NEWVectors or files. Pick a path.Start →
    Search & Discovery
    10 min read
    Updated 2026-09-17

    How Do I Find Out Where a Video Came From?

    Someone sent you a clip and you want to know who filmed it, when, and whether it is what it claims to be. Reverse image search on a frame is the first move and it fails more often on video than people expect. This covers what each method actually catches, how to pull the right frames, how to spot recycled footage and AI-generated video, and what to do when the clip is cropped or re-encoded.

    Reverse Video Search
    Video
    Verification
    Content Provenance
    Copyright
    Search & Discovery

    How do I find out where a video came from?



    Pull several still frames out of the clip and reverse image search each one. That single move solves most cases, because the web indexes images far more densely than video, and a frame from a widely shared clip usually appears somewhere as a thumbnail, an article image or a screenshot. Search at least one frame from the opening, one from the middle and one from the end, since a repost often trims the part the original is known by.

    When frames return nothing, the clip is either rare, heavily re-encoded, or genuinely new. At that point you switch from matching pixels to reading the video itself: what the audio says, what text appears on screen, what the weather and shadows imply, and whether the file carries any provenance data.

    What actually works, and where each method breaks



    MethodWhat it catchesCostWhere it fails
    Reverse image search on framesClips that circulated as articles, thumbnails or screenshotsFreeHeavy crops, re-encodes, or footage that never appeared as a still
    Keyword search on on-screen text and audioAnything with a caption, chyron, sign or spoken nameFreeSilent clips with no text, or languages the transcriber handles badly
    Video fingerprint matchingNear-duplicates across a library you control or a service indexesFree to moderateOnly finds copies inside the index it searches, so the open web is mostly out of reach
    Provenance metadataFiles that kept their signed capture recordFreeNearly every social upload strips or never had it
    Paid monitoring servicesOngoing appearances across platforms they coverSubscriptionCoverage gaps, and they answer "where is it now" rather than "where did it start"
    No single row is sufficient. The clips that resist the first method usually fall to the second, and the ones that resist both are the ones worth being careful about.

    Steps to trace a clip back to its source



    1. Save the highest-quality copy you can get. Screen-recording a playing video loses detail that frame matching depends on, so download the file where the platform allows it. 2. Extract frames rather than screenshotting one. Take a frame at the start, at each obvious scene change, and at the end. A scene change is where the visual content shifts, and it is where the most distinctive frames live. 3. Reverse image search each frame across more than one engine. Google Lens, TinEye and Yandex index different corpora and routinely disagree, so a miss on one is not an answer. 4. Read every piece of text the video shows you. A chyron, a street sign, a jersey number, a license plate format, a storefront. Search those as plain words, because text search reaches pages that image search does not. 5. Transcribe the audio and search distinctive phrases. A spoken name, a place, or an unusual sentence is often the fastest route to the original post. 6. Check the file for provenance data before assuming there is none. Content Credentials travel with some camera and editing-tool output and name the capture device and edit history. 7. Date the footage from the world inside the frame. Leaf cover, snow, sun angle, visible construction, signage for an event, or a model of car that did not exist yet. This is what separates new footage from an old clip being recycled.

    Why does reverse image search on a frame often fail?



    Because you are asking a system built for images to recognise a frame that has been through several rounds of compression. Every upload and download re-encodes the video, so the exact pixels of your frame have never existed anywhere else. Perceptual matching tolerates some of that, and it stops tolerating it when the clip has also been cropped, mirrored, letterboxed, sped up, or had a caption bar burned across it. Each of those is common on social reposts, and they stack.

    The practical response is to give the matcher more chances. More frames, from more points in the clip, across more engines. A single frame that returns nothing is weak evidence of anything.

    How do I tell if the video is old footage being recycled?



    This is the most common form of misleading video, and it usually needs no technical tooling at all. The footage is real and the claim about when and where it happened is false.

    Look for a date the clip cannot argue with. Bare trees in a story set in summer. A building that was demolished. A phone model in someone's hand. A shop sign that changed hands. Then search the event the clip is being attached to, and see whether the scene matches the place it supposedly shows. If a frame search returns a hit from years earlier, that is your answer, and it is the single fastest one.

    How do I tell if the video is AI-generated?



    Generated video has moved past the obvious tells, so a checklist of artifacts ages badly. Two things hold up better.

    The first is provenance. Several generation tools attach Content Credentials, and a file that carries a signed record of how it was made settles the question directly. Absence of that record proves nothing, since the vast majority of ordinary video has none.

