How Accurate Is AI at Finding a Photo's Location? What the Benchmarks Actually Show

Introduction

Consider two photographs handed to the same model. The first is a street corner in Lisbon, where the model identifies the neighborhood and lands within a few hundred meters. The second is an unremarkable stretch of green hillside somewhere in Southeast Asia, and that identical system relocates it to the wrong continent entirely. Nothing about the underlying software changed between those two attempts. What changed was the photograph.

This is the part that almost every page promising AI photo location leaves out. You will find plenty of writing that calls the technology accurate, and very little that says how accurate, measured how, on what kind of image. Between early 2025 and April 2026, several research groups published benchmarks that answer exactly that question — and their results are far more useful than any single number, because they show where the technology succeeds and where it collapses.

This article walks through what those benchmarks measured, what the numbers actually mean, and how to look at your own photo and estimate the odds before you upload it.

Three Metrics, and Why They Aren't Interchangeable

Before the numbers, a short glossary. The benchmarks below use different yardsticks, and mixing them up is the single most common error in writing about this topic.

Top-1 accuracy is the share of photos for which the model's first answer is correct, usually at country level. Median distance error is the distance between guess and truth for the middle photo in the set, measured in kilometers — a median rather than an average, because a handful of wrong-continent misses would otherwise drag the mean beyond usefulness. Elo rating is a relative score derived from head-to-head human preference votes; it ranks models against each other but says nothing about absolute correctness.

A model can look strong on one and weak on another. Keep the yardstick attached to every figure that follows.

The Short Answer: Median Errors Under 300 km

Ask a current vision-language model where a single photograph was taken, and it will typically land within a few hundred kilometers. That is good enough to identify a region. It is rarely good enough to identify a street.

The clearest figure comes from a study presented at the AAAI 2025 DATASAFE workshop. It evaluated foundation models on single-image geolocation against previously unseen imagery. Many achieved median distance errors below 300 km (Jay et al., arXiv:2502.14412, 20 February 2025).

Country-level results are stronger. A world-scale analysis published on 17 April 2026 tested nine vision-language models across three datasets, using ground-view imagery only. Its best performer, Qwen3-VL-4B, reached 74.79% Top-1 country accuracy on the GeoGuessr-50k set and 65.78% on CityGuessr (arXiv:2604.16248).

Push to finer scales, however, and performance drops sharply. The EarthWhere benchmark evaluated 13 models with web-search access across 810 globally distributed images. Half the set was country-level multiple choice; the rest demanded street-level reasoning. Its top scorer, Gemini-2.5-Pro, managed 56.32% average accuracy. The strongest open-weight entrant, GLM-4.5V, reached 34.71% (arXiv:2510.10880, 13 October 2025).

Those three figures are not comparable to each other, and that matters:

Anyone quoting a single "AI photo location accuracy" percentage is collapsing four different measurements into one. That, more than anything else, is how the topic became so murky.

For more on the mechanics underneath these scores, see our explainer on how AI geolocation works.

Why the Average Is Misleading

The gap between an easy photo and a hard photo is wider than the gap between the best model and a mediocre one. That single fact should reshape how you read every accuracy claim in this field.

The spread between models is real enough. In the world-scale analysis, Top-1 country accuracy ranged from 74.79% for the best model down to 19.23% for LLaVA-Vicuna-7B on the OSV5M dataset. But the same model, held constant, swings almost as much depending on what it is looking at — and on some axes, considerably more.

EarthWhere quantified one version of this directly, reporting performance varying by up to 42.7% across geographic regions. That is not noise. That is a systematic property of the technology, and it decomposes into three factors you can actually inspect in your own image: whether the scene is built-up or wild, what biome it sits in, and whether there is legible text anywhere in frame.

There is a second, subtler quality difference worth knowing about. The world-scale study also scored how reasonable each model's errors were — whether a wrong answer was at least visually justified by the image. Stronger models failed more sensibly: Qwen3-VL-4B produced visually defensible errors 44.4% of the time against LLaVA-Vicuna's 10.7% on the same dataset. A good model that misses tends to miss toward somewhere that genuinely looks similar. A weak one guesses.

