I’ve seen this tip floating around affiliate circles for a while now. The idea is that if you embed the right keywords into your video file’s metadata (the title, description, tags baked into the MP4 container) Facebook’s algorithm will read them and use that to better target your ads. Better targeting means lower CPMs, better delivery, higher ROI. Sounds logical. And given how black-box Meta’s system is, it’s the kind of theory that spreads fast because it’s impossible to definitively disprove in your own account.
So I dug into it properly. Here’s what I found.
What Facebook Actually Does When You Upload a Video
First, understand that Facebook does run your video through a detection stack the moment you upload it. This isn’t just a spam check. It’s a multi-layer fingerprinting system that Facebook actually open-sourced parts of back in 2019 (called PDQ and TMK+PDQF). When your video hits the servers, it gets checked at three levels simultaneously.
The first is file metadata: your EXIF data, creation timestamps, device info. Stripped or missing metadata actually flags the file as a re-upload rather than original content. The second layer is a perceptual hash, a compact fingerprint of the visual content that survives basic edits like resizing, color adjustments, or adding a watermark. The third and hardest layer to fool is AI feature embeddings, where deep learning models extract the semantic meaning of your content. Two videos that look nearly identical to human eyes can have completely different embeddings, but more importantly, two videos that look different can still match if they’re derivatives of the same source.
So yes, Facebook reads your metadata. The question is what they actually do with it for targeting purposes.
The Keyword-in-Metadata Theory: Why It Doesn’t Hold Up
Here’s the honest answer: Facebook’s ad delivery algorithm doesn’t use your MP4 container tags (the title, description, keywords you write into the file’s metadata fields) to make targeting decisions.
Meta’s targeting engine (they call the core system Andromeda) runs on behavioral signals. What users have watched, clicked, bought, how long they scroll, what they search, what apps they use, what websites they visit, what they’ve done across Instagram and Facebook and WhatsApp. It’s processing hundreds of millions of these signals in real time. The idea that it would also check whether your video file’s internal metadata says “auto insurance” or “cheap car insurance” before deciding who to show your ad to doesn’t match how the system actually works.
If you’ve tested this and seen results, what you’re likely measuring is something else: creative quality, audience warmth, offer timing, or just normal variance.
Where the “Unique Hash” Idea Has Real Legs
This is where it gets interesting, because there is a legitimate reason to make every video file unique. It’s just not what most people think it is.
If you’re scaling horizontally across multiple fan pages (which is the right move if you’ve been following what actually works with $1 ABO strategies), and you upload the same MP4 file to ten different pages, Facebook’s detection system can and does identify it as duplicate content. Pages that consistently push duplicate content see overall reach decline, not just on the flagged posts but across everything. It’s a page-level penalty, not just a post-level one.
Surface-level edits don’t fix this. Slapping a filter on it, trimming a second from the end, adjusting the speed, none of these fool perceptual hashing or the AI embedding layer. The only way to actually produce a unique file in the eyes of Facebook’s system is to change it at all three detection levels: metadata, pixel-level hash, and feature embeddings.
Practically, this means re-encoding with slightly different compression settings, injecting authentic-looking device metadata (camera model, GPS coordinates, capture timestamps), and making minor but real pixel-level modifications that shift the perceptual hash without degrading visible quality. Tools like MetaGhost are built specifically for this. You can also DIY it with FFmpeg if you know what you’re doing: re-encode with different bitrates, add subtle grain, adjust color balance at levels below what’s visible to the eye.
This matters a lot if you’re running the same winning creative across many pages at scale. It’s not about tricking the targeting algorithm. It’s about not getting your pages penalized for what Facebook classifies as duplicate content distribution.
What Actually Moves the Needle on CPM and Targeting
Since we’ve established that file metadata isn’t a targeting lever, here’s what is.
Creative quality is the biggest CPM variable you control. Meta’s auction model factors in your estimated action rate, basically how likely someone is to click, buy, or convert based on your ad. A creative that earns engagement drives that number up, which means you compete better in auctions without needing to bid higher. Lower effective cost per result, lower CPM.
Broad targeting consistently outperforms tight targeting right now. The data is pretty consistent on this: narrow audiences drive CPM up because you’re competing harder for a smaller pool, and Meta’s algorithm needs room to find the right people. The current best practice is to go broad and let the creative do the filtering.
Watch time and completion rate feed into delivery scoring. Facebook updated its video algorithm in mid-2025 to weight completion rates and audience retention more heavily. A video that gets watched to 75% signals relevance, and the algorithm rewards that with better delivery. This is worth knowing when you’re structuring your video ads. Front-loading value and getting to your hook fast matters beyond just the first-impression CTR.
There’s one exception where “keywords in the video” does actually matter. Facebook’s AI does audio transcription and OCR on your video content. The spoken words, on-screen text, and captions are processed and used for content classification. This is different from the file container metadata. It’s the actual content of the video. If your video ad is about auto insurance and someone says “get cheaper car insurance” or the text on screen mentions it, that feeds into how Facebook classifies and distributes the content. That’s a real signal, and it’s worth thinking about.
The Practical Summary
If you’re managing volume, lots of ad sets, multiple fan pages, reusing winning creatives across accounts, here’s what’s worth your time:
Make each video file genuinely unique at the hash level when you’re distributing the same creative across multiple pages. This prevents duplicate content suppression, which is a real penalty, especially as pages age and accumulate history.
Don’t bother editing MP4 container metadata fields with keywords hoping it’ll improve targeting. Facebook’s targeting engine doesn’t work that way.
Do pay attention to what’s actually said and shown in your video. Audio and on-screen text feed real classification signals.
And keep doing what moves the real numbers: broad audiences, high-quality creative that earns watch time, strong CAPI/pixel setup, and feeding the algorithm real conversion signals so it can find more of the right people.
The file metadata theory makes intuitive sense. It’s the kind of hack that sounds right. But it’s chasing a signal that isn’t there. The duplicate hash issue, on the other hand, is real infrastructure work worth building into your process if you’re running at scale.
