What 756,000 Short-Form Videos Told Us About What Actually Works

We analysed engagement signals across 756,000 short-form videos from TikTok, Instagram Reels and YouTube Shorts — both AI-generated and human-shot — alongside the internal corpus we built while operating Reviral. The patterns below are the ones that showed up over and over, not the "hot take" advice you'll see on creator Twitter.

The signals that actually correlated with reach

Five attributes correlated with above-median view counts across every niche we measured. None of them are surprising on their own — what's surprising is how consistent they are.

1. Hook framed as a question (+22% lift)

Videos whose first three seconds posed an explicit question ("Why do pros never use this grip?", "Which one is fake?") outperformed declarative hooks by ~22% on retention through second 5. Questions create a curiosity gap that holds attention long enough for the algorithm to register a strong watch-through.

About 12.2% of the corpus used a question hook. Replicating that shape is the single highest leverage script change a creator can make.

2. 9:16 vertical (78% of distribution)

78% of high-performing short-form was vertical. Square (1:1) and landscape (16:9) underperformed on every platform — even YouTube, which historically rewards landscape, now favors 9:16 in its Shorts feed.

3. Animated captions on by default (86.6%)

86.6% of high-performing short-form had word-level animated captions. The lift is partly accessibility (sound-off viewing) and partly motion — captions add visual rhythm even when the underlying clip is static.

4. Faceless content holds its own (55.3%)

Conventional advice says you need to show your face for the algorithm to trust you. The data disagrees: 55.3% of the corpus was "faceless" (B-roll, screen recordings, motion graphics, AI-generated characters) and faceless videos performed within 4% of face-on-camera on average. The actual differentiator was script quality and hook strength — not whether a human face was on screen.

5. Average length is longer than you think (115s)

The much-repeated "15-second sweet spot" doesn't survive the data. Average duration in the top-performing decile was 115.5 seconds. Sub-30-second videos were 16.3% of the sample but accounted for less than 9% of the high-performers. Long-enough-to-tell-a-story beat short-enough-to-be-snackable consistently.

What didn't correlate

The signals that didn't matter are almost more interesting than the ones that did. We saw no meaningful correlation between performance and:

• Posting time of day (within +/-2 hours of a niche's peak window)
• Hashtag count (1, 5, or 30 produced statistically indistinguishable results)
• Music genre (a strong hook beat the "trending audio" meta in 7 of 9 niches)
• Filter / colour grade (creators who switched grades mid-series didn't see retention drops)
• First-comment SEO (the long-debunked "seed the algorithm" tactic)

What we built around the data

Reviral's defaults are derived directly from this analysis. New users land on 9:16, animated captions ON, 30–90s duration, and a hook prompt that biases toward question-form. Every preset comes with a script style picker that maps to one of 13 proven openers (curiosity, contradiction, promise, listicle, story, …) rather than letting users build hooks from scratch.

We don't think the data lets you skip taste — you still have to pick a topic that resonates with your audience. But the defaults remove the categorically wrong choices, so the creator can focus on the parts that actually require judgment.

Try the cinematic preset with default settings and see how the data shows up in the output.