Axis 1: Urban Photos Beat Rural Ones by 15 Points

City photographs are substantially easier for AI to place, and the advantage holds across every model tested.

The world-scale analysis found a consistent gap of roughly 15 percentage points between urban and rural imagery across all nine models it evaluated. For the strongest model, that meant 83.8% Top-1 country accuracy on urban photos against 68.5% on rural ones (arXiv:2604.16248, 17 April 2026).

The explanation is information density. A single urban frame can accommodate numerous independent identifiers, such as shop signage, license plates, road markings, bus livery, architectural period, utility hardware, and street furniture. Each additional identifier constrains the geographic possibilities further. A rural frame, by comparison, might offer only vegetation, soil coloration, terrain, and perhaps a fence — fewer signals, and the ones remaining are distributed across substantially larger areas of the planet.

This is the same logic that makes a skyline so tractable; our guide on identifying a city from its skyline covers why built environments carry so much locational information. The inverse case — placing a trail or mountain photo — is harder for exactly the reasons the benchmark measures.

Axis 2: Biome Moves the Needle More Than Anything Else

Terrain type produces the largest measured swing in the entire literature. In the world-scale analysis, the best country-level accuracy achieved in boreal landscapes was 76.8% Top-1 on the OSV5M dataset. In tropical regions, the best any model managed was 20.5%.

A 56-point spread between biomes comfortably dwarfs the difference between frontier and mid-tier models. For example, an unremarkable model examining a Finnish forest road will reliably outperform an excellent model examining a Costa Rican one — the terrain, not the architecture of the system, determines the outcome.

Two mechanisms drive this disparity. The first is genuine visual convergence: humid tropical vegetation appears broadly interchangeable across widely separated regions such as Panama, Cameroon, and Malaysia, and the built environment visible in such photographs frequently carries fewer regionally distinctive architectural markers.

The second mechanism is coverage bias. The imagery these models learned from is not distributed evenly across the planet, so "difficult biome" and "under-represented region" overlap considerably. Consequently the benchmark measures a combined effect that it cannot fully separate. That caveat deserves attention before anyone concludes that tropical scenes are inherently unidentifiable.

If you want to understand what vegetation actually reveals, our guide to reading plants and biomes for location clues covers the signals a careful human looks for — many of which models are demonstrably still learning.

Axis 3: Readable Text, and Why Search Doesn't Rescue a Weak Photo

Legible writing in frame is the single highest-value signal in a photograph, and its absence cannot be compensated for by giving the model more tools.

That second half is the counterintuitive part. The EarthWhere team equipped all 13 evaluated models with web search, on the reasonable assumption that a model able to look things up would do better. Their conclusion was that "web search and reasoning do not guarantee improved performance when visual clues are limited" (arXiv:2510.10880, 13 October 2025). Search helps a model confirm a hypothesis. It does not help a model form one from an image that contains nothing to go on.

Tools do help when the photo cooperates. The DATASAFE study found that giving vision-language model agents access to supplementary tools reduced distance error by up to 30.6% (Jay et al., arXiv:2502.14412). The pattern across both results is consistent: augmentation amplifies signal, it does not manufacture it.

Script and signage sit alongside road markings, utility poles, and driving side in the standard toolkit — our OSINT guide to verifying where a photo was really taken works through how each one narrows the search.

Two Findings That Should Change How You Read Any Accuracy Claim

The benchmarks turned up two results that undercut common assumptions, including assumptions we would benefit from you holding.

Bigger is not reliably better. The world-scale analysis recorded an inverted-scaling result: Qwen3-VL-8B scored 7.49 percentage points below the 4B variant of the same family on GeoGuessr-50k. Parameter count is not a proxy for geolocation skill, and a tool built on a larger model is not, on that basis alone, more likely to place your photo correctly.

