An AI viral score is not magic. It is a structured guess at how a real viewer will react in the first 10 seconds, calibrated against thousands of videos that did and did not pop.
What gets measured
- Hook strength: pattern, specificity, and time-to-payoff.
- Pace: cut frequency, dead frames, energy continuity.
- Clarity: can a viewer explain the video in one sentence after watching once?
- Share trigger: presence of a quotable line or surprising claim.
- Caption-hook alignment: does the on-screen text reinforce or fight the spoken hook?
Why scores move between runs
Large language models are non-deterministic by design. Two runs on the same clip might score 78 and 82. That is not a bug. It is the model sampling near a real signal. Treat any single score as ±5 points. Trends across edits matter more than absolute numbers.
How to use the score
Score before you record. Score after edit one. If the number does not move at least 8 points, you fixed the wrong thing. Use the feedback to target a specific weakness, then re-score. Two iterations is usually enough.