Step-by-Step: Uploading Your First Photo to Raven

So you have heard that Raven can look at a photo and guess where in the world it was taken, and you would rather try it than take our word for it. Good — that is exactly what it is built for. The whole flow, from picking a photo to reading your result, takes about a minute. It still helps to know what to expect at each step, particularly the parts that are easy to get wrong on a first attempt, such as choosing a photo that gives the model almost nothing to work with.

This walkthrough covers the four steps in order, then the questions that usually come next: what the confidence figure means, what happens to your image afterwards, why a result sometimes comes back deliberately vague, and what to try when it does.

It is worth being precise about the category. Raven is not a lookup service and not a forensic tool. Nothing is matched against a database of your photographs, and there is no index of the world's images being searched behind the scenes. The model is shown one picture and reasons about it the way a well-travelled person might, only faster and with a much wider frame of reference. That framing matters because it sets expectations correctly: a well-reasoned estimate from visible evidence is a different thing from a coordinate, and it can be confidently wrong.

The single biggest lever over how impressive your first result feels is the photo you choose. Raven reads environmental context, so a frame with some visible outdoors gives it far more to work with than a close-up of an object against a blank wall. Three kinds of photo tend to produce the most satisfying first run: an old travel photo whose real answer you already know, so you can mark the guess against your own memory; a still from a film or series with a distinctive outdoor setting; and a friend's holiday photo from a trip they are being cagey about.

• —Street scenes — shopfronts, kerbs, bollards, road markings and parked cars all in one frame.
• —Coastlines and hillsides — rock colour, vegetation and the angle of the light do a surprising amount of work.
• —Building facades — window proportions, roof pitch, balcony style and render colour are strongly regional.
• —Anything with legible text — a fascia, a number plate, a bus destination board, a menu in a window.

If you want the longer version of this decision, the companion guide to the best and worst photos to upload for accurate guesses works through the categories in detail, and there is a shorter pre-upload checklist in the quick-start guide to getting accurate guesses.

Before your first upload, Raven asks you to sign in with Google. This is not a gate on the rest of the site — the blog, the FAQ and the legal pages are all open — it is specific to the analysis feature, because each guess costs something to run. The sign-in uses the standard delegated flow described by OAuth, whose current version was standardised in 2012: you approve the sign-in on Google's own page, Raven receives confirmation of who you are, and no password is ever typed into this site. There is no profile to complete and nothing to configure afterwards.

Drop your photo onto the upload area or tap to browse for it. Behind the animation, the file is validated — the type is checked from the file's own leading bytes rather than trusting whatever label the browser attached — converted for transport, and passed to the model with a single instruction: work out where this was taken and say how sure you are. Most runs finish in a handful of seconds. If a run fails, retrying with the original file rather than a screenshot of it resolves the majority of cases.

Your result comes back as a location — sometimes a specific city or district, sometimes a broader area when the photo is genuinely ambiguous — paired with a confidence indicator rather than a bare assertion. That indicator is the part people skip and the part that carries the most information. A photo full of distinctive clues, such as a recognisable roofline, a legible sign and a particular species of street tree, earns a narrow, confident guess. A hazy landscape with a plain sky and generic scrub will come back vaguer, and that is the correct behaviour rather than a failure. A precise answer drawn from a photograph that does not contain precise evidence would be worse, not better.

The instructive experiment here is to deliberately feed it photographs with almost nothing in them and watch the hedging increase. We ran exactly that test in our write-up of testing AI with nearly impossible photos, and the pattern was consistent: as evidence thins, the answers widen. A model that stays vague on a blank wall is behaving well, even though it makes for a dull demonstration.

That single design decision is what makes the tool comfortable to use on genuinely personal photographs rather than only on images you would happily post publicly. It is also worth knowing what your files carry independently of Raven. Photographs taken on a phone frequently contain Exif metadata, a container standardised back in 1995, which can include the camera model, the exact timestamp and, if location services were enabled, GPS coordinates. Raven's guess comes from the visible picture rather than those fields, but the fields still travel with the file wherever else you send it. If you want to see what yours hold, or remove them, the free ExifTool utility reads and strips them locally. The broader picture is covered in our overview of AI and photo privacy.

A wrong answer is not necessarily the end of the exercise; it is often information about the photograph. Four things reliably help. Use the original file rather than a screenshot, which adds a second layer of compression over the first. Loosen the crop, since cropping in to help the model usually removes the context that would have narrowed things down. Prefer daylight frames, where colour and shadow direction read cleanly. And when several photographs exist from the same moment, choose the one with the most going on around the edges rather than the best composed one — an awkward frame that happens to catch a street sign in the corner routinely beats an immaculate portrait that crops it away.

Once the mechanics are familiar, the interesting part is the experiments. Our collection of fun ways to test Raven with travel photos sets out several, from a head-to-head guessing duel to a family geography round played on a television. Raven itself lives in a browser, which suits going back through an old camera roll on a laptop. If the curiosity strikes mid-trip instead, the free Geospy AI app for iPhone carries the same idea into your pocket.

That is the whole flow: choose a photo with something to look at, sign in once, upload, and read the guess alongside its confidence rather than as gospel. The first guess is free, so the only real reason to hesitate is deciding which photograph you have always wondered about most.