How does the YouTube algorithm work? How your home feed actually chooses videos
You subscribed on purpose. You picked those channels because you wanted their uploads, yet the home feed surfaces them late, out of order, or not at all. That is not a glitch in your account. It is documented behavior: the home feed is a prediction machine, not a subscription list, and subscribing is not the same as seeing.
How does the YouTube algorithm work? The answer in YouTube's engineering posts and published research is structural rather than mystical. Two stages assemble and score your feed. Three predicted outcomes decide the scoring. Freshness opens a testing window for every upload. And your subscription list barely figures into any of it. This article walks through those mechanics, dismantles the popular myths built on top of them, and closes with the signals a viewer can actually influence and the curation routes that skip ranking entirely.
How does the YouTube algorithm work? A two-stage answer
Consider this the YouTube recommendation system explained in the terms its engineers use: a pipeline of two stages with different jobs, each optimizing for something different. The split exists because the catalog is enormous and fine-grained scoring is expensive. Coarse filtering first keeps the workload manageable.
What appears on the home screen is predicted for each viewer, moment by moment.
Stage one: candidate generation
Candidate generation makes a broad, coarse cut through the catalog. Its job is recall, not precision: gather every video that could plausibly be worth showing this viewer into a candidate pool. The filtering is deliberately rough because a later stage exists to be picky. What matters is that the pool is broad enough to contain good options.
Stage two: ranking
The ranker then scores the candidate pool in fine detail. For each user-video pair, it predicts how this specific viewer will respond: will they click, how long will they stay, and will they consider the recommendation valuable. The highest-scoring candidates become the videos you actually see.
The division of labor is the point. Candidate generation optimizes broad recall; ranking optimizes precision. The architecture has a consequence viewers feel constantly: a video that never becomes a candidate can never be ranked, no matter how well it would have suited you. Some filtering happens before any of your recent behavior can even register. Everything else in this article happens inside one of these two stages.
YouTube home feed ranking signals: what the ranker predicts
Once ranking begins, the system estimates three headline outcomes for every user-video pair:
- Expected click-through rate, or CTR: the share of impressions a video converts into clicks.
- Expected watch time: how long the viewer actually stays.
- Expected satisfaction: whether the viewer will report the recommendation as valuable, captured through direct survey-style feedback.
These are predictions made in advance, not raw averages of how a video performed elsewhere. That distinction is why the same video can sit near the top of one viewer's home feed and never appear in another's: placement is computed per viewer, per video, before anything happens.
No single signal rules the others. The system blends predicted engagement, meaning clicks and watch time, with predicted satisfaction, and the blend is personalized. The inputs are also gathered by different means: clicks and watch time are inferred from behavior, while satisfaction is measured by asking viewers directly.
The practical takeaway is that the set of home feed ranking signals is smaller and more legible than most mythology claims. CTR, watch time, and satisfaction do most of the work. Everything else the feed does is downstream of those predictions.
YouTube watch time vs click-through rate: the tension at the core
CTR rewards the hook. Watch time rewards the payoff. The system predicts both for every candidate, and neither signal is enough on its own.
Click-through rate alone is easy to inflate, and sensational titles and thumbnails are the standard way to do it. That is the classic clickbait problem: a feed optimized on clicks alone would drown in bait. Watch time is the counterweight. In the system's accounting, a click only pays off if the viewer stays, so a video that wins the click and then loses the viewer creates a measurable mismatch.
That mismatch is the documented anti-clickbait dynamic. When clicks run high but watch time disappoints, the system's predictions fail against actual behavior, and future placement adjusts. No human reviewer has to intervene. The prediction error itself is the correction.
The two metrics answer different questions: did the video earn the click, and did it deserve it. For a viewer, the mechanics translate directly into habits. Your click trains the CTR prediction, so clicking similar thumbnails teaches the feed to predict you will click them again. What you do after the click trains the watch-time prediction, so finishing a video and bailing on it early send opposite messages. The feed you see tomorrow is the output of both lessons.
Satisfaction surveys: the feedback clicks cannot capture
Behavioral metrics are blind to regret. A disappointing video still generates a click and still accumulates watch time, so an enjoyable watch and a regretful one can look identical through behavior alone. That is a structural blind spot for any engagement-driven system.
A quick rating can say what a click and a few minutes never could.
Satisfaction surveys exist to close the gap. Instead of inferring value from clicks and minutes, the system asks viewers directly whether a recommendation was worth their time, through survey-style prompts and rating feedback. Those responses train the ranker alongside clicks and watch time as a documented third input.
The existence of the channel matters for interpretation. A system that wanted nothing but engagement would have no reason to ask. Survey feedback is the corrective layer for recommendations that rack up engagement while leaving viewers feeling cheated: clickbait, repetitive formats, videos people click and regret. It is the clearest sign that the home feed's stated goal extends beyond raw engagement.
