← Blog · Growth · Algorithm · YouTube Shorts
How YouTube Shorts actually work: the 6-stage algorithm explained
7 min read · The distribution pipeline that decides your reach in the first 4 hours.
Most creators think YouTube shows their Shorts to everyone and lets the audience decide. That's not how it works. YouTube runs your video through a 6-stage automated distribution pipeline — and your reach is decided in the first 4 hours based on pure behavioral data.
There is no audience vote. There is no slow build where word of mouth carries you. Each Short is tested, measured, and either advanced or frozen by an automated system that never sees a face or reads a comment. Understanding that pipeline is the difference between guessing and engineering your reach.
The 6-stage distribution pipeline
Every Short you publish travels through the same six checkpoints, in order. At each stage the algorithm asks a single question — does the behavioral data justify wider distribution? — and your video either advances or stops.
Ingestion & Filter
Vector mapping
1Target Seed Loop
Stress test
2Niche Expansion
The spike
3Linear Cap Lock
The plateau
4Global Breakout
International feeds
5Evergreen Search
Long-tail indexing
6The diagram shows the one-way path every Short follows: from automated ingestion, through the seed test and expansion spike, into the cap lock, and only then — if the data holds — out to global breakout and finally evergreen search.
Stage 1 — Ingestion and vector mapping (0 to 60 minutes)
The moment you press publish, your video enters a processing sandbox where no human is involved. YouTube runs speech-to-text on your audio and tokenises every spoken word, then runs OCR across your thumbnail and any on-screen text to read what's written. Together these signals build a content vector — a mathematical fingerprint of what your video is about and who it's for.
That vector is what decides which audience receives your video in the seed test. It is assembled entirely from machine reading: the speech, the on-screen words, the visual context. No viewer has seen a single frame yet.
The words you speak and the text on screen in the first 60 seconds determine which audience YouTube targets. This is your content vector.
Stage 2 — The seed cohort stress test (hour 1 to hour 4)
YouTube drops your video into the live feed of 10 to 100 highly targeted people drawn from your content vector. At this stage your subscriber count and channel history carry zero weight — they are not part of the calculation. The algorithm measures one ratio above all others: stayed to watch versus swiped away.
The critical window is 1.5 seconds. If a viewer swipes away inside that window, the system logs a negative interaction vector against your video. If the swipe-away rate across the seed cohort crosses roughly 70 to 80 percent, distribution freezes permanently and the video never advances.
This is why a zero-subscriber channel can outperform a million-subscriber channel on Shorts. The seed test ignores your history entirely.
Stage 3 — The cohort expansion whirlpool (the spike)
If your seed metrics pass the gate, YouTube opens distribution to larger lookalike cohorts modeled on the viewers who stayed. This is what creates the near-vertical spike you see on real-time analytics charts — the moment reach stops crawling and shoots upward.
In this stage the algorithm is testing whether your video holds appeal beyond the initial niche seed audience. Distribution keeps widening cohort by cohort until retention finally breaks down in the broader pool, at which point the expansion slows.
Stage 4 — The cap lock (the plateau)
When swipe rate and completion rate fall below their thresholds, the system freezes distribution. This is not a permanent death sentence — it is a consolidation pause while YouTube recalculates your targeting against the new, broader behavioral data it just collected.
The video then enters a 24 to 48 hour analytics sync before Stage 5 is even considered. During this window your real-time view count flattens, which is exactly what the name "cap lock" describes.
| Video duration | Target stayed to watch | Minimum average % viewed |
|---|---|---|
| 10–15 seconds | above 80% | above 100% (loops) |
| 30–45 seconds | above 75% | above 80% |
| 45–60 seconds | above 60% | above 50–70% |
Most videos never leave Stage 4. The ones that do have both a low swipe rate and a high completion percentage — simultaneously.
Stage 5 — The global breakout vector
If your video keeps performing above threshold after the Stage 4 consolidation, YouTube pushes it into international feeds well beyond your original audience. This is where the hundreds of thousands or millions of views happen — the breakout most creators imagine when they picture "going viral."
Most creators never reach this stage, because Stage 4 caps them first. Breakout is not the default outcome; it is what's left after a video clears every prior gate.
Stage 6 — Evergreen search and browse indexing
Eventually feed distribution stops completely. The video then shifts into long-tail search traffic, recommended sidebars, and desktop browse — the slow, durable trickle that can run for months.
This phase is driven entirely by the keyword density established back in Stage 1 — what you spoke aloud and what appeared on screen. The content vector that targeted your seed audience is the same signal that surfaces you in search a year later.
A real case study — 13 views to 736 views
A video was published and got 13 views. It failed Stage 2 outright — the wrong content vector sent it to the wrong seed audience, and the swipe-away rate told YouTube to freeze distribution before it ever had a chance.
The same video was then republished with three changes: a new title using creator vocabulary, new hook text on the opening frame, and new spoken vocabulary that replaced technical terms with the words creators actually use every day. Nothing in the footage itself was re-shot.
The result was 736 views. Stage 3 fired and the spike happened at hour 4, exactly as the model predicts. The content vector had been fixed, the seed cohort was now correct, and the algorithm did the rest on its own.
The video didn't change. The words changed. And the words determined the audience. That's not a theory — that's a live experiment with documented results.
What this means for your next video
One — your spoken words in the first 30 seconds build your content vector. Use the exact vocabulary your target audience searches for, not the technical terms you'd use to describe your work to a colleague. Their words, not yours.
Two — your first frame is your real thumbnail for Shorts. There are no thumbnails in the feed. Whatever appears at frame zero is what the viewer sees and judges in 1.5 seconds, so it has to earn the stay instantly.
Three — the seed test runs in the first 4 hours, so don't judge your video in the first 15 minutes. The data that decides its fate simply doesn't exist yet, and reacting early tells you nothing.
Four — a high swipe rate kills distribution permanently at Stage 2, and a low completion rate kills it at Stage 4. You need both a low swipe rate and a high completion percentage to pass, and they're measured at different gates.
Five — zero-subscriber channels have the same Stage 2 opportunity as million-subscriber channels. The seed test ignores history entirely, which means every video starts on a level field.
Before your next video goes live — run it through WinRec.
It checks your spoken keyword density, your content vector alignment, and predicts your retention curve before a single person watches it. Try WinRec free.
Try WinRec free at winrec.online.