The YouTube Algorithm Explained: What It Actually Optimizes For

Almost everything creators argue about online comes down to a misunderstanding of one thing: what YouTube's recommendation system is actually trying to do.

Once you understand that, most of the contradictory advice sorts itself out. You stop asking "does the algorithm like long videos?" and start asking the only question that has an answer: what is this system predicting, and what evidence does it have?

There is no "the algorithm"#

The first correction is grammatical. There is no single algorithm. There are several systems doing different jobs on different surfaces, and they disagree with each other constantly.

The home page recommender, the suggested-videos recommender, the search ranker, the Shorts feed and the notification system are separate. A video can be invisible on one and thriving on another. When someone tells you "the algorithm changed", what they usually mean is that one surface started behaving differently for one type of content — which happens all the time and is rarely a policy decision aimed at you.

This distinction is not pedantic. It changes what you do. If videos were pushed, your job would be to please a system. Because they are pulled, your job is to be the obvious next thing for a describable group of people.

The two-stage machine#

The publicly documented architecture — described by YouTube engineers in the paper Deep Neural Networks for YouTube Recommendations (opens in a new tab) and still the best available account of the shape of the system — works in two stages.

Stage one: candidate generation. From a catalogue of billions of videos, narrow it to a few hundred that might plausibly interest this viewer. This stage is coarse and fast. It leans heavily on collaborative signals: people who watched what you watched also watched these.

Stage two: ranking. Take those few hundred candidates and order them for this specific person, right now, on this specific surface. This stage is fine-grained and uses far more features — watch history, video metadata, how recently it was published, how similar viewers behaved.

Why does this matter to you? Because the two stages fail in different ways, and the failures look identical from Studio.

If you cannot get into the candidate pool, you get almost no impressions. That is a topic and audience problem — YouTube does not have a group of people it associates with this subject and with you. If you get into the pool but get ranked below everything else, you get impressions and no clicks. That is a packaging problem.

This is exactly why I always start a diagnosis with impressions, as I lay out in the five-stage system. Impressions tell you which of the two stages you are failing at.

What the ranking model is predicting#

Here is the part that most advice gets backwards.

Click-through rate and average view duration are not goals. They are inputs. What the system is trying to predict is something closer to: how much of this video will this person watch, and will they be glad they did?

Everything else follows from that. YouTube publishes its own overview of this on the how recommendations work (opens in a new tab) page, and the consistent theme is that they are optimizing for viewer satisfaction over the long run, not for any single engagement metric.

Which explains a set of otherwise confusing observations:

• A high click-through rate can hurt you. If a thumbnail wins clicks and the video then loses people at 0:20, you have given the system evidence that recommending your video makes viewers unhappy. You would have been better off with fewer, better-matched clicks.
• Watch time alone is not the target either. A ten-minute video that holds 90% is a stronger signal than a forty-minute video that holds 20%, even though the second one produced more raw minutes for some viewers.
• The same video performs differently for different people. Because ranking is per-viewer, "how is my video doing" is not a well-formed question. It is doing well for some audiences and badly for others.

The four surfaces, and why they disagree#

Home#

The viewer was not looking for anything. Your thumbnail is competing with eleven others from channels they already like. Home traffic is driven by packaging strength and by whether YouTube believes this viewer has an appetite for your subject today.

Home is where breakout videos happen, because the pool of eligible viewers is enormous. It is also the least predictable surface. You cannot aim at it precisely; you can only make things that are more interesting than the alternatives.

Suggested#

The sidebar, and whatever autoplays next. This is where most established channels get most of their views. Suggested is fundamentally about adjacency: is your video a sensible next thing for someone who just finished a related video?

Suggested traffic is the most compounding kind, because it attaches your channel to topics rather than to your subscriber list. If you want more of it, make videos that sit naturally beside videos that are already popular in your subject.

Search#

The one surface you can deliberately aim at. Someone typed a query and the ranker is matching intent. Relevance comes largely from your title and from what you actually say on camera; engagement decides the order among relevant results.

Search is capped but durable — a good search video earns views for years. I have written the practical version of this separately in the YouTube SEO guide.

Shorts#

A separate feed with separate logic, closer to a swipe-based recommendation loop than to browse. Crucially, performing well in Shorts does not transfer automatically to long-form, which is the single most common trap for channels that grow fast on vertical video. I covered the mechanics and the risks in the Shorts strategy guide.