Freshness and impressions: every upload's testing window
Fresh uploads receive an early batch of impressions, and that exposure functions as a testing window: the system gathers real CTR and watch-time data from actual viewers instead of relying on predictions alone. Freshness is an input in its own right, which is why recent videos earn exploration opportunities precisely because their performance is still unknown.
Each fresh upload receives an early wave of impressions that acts as a live test.
Impressions are the currency throughout. Each one is a chance to earn a click, and the clicks and watch time that follow loop back into future placement. A video's continued visibility depends on how its early data compares with what the system predicted.
This loop explains the surge-then-fade rhythm every viewer recognizes. A new upload appears prominently, collects data, and then either holds its position or slips out of the feed depending on whether it outperformed its predictions. That is not punishment, and it is not promotion. It is the testing window closing.
Where subscriptions actually rank: barely in the home feed
The home feed optimizes for predicted engagement, not subscription fidelity. A subscription is one input among many, and on this surface a weak one: it guarantees neither placement, nor order, nor visibility. An upload from a channel you deliberately followed competes against every other candidate on predicted clicks, watch time, and satisfaction, and it can lose.
Subscription content has dedicated surfaces. Notifications exist to tell you about uploads, and the subscriptions tab lists them. The home feed is a different surface playing a different game: its objective is what it predicts you will watch now, not what you committed to watching when you subscribed.
For power viewers who curated their subscriptions with intent, this gap is the core frustration. The system's best prediction can outrank your explicit choice. It is also why curation-minded viewers eventually look elsewhere: if the home feed will not honor the subscription list, the options are to accept ranked delivery or to build routes that do not depend on ranking at all.
Myths vs documented mechanics
Each myth below collapses against mechanics already covered, so the table only makes the contrast explicit.
The myth | What the documented mechanics show |
|---|---|
Subscribing guarantees home-feed visibility | A subscription is one weak input; the home feed optimizes predicted engagement |
The feed is random or purely chronological | It is a two-stage predictive ranking system: candidate generation, then ranking |
Only clicks matter | Watch time and satisfaction carry real weight alongside CTR, and together they counter clickbait |
The feed is a black box you cannot influence | Clicks, skips, watch behavior, and survey responses all train the predictions |
None of these verdicts requires insider knowledge. Each follows from the two-stage architecture and the three predicted outcomes the system optimizes.
The signals a viewer can actually influence
Sorting influenceable signals from architectural ones sets honest expectations. You cannot reach into candidate generation, because that stage filters before any per-video choice of yours can register. What you can do is steer the ranker, gradually, through three levers.
- Clicks and non-clicks. Every impression you skip quietly teaches the ranker what not to predict for you. Clicking is the louder signal, but a consistent pattern of skipping shapes CTR predictions too.
- Watch behavior. Finishing, abandoning, and re-watching a video all feed the watch-time prediction. Finishing videos you value moves the feed toward them; lingering on content you resent trains it the wrong way. Your watch history is the training data behind these predictions, which is why what actually happens when you clear it matters more than most viewers assume.
- Direct feedback. Survey responses and rating prompts are the loudest lever a viewer holds, because they state value explicitly instead of leaving the system to infer it. The same family of controls includes channel-level options like don't recommend channel, which takes some handling to make it stick.
The levers interact. A click raises the CTR prediction, but the same click followed by an early exit lowers the watch-time prediction, so a single action can pull the feed in two directions at once. And none of this is a switch. Influence arrives as gradual steering of predictions across many interactions, which is why feed changes feel slow. The system is updating its model of you, not executing your instructions.
Curation routes that bypass ranking entirely
For uploads you refuse to miss, ranking-independent routes already exist: the subscriptions tab, notifications, and self-built curation. Understanding the algorithm helps you coexist with the home feed. These routes let you stop depending on it.
Lists, notifications, and self-built feeds sidestep ranking altogether.
Self-built curation is what Yougroup does. The open-source Chrome extension groups channels into themed Lists, such as research, recipes, or interviews, and merges every list into one deduplicated feed, so an upload from a channel that appears in two lists shows up once. The feed is built on YouTube's public RSS uploads, which means no Yougroup account, no hosted backend, and no API key; an optional key adds richer details such as duration and view counts. Playback runs through queues you control: sort by newest, popular, or interleaved, mark uploads watched locally, and open the queue directly on YouTube. The local-first design keeps all data in browser storage, with no server-side sync and no analytics.
If you want a wider playbook for feed control that does not rely on the recommender, start with taking back control of your YouTube feed without the algorithm. For the channels you chose on purpose, curation is the route that guarantees you will actually see them.