How AI Clip Scoring Works: Six Signals That Predict Short-Form Performance

The most important three seconds you will ever edit

A short-form clip does not live or die by its length. It lives or dies in the first three seconds. Viewers who skip at second two never see the resolution, never hear the punchline, never click through. The hook is not an introduction — it is the entire audition.

This is the premise behind DecaTrend's clip scoring system. Rather than ranking clips by length, resolution, or how they look in a grid, the Virality Engine evaluates each clip across six discrete signals. Each signal targets a different moment in the viewer's decision to keep watching.

The six signals

Hook strength measures how sharply the clip opens. A hook scores high when the opening frame contains motion, a recognizable face, or speech that begins mid-thought — something that creates a question in the viewer's mind before they consciously decide to pay attention. Clips that open on a static title card or silence score low regardless of how good the rest of the content is.

Retention potential estimates how likely a viewer is to watch past the fifteen-second mark. The model uses pacing — how quickly the scene changes, how often the speaker pauses, whether there are visual or audio cuts that reset attention. A clip with a 45-second static wide shot of someone speaking can score high on hook and low on retention. That is useful information.

Emotional arc looks at whether the clip moves through a recognizable emotional shape: tension to relief, confusion to clarity, calm to surprise. Flat emotional content — a speaker explaining something at one steady pace and tone — scores lower than content with a clear pivot. This is not about being dramatic. It is about change. Viewers remember change.

Caption density tracks how much of the clip has readable caption coverage. Platforms auto-mute video for most users on the first playthrough. A clip where captions appear on-screen for less than 60% of its runtime is invisible to muted viewers. High caption density is not a creative choice at this point — it is table stakes.

Audio energy measures the average RMS loudness, dynamic range, and whether the audio track has any significant dead air. Low audio energy often correlates with a recording that sounds like it was captured on a laptop microphone in a reverberant room. The score does not penalize quiet content — it penalizes content where the audio gives viewers a reason to scroll.

Trend fit compares the clip's content type, format, and pacing against the current signal library. If the clip matches an archetype that is gaining traction — a "confession + pivot" structure, a single-question interview format, a silent-process video — it scores higher. Trend fit is the most time-sensitive signal and the one that changes most between batches.

How DecaTrend computes the score

Each signal produces a normalized value between 0 and 1. Those six values are combined into a weighted composite score between 0 and 100. The weights are not equal — hook and retention carry more weight than caption density because a clip with a weak hook will not be watched long enough for captions to matter.

The signal weights are tunable per content type. A gaming clip and an interview clip are evaluated differently because the signals that predict performance differ by format. A gaming clip with low speech density is not penalized for caption coverage in the same way a talking-head clip would be.

The composite score is computed per-clip, per-batch. If you upload the same video twice, the scores may differ slightly if the trend signals have shifted between uploads.

Why no model can guarantee a viral clip

The score predicts the probability of strong early performance based on the signals that have historically mattered. It does not predict virality, because virality is not predictable. It is the product of the right clip reaching the right audience in a window when that content is novel — and novelty is not something a scoring model has access to.

The score is most useful as a prioritization tool. If you have 12 clips and enough time to caption and post three, sort by score and start at the top. The score does not tell you a clip is good. It tells you which clip has the best structural ingredients for strong short-form performance given what the model knows right now.

What to do with the score

Post the highest-scoring clip first and treat the next 48 hours as a test. If it significantly underperforms — less than half the reach of your last ten clips — the trend fit signal may be stale or your hook may not be landing despite a high score. Recut the opener by trimming the first two seconds and repost.

Use low-scoring clips as raw material for montages rather than standalone posts. A clip that scores 42 may still contain a great 8-second moment that belongs in a highlight reel.

Do not discard clips below 60. The score is not a binary pass-fail. A 62 clip with strong personal brand alignment for your specific audience can outperform an 88 clip that targets a general trend your audience does not respond to.

The score is one input. The creator is still the editor.