What actually counts as a signal#

Roughly in order of how much they matter:

1. Watch behavior on your video — how long people stay, whether they finish, whether they rewatch parts. This is the core evidence. The audience retention report is your window into it.
2. What viewers do next — do they watch another of your videos, or leave YouTube entirely? Session behavior matters because YouTube cares about the whole visit, not one video.
3. Click behavior in context — did people choose your video when it was offered, and did that choice work out for them?
4. Explicit feedback — likes, "not interested", "don't recommend channel", subscriptions, survey responses.
5. Topic and metadata relevance — title, description, spoken content, captions.

Notice that the top three are all viewer behavior after the fact. You do not control them directly. You influence them by making a video that matches the promise you made.

What does not count#

I want to be blunt about this list, because it accounts for an enormous amount of wasted effort.

• Tags. YouTube has said publicly that tags play a minimal role in discovery and are mostly useful for common misspellings of your topic.
• Hashtags. Up to three appear above your title and link to a hashtag page. They are a navigation feature, not a ranking lever.
• Upload time. Slight day-one effect because your subscribers are awake. No effect on lifetime ranking. A good video posted at a bad hour still finds its audience — the recommendation system will keep testing it for weeks.
• Video length hitting a magic number. There is no threshold that unlocks anything. Make it as long as the idea deserves.
• Engagement pods, comment swaps and bought engagement. These do not merely fail; they actively hurt. They teach the candidate-generation stage that people with no interest in your topic are your audience, so your next video gets shown to the wrong pool. They also fall under YouTube's spam policies (opens in a new tab), which puts your channel at risk for a signal that was never going to help.
• Deleting old underperforming videos. Videos are evaluated individually. Your old flops are not dragging anything down.

Why videos "die", and why some come back#

A common experience: a video does well for four days then flatlines. Nothing was penalized. What happened is that YouTube exhausted the pool of viewers it was confident about, tested a wider pool, got a weaker response, and stopped widening.

The opposite also happens. A video sits at 300 views for eight months and then suddenly climbs. Usually this is because something changed on the demand side — a topic became relevant, or another video created a suggested-traffic pathway into yours. This is why older videos deserve packaging refreshes: they may now be eligible for an audience that did not exist when you published.

Reworking the title and thumbnail on an older video with proven demand is one of the cheapest wins available to any channel. It costs an hour and it changes the second multiplier in the views equation without you filming anything.

How to actually work with the system#

Five things follow from everything above.

Be describable. If you cannot say in one sentence who your videos are for, neither can the candidate generator. Vague channels get thin impressions because there is no coherent group to match them to.

Match your packaging to your content precisely. Over-promising is not a clever trick that costs you nothing; it is actively training the system to stop recommending you.

Make videos that sit next to other videos. Suggested traffic is adjacency. Ask what your target viewer just finished watching, and make the natural follow-up.

Judge at 28 days, not 28 minutes. The system tests over weeks. Reacting on day one to a slow start is how people end up changing five variables and learning nothing.

Change one thing at a time. This is the discipline that separates people who improve from people who churn. If you change the topic, the format, the title style and the thumbnail style at once, a good result teaches you nothing you can repeat.

Common questions, answered plainly#

"Does the algorithm suppress small channels?" No. Recommendation is per-video. Small channels have less data attached to them, so the system tests more cautiously and with smaller pools. That is a cold-start problem, not a penalty. It resolves with videos that hold attention.

"Did I get shadowbanned?" Almost certainly not. Check Studio for a "limited or no ads" flag or a Community Guidelines strike — those are visible and explicit. If neither is present, what you are seeing is a normal weak response to a video.

"Should I niche down forever?" Niche down until the system knows who to show you to, then expand deliberately, one adjacent topic at a time, watching whether returning viewers hold. Expanding is not forbidden; expanding randomly is.

"Do subscribers still matter?" Less than people think for reach, more than people think as a signal of quality. Subscribers who do not watch make your audience data worse. This is exactly why buying subscribers backfires, and why I would rather have 1,000 people who watch than 10,000 who do not — a point I make at more length in getting your first 1,000 subscribers.

Where this leaves you#

The recommendation system is not an obstacle and it is not a lottery. It is a matching engine with a very specific job: put the right video in front of the right person at the right moment, and be right often enough that they come back tomorrow.

Everything useful you can do falls into two categories — give it a clearer signal about who your video is for, and make the video worth the recommendation once it takes the risk.

If you want the operational version of that, the five-stage system is the process I run on every channel. If your problem is more immediate than strategic — the videos are simply not getting seen — start with why your YouTube videos get no views instead.

For primary sources, How YouTube Works (opens in a new tab) and the Creator Insider (opens in a new tab) channel are where YouTube says things officially. Everything else — including this article — is interpretation, and you should treat it that way.