Confidence reads as accuracy, and it isn't. GeoArena evaluates 17 models through pairwise human votes rather than distance-to-truth, and in analyzing those votes the researchers found systematic stylistic preferences: raters favored longer responses (coefficient +0.526) and answers that included GPS coordinates (+0.06) (arXiv:2509.04334, platform live since June 2025).

That finding applies to us as much as to anyone else. A geolocation result rendered as precise decimal coordinates feels more authoritative than the same underlying guess expressed as a region, and the feeling is unrelated to whether the guess is right. When you use any photo location tool, including ours, treat the format of the answer as a presentation choice and not as evidence. The useful question is never how confident the output looks — it is whether the image contained enough to support the conclusion.

Will AI Find Your Photo? A Practical Predictor

You can estimate your own odds before uploading anything by scoring your photo on the factors the benchmarks actually measured.

Roughly, a photo hitting four or five of the left-hand column is in the range where the best models identify the correct country most of the time and often land within tens of kilometers. A photo sitting in the right-hand column on three or more rows is in the range where the same models return the wrong country more often than not.

One important distinction before you conclude a tool has failed you: everything above concerns location inferred from pixels. If your photo carries GPS metadata, the location is already written into the file and no inference is required — a completely different and far more precise path. You can read those tags yourself with our free EXIF location viewer.

You can test where your own image falls on this table with our free photo location finder.

Frequently Asked Questions

How accurate is AI at finding a photo's location?

It depends on what you're measuring. On distance, foundation models achieve median errors under 300 km on single-image geolocation (Jay et al., 2025). On country identification from ground-level photos, the best of nine models tested in April 2026 reached 74.79% Top-1 accuracy (arXiv:2604.16248). On harder mixed-scale tasks that include street-level reasoning, the best of 13 models managed 56.32% (EarthWhere, 2025). These are three different questions with three different answers.

Can AI get exact GPS coordinates from a photo?

Rarely from the image alone. Models will often output coordinates, but those are the format of an estimate rather than a measurement — median errors in the hundreds of kilometers mean the decimal places are not meaningful. Exact coordinates come from EXIF metadata embedded by the camera, which is a different mechanism entirely. Check your own file with the EXIF location viewer, or see how to remove that metadata before sharing.

What kind of photo works best?

Urban, daylight, wide framing, visible signage, distinctive architecture, and a temperate or boreal setting. The predictor table above lists the measured factors in full.

Is AI better than a human at this?

On Street View panoramas, current top models now edge out expert humans on precision. The GeoBench leaderboard shows its leading model at a 109 km median distance against a professional player baseline of 220 km, while the human still leads narrowly on country accuracy at 90% versus 89% (GeoBench, ACW map, n=100, retrieved 23 August 2026). Note the task: those are interactive Street View panoramas, not arbitrary uploaded photographs, so the comparison does not transfer directly to a phone snapshot.

Can someone find my location from a photo I posted?

Sometimes, and the researchers themselves flag this. The DATASAFE authors note that modern foundation models can act as capable geolocation tools without having been trained for the task, and that their growing accessibility carries implications for online privacy (Jay et al., arXiv:2502.14412). The predictor above works in both directions: a photo that is easy for you to locate is equally easy for someone else. Our photo location privacy guide covers what to strip before posting.

Conclusion

AI photo geolocation is genuinely capable and genuinely uneven, and the honest answer to "how accurate is it" is a range rather than a number. Median distance errors under 300 km and country accuracy around 75% describe the technology at its measured best. A tropical hillside with no signage sits at the other end of that distribution, where the same systems return the wrong continent.

The practical consequence is that the photograph, not the model, is the main variable. Before you judge a tool, look at what you handed it: built-up or wild, temperate or tropical, text or no text. Those three questions predict your result better than any product comparison will.

If you want the human version of the same skill — the clues a careful observer reads before any software gets involved — start with our step-by-step guide to identifying a location from a photo.