine as the production engine.
- Architecture: three stages — a frontier vision model (eyes), a System One decision model (judgment), and a GeoNames gazetteer resolver (map).
- Country accuracy: 84% on our 69-photo town benchmark. The vision model alone scores 81%; the decision layer and knowledge base add the rest.
- Distance: 2 km median error; 35% of results within 1 km; 61% within 25 km; 93% within 2,500 km.
- Knowledge base: 2,265 curated GeoGuessr-style meta clues across 136 countries, including explicit "how to tell country X from country Y" rules.
- Resolver: GeoNames gazetteer with 34,000+ cities — region, then city, then coordinates.
- Calibrated confidence: on our earlier benchmark, high-confidence answers were right about 80% of the time versus about 9% for low-confidence answers.
- Speed and cost: about 15 seconds per analysis (12–21 s measured); roughly one US cent of our own model cost. The service is free to use.
Architecture: Eyes → Judgment → Map
v4.0 splits the work into three stages, gives each job to the component best suited to it, and passes structured evidence between them. Seeing is a perception problem. Deciding between look-alike countries is a knowledge problem. Turning a country and a town name into coordinates is a lookup problem. Three problems, three tools.
Stage 1 · Eyes
Frontier vision model
The vision stage reads the photo and produces ranked country candidates, each with written evidence, plus a nearest-town estimate. It is the perception layer — everything downstream works from what it sees.
Stage 2 · Judgment
System One decision model
Our decision layer disambiguates among the top candidates using a curated knowledge base of 2,265 GeoGuessr-style meta clues across 136 countries — licence plates, scripts and stop-sign wording, driving side, bollards, utility poles, architecture — and outputs a calibrated confidence.
Stage 3 · Map
Gazetteer resolver
A GeoNames gazetteer of 34,000+ cities resolves the region, then the city, and returns coordinates. The map step is a lookup rather than a guess — and lookups do not hallucinate.
The knowledge base is the part we are proudest of. It is not a pile of facts; it is organised around the questions that actually decide a photo. When the vision model narrows a photo down to two neighbouring countries that share a script, a road style, and a climate, the decision layer does not need to know everything about both — it needs the handful of rules that tell them apart. Those explicit tell-apart rules are what the 2,265 clues encode, and they are why a decision model can outperform the perception model that feeds it.
Why a decision layer beats a bigger model
The obvious move in 2026 is to throw a larger model at the problem. We tried the honest version of that experiment: run the frontier vision model on its own and score it. It gets the country right 81% of the time on our benchmark — genuinely impressive. Then we put the decision layer and its knowledge base on top. Same photos, same vision model: 84%. The extra points come from exactly the cases the layer was built for — photos where the top candidates look alike and only a rule can separate them.
Vision model alone
81%
country accuracy on our benchmark
Vision + decision layer
84%
+3 points from judgment, not scale
The second win is bigger than the first, and it does not show up in an accuracy number. The decision layer reports a calibrated confidence. On our earlier benchmark, when it said it was confident, it was right about 80% of the time. When it said it was not, it was right about 9% of the time. That gap is the whole point. A confidence score that actually tracks correctness is a trust signal you can build on: auto-accept the high-confidence results, and route the low-confidence ones to a second pass or a human analyst. A bigger model gives you a better guess. A decision layer tells you when to believe it.
Benchmarks
Country accuracy
84%
on our 69-photo town benchmark
Median error
2 km
on our 69-photo town benchmark
Within 25 km
61%
on our 69-photo town benchmark
Within 1 km
35%
on our 69-photo town benchmark
Also measured: 93% of results within 2,500 km, and the dev/test halves agree — 60% and 63% within 25 km respectively.
How we measure
These numbers are from our own benchmark, and we want to be precise about what that means. We selected 69 real photographs of ordinary towns from Wikimedia Commons mechanically, by coordinate — no hand-picking, and deliberately no famous landmarks. The set is held out from development and split into dev and test halves; the two halves agree, which is what we look for before we trust a figure. This is a town-photo benchmark, not a Street View benchmark, so we do not compare it against Street View results such as PIGEON — that is a different task with different imagery. When we say 84%, we mean 84% on this set.
Speed & cost
Three stages, one bill. An analysis takes about 15 seconds end to end — we measured between 12 and 21 seconds — and costs us roughly one US cent in model costs (the demo and accounts are free). That is fast enough for interactive use in the dashboard and cheap enough to run across a whole dataset.
Time per analysis
~15 s
12–21 s measured, all three stages
Our model cost per analysis
~1¢
what it costs us; the service is free
Conclusion
v4.0 is live today as the engine behind every GeoLocator analysis. It sees with a frontier vision model, decides with a knowledge base built from thousands of real-world clues, and resolves to coordinates with a 34,000-city gazetteer. It tells you how sure it is — and that number means something. Bring a photo of somewhere ordinary and try the demo.
New to this? Start with what geo-estimation is or our guide to finding where a photo was taken.