ChatGPT Got Scarily Good at Locating Photos. The Viral Prompt Wasn’t Why

In April 2025, Kelsey Piper gave OpenAI’s new o3 model a photo of a beach. With no further questions, it named the exact spot: Marina State Beach, in Monterey County, California. Commenters assumed a trick. Perhaps the file still held its location data, or o3 remembered where she lived from earlier chats, or it traced her IP address.

Piper had also been using a long prompt, built up over time, which she said significantly improved o3’s performance. Scott Alexander, who writes the blog Astral Codex Ten, ran it on a set of test pictures, mostly old photos of his own that had never been online, stripped of metadata and flipped to make matching harder. A shot of bare rocks and the flag of an imaginary country, taken on a Himalayan peak, came back as “Nepal, just north-east of Gorak Shep, ±8 km”. That was exactly right.

The prompt was shared widely as “the GeoGuessr prompt”, and many people who tried it came away impressed. A year later, someone finally ran the controlled test. On average, the prompt did not make o3 any better.

Scott Alexander reproduced it in full in Testing AI’s GeoGuessr Genius. Its main parts:

• Ground rules. “No metadata peeking. Work only from pixels (and permissible public-web searches).” It also tells the model not to reason from the user’s IP address.
• Raw observations. Up to ten bullet points of what can literally be seen or measured, such as colour, texture, shadow angle and the shapes of letters, with interpretation held back.
• Clue categories. Climate and vegetation, landforms, the built environment, culture and infrastructure, and the light.
• Five candidates. A ranked shortlist in which the first and fifth places must be at least 160 km apart.
• Disproof. For the leading guess, spell out what would kill it. “You are an LLM, and your first guesses are ‘sticky’ and excessively convincing to you,” the prompt warns.
• Lock-in. Coordinates or the nearest named place, plus an uncertainty radius in kilometres.

It even includes a formula for estimating latitude from the length of a shadow, a technique with its own guide. It reads like a checklist for a careful human geolocator, which is part of its appeal.

Goedecke’s short prompt simply told o3 to geolocate the photo from visual evidence and return coordinates, a confidence score and the clues it used. The long version got the same answer format added. Both ran through OpenAI’s API without web search, so the test measured the protocol’s reasoning steps, not the search steps it also describes.

The long prompt did have wins. It sometimes kept a guess in the right country when the short prompt drifted to the wrong part of Europe, and it placed one indoor photo, a shopping arcade in Moscow, within 200 metres, where the short prompt said Saint Petersburg. But its harmful swings outnumbered and outweighed its helpful ones, Goedecke’s report found.

Know the test’s limits. It is one benchmark, built by one person for about 15 US dollars. Most of its photos (180 of 200) came from Geograph, a project that documents Britain and Ireland, with the rest from Wikimedia Commons, and public photos may already sit in a model’s training data. Goedecke checked that none carried EXIF metadata, and argues that a model that had memorised the photos would have scored better.

Piper built the prompt by asking o3, after each mistake, how it could have avoided it, then adding the answer. Goedecke’s point is that this loop flatters the prompt: models “will almost always say ‘yes, that helped a lot!’ when you ask them if a particular prompt tweak made things better.” The only way to know, he writes, is a benchmark.

Alexander’s own run shows why single examples mislead. With the prompt, o3 put a featureless plain in the right 300-by-100-mile region of Texas and New Mexico but, asked for a point, missed by about 110 miles. It could only call a California dorm room “a dorm room on a large public university campus in the United States”, put a patch of Michigan lawn in the Pacific Northwest, and took a brown stretch of the Mekong for the Ganges. His old house in Michigan came back as Richfield, Minnesota.

Striking hits are what people share. A test on hundreds of photos says more than any single screenshot, however uncanny.

“Whatever o3 had that made it good at this task hasn’t transferred to newer models,” Goedecke concluded. Which model you actually get matters too: after GPT-5 arrived, OpenAI removed the option to pick older models such as o4-mini-high in ChatGPT, Bellingcat reported. How ChatGPT compared with Google Lens and other chatbots has the rest of Bellingcat’s results.

1. Step 1Read the metadata firstCoordinates stored in the file are measured, not guessed, so check for GPS data before anything else. Photos saved from social apps have usually lost it.
2. Step 2Start a clean chatIn ChatGPT, open a temporary chat and choose Unpersonalized before the first message; OpenAI says that mode does not use memory, custom instructions or plugins. Bellingcat found that account history could steer answers, and ChatGPT may still estimate your rough location from your IP address.
3. Step 3Use the sharpest copySmall details decide the result: in one Bellingcat test, the exact address came from writing on a mailbox. Upload the original rather than a compressed repost, and add a close-up crop of any text.
4. Step 4Ask plainly, and ask for evidenceA short request did at least as well as the 1,100-word protocol. Try: Where was this photo taken? List the visible clues you used, give three candidate places with a confidence for each, and say what would rule each one out. The extra lines are not there to raise accuracy; they give you something to check.
5. Step 5Check the candidates yourselfLook for two or three independent matches, such as a sign, a roofline and the shape of a hill, in Street View or satellite imagery. For landmarks and tourist spots, run a reverse image search too: Google Lens beat most chatbots on those in Bellingcat’s tests, and identifying a landmark covers that route.
6. Step 6Keep the uncertaintyEvery model Bellingcat tested was at some point entirely wrong. Note how confident the answer was, and never treat a chatbot’s guess as evidence about a person.

The same order works for any method, AI or not: measured data first, a hypothesis second, confirmation last. The four ways to find where a photo was taken sets it out in full.

The prompt borrows the game’s name, not its rules. It was written for real photos, which, it notes, may come from private land, someone’s backyard or an off-road trip. Whether AI can beat top GeoGuessr players, and how to use it for practice between rounds, has its own guide.

GeoSpy works from what is visible in the frame, answers at city level rather than with an address, and does not store uploaded photos. On geospys.com, analysis comes with GeoSpy Premium: 7.99 US dollars a week or 29.99 US dollars a year (₺389 or ₺1,460 in Türkiye), unlimited under fair use, and you can cancel any time. Choose the yearly plan, or start with the free trial in the GeoSpy iPhone app.