    The second is corroboration. A real event filmed by a real person almost always leaves more than one trace: another angle, a local report, a named location that exists, a person who can be found. A generated clip tends to be the only artifact of the thing it depicts. If a dramatic scene has exactly one source and no independent record, treat that as the signal, ahead of any pixel-level tell.

    Our guide on searching and managing AI-generated video and images covers the indexing side of this for teams handling generated content at volume.

    What if the clip is cropped, re-encoded, or has a watermark burned in?



    Crop the interference back out where you can. If a platform bar covers the bottom third, crop to the clean region and search that. If the clip is letterboxed, remove the bars first, since they change the frame's proportions and can throw a perceptual match.

    A burned-in watermark is often useful rather than an obstacle. It names a platform, an account, or an aggregator, and that is a place to go looking. The watermark tells you where the copy you have came from, which is frequently one step away from where it started.

    For the mirror image of this problem, where the video is yours and you are looking for people who took it, see how to find out if someone reposted your video. If the question is whether you are allowed to use a clip you found, see how to check if a video is copyrighted.

    How do teams do this across a whole library?



    Everything above scales to one clip and an afternoon. It does not scale to an archive, a daily feed of user uploads, or a moderation queue, because the work is manual and the answer is only as good as the frames a person happened to pick.

    At volume the method changes shape. Instead of searching the open web one frame at a time, you index your own content once and query it: split each video into scenes, embed representative frames, transcribe the speech, read the on-screen text, and keep all of it in one searchable index. A new clip then becomes a query against everything you already hold, and the answer comes back with a timestamp rather than a page of links.

    Mixpeek is built for that layer. It does token-level multimodal indexing and retrieval directly over your object storage, so your video stays in your own bucket and the index is built beside it. A single pipeline pulls scenes, frame embeddings, transcripts and on-screen text out of each file, and one query reaches all of them. The mechanics of the matching step are in our explainer on how reverse video search works, the frame-selection tradeoff is in video frame sampling for embeddings, and the segmentation step is in video scene segmentation.

    If you are choosing a tool rather than building the layer, the best copyright detection tools and the best video search tools compare the options on what each one actually catches.

    Frequently Asked Questions



    Can I reverse search a whole video file instead of frames?



    Not on the general web. Google, TinEye and Yandex index images, so a video has to be reduced to frames before any of them can help. Services that match whole videos exist, and they search their own index rather than the open web, which means they answer "is this in my library" rather than "where on the internet did this start".

    How many frames should I search?



    At least three, and more if the clip has distinct scenes. One from the opening, one from each obvious scene change, and one near the end. Reposts frequently trim the beginning or the end, so the frame that matches the original is often not the one you would pick first.

    Does the video's metadata tell me where it was filmed?



    Almost never, once it has been through a social platform. Platforms strip or rewrite file metadata on upload, so the copy you downloaded carries the uploader's processing rather than the camera's record. Signed Content Credentials are the exception, and they survive only when every tool in the chain preserved them.

    The frames return nothing anywhere. What does that mean?



    Three possibilities, and they are not equally likely. The clip may be genuinely new and unpublished. It may have been altered enough that perceptual matching fails. Or it may exist only inside places search engines do not index, such as private groups and messaging apps. Move to the audio and on-screen text before concluding it is new.

    Is any of this proof?



    No. Frame matches, absent metadata and world-detail dating are all evidence, and each one can be wrong on its own. Treat a single method as a lead and look for a second, independent line that agrees with it. Organisations that do this professionally, such as the open-source investigation community, work that way for exactly this reason.

    Further reading



  1. InVID and WeVerify verification plugin, a browser toolkit that extracts keyframes and runs multi-engine reverse searches
  2. Content Credentials and the C2PA standard, the provenance record some cameras and editing tools attach
  3. TinEye and Google Lens, two reverse image engines with meaningfully different corpora
  4. Bellingcat, open-source investigation methodology applied to real cases
  5. Managed Mixpeek

    Put multimodal search to work

    Connect a bucket and Mixpeek runs the whole multimodal search pipeline for you: extraction, indexing, and search over your own objects. No models to wire up, nothing to host.

    Start with Managed
    MVS · bring your own

    Already have vectors?

    Keep your embeddings on your own cloud and run dense, sparse, and BM25 search directly on object storage. From $25/mo.

    Start with MVS

    Automate Copyright Detection

    Stop checking content manually. Mixpeek scans images, video, and audio for IP conflicts in seconds.

    Try Copyright CheckLearn About IP Safety