How do I find what makes my best-performing video ads work?
Break every ad into scenes, tag what each scene shows, attach each ad's results, and compare your winners with the rest of the ads that ran under the same objective and budget. The attributes that keep showing up in winners and rarely in the others are your candidates: a face in the first seconds, the product on screen early, a spoken offer, a particular setting. Treat that list as a hypothesis and confirm it with an A/B test, because the ads that won also got more spend, and spend alone changes results.
What a list of top ads misses
Watching your five best ads shows what they have, and the other 95 ads may have it too. If every ad in the account opens on a face, a face in the hook explains nothing. A pattern only counts when it separates winners from the rest, which means you need the same tags on every ad, including the losers.
Two more things distort a top-ads list:
What should I tag in each video ad?
Tag things you can see or hear and could change in the next ad. Each tag should be per scene, with a timestamp, so you can ask when something happened as well as whether it did.
| Attribute | What to record | Why it matters | How to tag it at scale |
|---|---|---|---|
| Hook | What is on screen and said in the first seconds | Attention is decided early; Google's ABCD research puts it first | Scene split plus a description of the first scene |
| Product on screen | Whether and when the product or packaging appears | Separates ads that show the product from ads that only talk about it | Visual search for the product, or a taxonomy of product categories |
| People | Faces, how many, talking to camera or not | Testimonial and creator styles behave differently from product shots | Face detection per scene |
| On-screen text | Offer, price, captions | Many feeds autoplay muted | OCR per frame |
| Voice and sound | Voiceover, music, spoken offer | Sound-on and sound-off placements differ | Transcription of the audio track |
| Setting | Kitchen, street, studio, outdoors | Context can carry the message | Scene description or a taxonomy of settings |
| Call to action | What the ad asks for and when | The last step before a click | Transcript and OCR near the end |
| Format | Length, aspect ratio, number of cuts | Pacing differs by placement | File metadata and scene count |
Google's ABCD framework (Attention, Branding, Connection, Direction) is a useful checklist for this table. Google reports that ads following it showed a 30% lift in short-term sales likelihood and a 17% lift in long-term brand contribution in its work with Kantar (Think with Google).
Step by step
1. Export ads with their results. For each ad: an id, the video file, the copy, the campaign objective, placement, impressions and the metric you care about. Ad platforms export this from their reporting tools. 2. Split each video into scenes. Scene boundaries let you say "the product appears in scene 2 at 0:04" instead of "the ad contains the product". 3. Tag every scene. Use the attributes above. A fixed list of labels (a taxonomy) keeps tags comparable across hundreds of ads; free-text descriptions are useful for exploring but hard to count. 4. Group similar scenes. Clustering scenes by what they show surfaces patterns you did not think to tag, such as a recurring "unboxing on a kitchen counter" shot across brands. 5. Compare winners with the rest. Within one objective and format, mark the top group by your rate metric. For each tag, compare how often it appears in winners and in the others. Look for large gaps backed by many ads. 6. Test the pattern. Make two versions of one ad that differ only in the attribute, and run them as an A/B test, which Meta, Google and TikTok all support (Meta's A/B testing guide).
Common mistakes
Doing it with Mixpeek
Mixpeek indexes video ads from object storage and keeps every scene searchable, tagged and grouped, so the comparison above runs over hundreds of ads without anyone watching them. The Creative DNA template sets this up in one apply:
The template stores the creative and its tags. Your results stay in your ad platform: join them to the scenes on the ad id, then run the winners-versus-rest comparison above. See pricing for what indexing costs per minute of video.
Frequently Asked Questions
How do I find what makes my best-performing video ads work?
Tag every scene of every ad with the same attributes (hook, product on screen, people, on-screen text, sound, call to action), join each ad's results, and compare the top ads with the rest inside one objective and format. Attributes that appear far more often in winners are candidates; confirm each one with an A/B test.
What should I analyze in a video ad's first seconds?
What is on screen, whether a person or the product appears, whether there is text or a spoken line, and how fast it cuts. Google's ABCD framework lists attention as the first principle and suggests hooking the viewer from the start (Think with Google).
How many ads do I need to find a real creative pattern?
Enough that each group you compare has dozens of ads with enough impressions to trust their rates. With a dozen ads, chance differences look like patterns. Below that, use the tags to choose A/B tests rather than to draw conclusions.
Can AI tag video ads automatically?
Yes. Scene detection, multimodal embeddings, face detection, OCR and transcription cover most attributes in the table above, and a taxonomy turns them into fixed labels you can count. Spot-check a sample by hand before trusting the counts.
Why do my winning ads stop working?
Audiences see them repeatedly and respond less, and the platform keeps spending on them. Knowing which attributes made an ad work lets you build the next one with the same ingredients instead of rerunning the old one.