How we index every video: the descriptors behind what we make next
Godswill Iyke Dave3 min read
Two videos can get the same views for opposite reasons. Unless you know what was in each one, the numbers teach you nothing. So every film is watched and described, the same way every time, before its results are read.
Two reads of every film
A model watches the whole film with its sound, twice. The first read is about structure, mostly the opening seconds. The second is about meaning: the message, the sound and the arc. Both answers are structured, so every film in every company is described in exactly the same terms.
Read one: structure
| Descriptor | What we record |
|---|---|
| Hook type | Question, bold claim, problem, product reveal, face to camera, demo, pattern interrupt, text card, scenic, cold open of a story |
| The first three seconds | What happens, in a sentence; how much movement; whether a face, the brand, the product or text appears; and what is heard: silence, music, a voice, sound effects |
| Pacing | Calm, moderate, fast or frenetic |
| Shots | How many distinct shots |
| People | Most on screen at once, a face in the first three seconds, a close-up, someone talking to camera |
| Setting | Home, office, studio, shop, street, market, nature, vehicle, school, computer-generated, mixed |
| Product | The second it first appears, the share of the film it is on screen, and whether it is shown in use |
| Brand | The second the logo first appears, and how many times the brand is said |
Read two: meaning
| Descriptor | What we record |
|---|---|
| Summary and keywords | What the film is and does, and up to 25 words to find it by |
| Industry, product and audience | The sector, what is sold, and who it seems made for |
| Tone | Warm, funny, dramatic, inspiring, urgent, calm, luxurious, playful, serious |
| Format | Story advert, product demo, testimonial, explainer, short film, music-led, announcement |
| Message | Whether it says one thing, the claim it makes (quality, price, convenience, status, trust, emotion, novelty, social proof) and whether it has an offer |
| Voice | Whether there is one, its language, its tone, and the second it starts |
| Music | Genre, tempo, and its shape: flat, building, dropping, peaking at the end |
| Call to action | Whether there is one, shown or said or both, the second it comes, and its words |
| Emotional arc | For each quarter of the film, how positive and how intense it feels |
| Text on screen | The words that appear |
| Matches the brief | Whether the film says what it was meant to |
The story it was made from
For films our engine makes, we also keep what each one was made from: the brief, the scenes as written and the voiceover. So we can compare what a film was meant to do with what it actually does on screen, and with what it did in the world.
Two embeddings
Each film also becomes two embeddings, one from its description and one from the video itself, each a list of 1,536 numbers placed so that similar films sit close together. That makes questions like show me the films most like our best performer a single search, across thousands of videos.
From descriptors to decisions
Joined to each post's numbers and each customer's journey, the index answers the questions that decide the next brief:
- Does a face in the first three seconds keep this audience watching, or does the product reveal do better?
- Does a price claim beat a quality claim for this product, in this country?
- Does the call to action arrive before most people stop watching?
- Does music that builds lift how much of the film people finish?
- Which hook types bring sign-ups, not just views?
The answers become the next briefs, which become the next films, which are indexed in turn. That loop is why a cycle's tenth week performs better than its first.
Kept apart, company by company
The descriptors and the embeddings are ours. The platforms' numbers belong to the company whose accounts they came from, and they only ever inform that company's work. What our data looks like.
