How to Find Where a Video Was Filmed

A ten-second phone clip at 30 frames per second is 300 photographs of the same place, plus a soundtrack. That makes video a richer source than any single photo — once you get past the fact that the copy in front of you is rarely the original file.

The method below runs from the file outwards: metadata if you have the original, then what the platform shows, then frames, searches, clues and a final check. Each step is cheaper than the one after it, and any of them can end the search.

1. Check the original file for GPS tags, if you have it.
2. Read what the platform shows: location tag, caption, comments.
3. Pull sharp frames with FFmpeg or a paused screenshot.
4. Reverse search the best frames for earlier uploads.
5. Read the clues across frames, and listen to the audio.
6. Run the clearest frame through AI geolocation for a probable city.
7. Verify against map imagery before you conclude.

The camera needs permission first. On iPhone that is Settings → Privacy & Security → Location Services → Camera; Android camera apps have their own location or geotag switch. Apple’s tag holds an ISO 6709 coordinate string, and Android’s developer documentation says its recorder stores geodata in the udta box in the same ISO 6709 format.

Read it with ExifTool

ExifTool is free for Windows, macOS and Linux. These three commands were tested with version 13.59:

• exiftool -G1 -s "-*gps*" "-*location*" clip.mov lists every location tag, whichever box the phone used. [Keys] in the output marks iPhone-style tags and [UserData] Android-style ones. iPhones also store LocationAccuracyHorizontal, the phone’s own estimate of its error in metres.
• exiftool -n -GPSPosition clip.mov prints plain decimal coordinates such as 48.8584 2.2945, ready to paste into a map.
• exiftool -CreationDate clip.mov shows an iPhone’s capture time with its clock’s UTC offset, such as +02:00.

That UTC offset is worth a glance even when GPS is missing. It is a time zone rather than a place, and a phone left on its home zone will mislead, but it can rule out most of the world in seconds.

Already installed FFmpeg for step 3? Its companion ffprobe shows the raw value: ffprobe -v error -show_entries format_tags clip.mov. Look for a string like +48.8584+002.2945+035.000/ — latitude, longitude, then altitude in metres, with plus meaning north or east.

Dashcams, drones and action cameras

These often record a track — a new position every second or faster — rather than one point. ExifTool currently reads 124 types of this timed GPS, but only when asked. If it warns that the ExtractEmbedded option may find more tags, run exiftool -p gpx.fmt -ee3 clip.mp4 > track.gpx to save the route as a GPX file that most mapping tools open. The gpx.fmt file ships in the fmt_files folder of the full ExifTool download and must sit in the folder you run the command from.

YouTube’s help pages describe the process: a new upload is first processed in low quality, and higher resolutions follow later. Viewers stream fresh encodes, and the location tags in the creator’s file are simply not part of them. TikTok and Instagram re-encode uploads too, which is also why photos lose their location when shared.

So a “video location checker” cannot recover GPS from the copy you see; at best it repeats what the platform publishes about a post. If you are verifying a clip someone sent you, ask for the original file, sent as a document or a cloud link rather than as an ordinary video message. The publish-versus-transfer rule decides whether metadata survives, for video as for photos.

• Location tag. TikTok and Instagram let a poster attach a place when publishing, and tapping it lists other public posts tagged there. On YouTube, creators set a video location in YouTube Studio; it shows on the watch page, and clicking it searches for public videos from the same location.
• Caption, description and hashtags. Travel and street clips are usually labelled by the people who filmed them.
• Comments. Viewers are quick to say “that’s the corner by the station” — sometimes wrongly, but it gives you a hypothesis to test.
• Spoken words. On YouTube, expand the description and choose Show transcript, then search it for place names that never made it into the caption.
• Watermarks. A repost often carries the original uploader’s watermark or handle, and the first upload may still have the caption and tag that the repost dropped.

Treat all of it as testimony. People tag the nearest landmark, the city rather than the street, or somewhere they never were, and the Instagram guide explains how far a tag deserves trust.

A screen recording is fine: every method from here on reads pixels, not metadata. Set playback to the highest quality first, and hide captions and controls where you can — they sit on top of the very pixels you want to read.

FFmpeg is free and handles the rest. Each command below was tested with FFmpeg 8.0.1 on real street footage:

1. Step 1Sample one frame per secondffmpeg -i clip.mp4 -vf fps=1 frame_%03d.png writes frame_001.png, frame_002.png and so on. For a long video, fps=1/5 takes one frame every five seconds.
2. Step 2Grab the keyframesffmpeg -i clip.mp4 -vf "select='eq(pict_type,I)'" -fps_mode vfr key_%03d.png keeps only the frames stored as complete pictures. It worked on both H.264 and VP9 files, while the often-quoted -skip_frame nokey shortcut silently exported every frame of a 60 fps VP9 video — 1,800 images from 30 seconds.
3. Step 3Catch the cuts in edited clipsffmpeg -i clip.mp4 -vf "select='gt(scene,0.3)',showinfo" -fps_mode vfr cut_%03d.png saves the first frame after each cut and prints its time as pts_time. FFmpeg’s documentation calls a threshold between 0.3 and 0.5 a sane choice. On a continuous 65-second walking shot it found nothing, so keep it for montages.
4. Step 4Export the best momentffmpeg -ss 00:00:12.4 -i clip.mp4 -frames:v 1 best.png saves the frame at 12.4 seconds. Nudge the time by a few tenths of a second to dodge motion blur.
5. Step 5Make a contact sheetffmpeg -i clip.mp4 -vf "fps=1/3,scale=480:-1,tile=5x4" -frames:v 1 sheet.png puts one frame from every three seconds of the first minute on a single image — the fastest way to spot which moments deserve a closer look.

PNG avoids a second round of compression. In our tests on downloaded street footage, PNG frames came to 1.5–2.6 MB at 1080p and 3.9–5.1 MB at 4K, and the same frames saved as .jpg with -q:v 2 stayed under 1 MB — the fallback if a camera original ever produces a PNG over the analyser’s 10 MB limit.

No command line? On YouTube, pause and press the full stop or comma key to step one frame forward or back, then screenshot in full screen. The free InVID-WeVerify plugin for Chrome and Firefox splits a linked or uploaded video into keyframes and sends any of them to reverse image search.

Reverse search matches images, not places, so it answers a narrower question: has this scene, or this exact footage, been published before? For viral clips the answer is often yes, and an earlier upload under a different caption is both a location lead and a warning that the current caption may be wrong.

Video gives you several attempts where a photo gives you one. Search three or four different moments — the wide opening shot, the clearest sign, the oddest building — because each is a separate chance to match. A YouTube thumbnail is worth a search too, although a custom one may not be a frame from the video at all.

Which engine suits which subject matters, and Google Lens and Yandex tend to fail on different scenes. On a phone, the mobile routes need no app.

Read each frame as you would a photo: script and signage, road markings, licence plates, architecture, vegetation, then the light. The country clues guide explains why regulated things such as road markings outrank landscape. What video adds is movement and repetition:

• Direction of travel. In a still, a stationary car may be parked facing either way. In a video, two-way traffic in motion shows the driving side beyond doubt, even when the clip is blurry — just avoid one-way streets.
• The same sign, several times. Glare, blur or a passing bus may hide text in one frame and reveal it in the next. Step through frame by frame before giving up on it.
• Pans and turns. A camera that sweeps round shows the hills, the coastline or the skyline that a single frame leaves out.
• The sun across the clip. As the camera turns, you see which way shadows fall along the street. Shadows give a hemisphere and a rough time of day when vegetation cannot.

Listen to the soundtrack

First make sure the audio came from the scene. On TikTok and Instagram, the audio label should read “original sound” or “Original audio” with the poster’s own name; a song title or another account’s name means the sound was borrowed. Even original audio can be dubbed before upload, so treat music and voice-over with suspicion.

• Language and accent. A language narrows the list of countries at once, and a dialect can narrow it to a region.
• Announcements. Station, platform and airport announcements name places by design. Transcribe the stop and search for it.
• Signature sounds. Emergency sirens differ by country, and a call to prayer, a tram bell or church bells each narrow the region.
• Electrical hum. Mains power runs at 50 Hz across most of the world and 60 Hz in North America and a few other countries, such as Brazil and South Korea. A steady tone at either frequency in a spectrogram splits the candidates in two — when it survives, since phone noise reduction often removes it.

Choose the frame with the most kinds of evidence in it at once — road, buildings, signs and vegetation together beat a close-up of any one of them. Crop off captions, subtitles and platform buttons first; they cover evidence and add none.

Then run a second and a third frame from different moments. If they point to the same city, confidence goes up; if they scatter, the scene is ambiguous and the honest result is a region. How the model reads a scene explains why some frames work and others do not.

Matching the clues that led you to a place proves nothing, because you already knew they agreed. The test is whether unused details line up too. Video makes this easier than a photo does: take a frame from ten seconds later and check that the next building along is the right one.

Timing is the other half. Footage recycled from an older event is common, so the earliest copy from step 4, the state of the trees and the direction of the light should all fit the claim. The OSINT workflow sets out the full discipline, including what to record, and the verification guide covers checks that work frame by frame.

Where imagery is thin — Street View coverage is uneven — the right conclusion is a region with a stated confidence, not a street.

Two independent sources that agree — a tag and a Street View match, say — make a finding; one on its own is a lead. For still images, the main photo guide compares the same methods side by side.