YouTube’s Algorithm Tracks 6 Signals to Pick Your Next Video

The honest answer to how recommendation algorithms work is that nobody fully knows, including the teams that build them. It is a design feature. These systems learn from behaviour at a scale no human can audit, optimising for signals (watch time, replays, shares) that correlate loosely with satisfaction and more tightly with engagement. Most coverage of this targets two audiences: creators learning to feed the machine, or critics building a case against it. The viewer in the middle, whose media diet these systems direct without explanation, rarely gets a clear account of what is happening or what to do about it.

the two-stage machine

Most people imagine the algorithm as a single thing making a single decision. Pick something for this person to watch. What runs underneath is closer to a two-round tournament.

Round one is retrieval. The system starts with a library of millions of videos and needs to reduce that to a workable shortlist, typically a few hundred candidates. It does this fast and roughly, using your watch history and the behaviour of people who watched similar things. The goal at this stage is recall, not precision. Get the plausible options into the room.

Round two is ranking. A second, more computationally expensive model scores the candidate shortlist using dozens of signals: predicted watch time, likelihood of a like, likelihood you will leave the platform, time of day, how recently the video was uploaded. The system combines these into a score and sorts the list. What you see is the top of that sorted list.

Understanding how recommendation algorithms work means recognising that these two stages have different goals. The retrieval model optimises for breadth; the ranking model optimises for predicted engagement. You are not the primary input to either model. Your watch history is. The distinction matters because it explains why the system can feel both personalised and weirdly off. It knows what you watched. It has a much weaker signal for whether you wanted to watch it.

That second gap is where the more interesting questions begin.

signals and objectives

The ranking model has a list of things it can measure and a list of things it cannot. Watch time it can measure. Completion rate it can measure. Whether you felt good about that hour of television, it cannot.

The signals available to the algorithm are behavioural proxies: you watched this for twelve minutes before switching, you rewatched that clip twice, you searched for something related at 11pm on a Wednesday. The system treats all of these as evidence of preference. Some of them are. Others reflect boredom, or spite, or the particular inertia of late-night scrolling.

Your goal is to watch something good. The system’s goal is to maximise predicted engagement in your next session. Those outcomes overlap often enough to feel like the same thing, and much of the time they do. But research into YouTube’s recommendation model found that completion rate and click-through rate are among the highest-weighted signals, which creates structural pressure toward content that holds your attention rather than content that earns it. The gap between those two is how recommendation algorithms work in practice, as distinct from how they are marketed.

That gap follows from engineering constraints. The system optimises for what it can measure. The problem is that the things it can measure are not the things most viewers would choose to have measured on their behalf.

the filter bubble question

The filter bubble theory arrived with Eli Pariser’s 2011 book and has not left the conversation since. The premise is intuitive: algorithms learn what you like and show you more of it, walling you off from perspectives that might challenge you. The version you hear most often is dramatic. Recommendation engines are sorting us into ideological silos, feeding extremism, making it impossible to agree on basic facts.

The research is more awkward. Some studies find that people on algorithmic platforms see a wider range of content than people relying on their existing social networks, because the algorithm surfaces things outside your immediate circle. Others find filter effects concentrated among people who already seek out partisan content deliberately. The evidence points to a conditional effect rather than a universal one: the algorithm amplifies what you bring to it more than it conjures the problem from scratch.

None of that makes the concern unfounded. Understanding how recommendation algorithms work in practice means accepting the awkward middle. The system nudges you. It finds patterns in your behaviour and reinforces them. Whether that tips into something harmful depends on the platform, the content category, and the person. Sweeping claims in either direction overstate what the data shows.

Algorithmic recommendation narrows your attention without fully closing your view. The filter shapes your experience in real ways. The bubble metaphor overstates how sealed that experience becomes.

Australia’s regulatory moment

While most of the world was still arguing about whether recommendation algorithms cause harm, Australia did something unusual: it regulated them.

The Online Safety Amendment (Social Media Minimum Age) Act 2024 makes Australia the first jurisdiction in the Asia-Pacific to treat algorithmic recommendation as a distinct, regulated feature rather than a byproduct of content moderation. The legislation requires platforms to disable recommender systems for users under 16 and gives the eSafety Commissioner new powers to audit how those systems operate. It is, by any measure, a significant shift in how regulators are thinking about the problem.

The framing matters. Previous approaches targeted content: remove the harmful post, label the misinformation, restrict the account. Australia’s approach targets the mechanism. It says the problem isn’t just what the platform shows you, but how it decides to show you more of it. Understanding how recommendation algorithms work is now, in Australia at least, a question with legal consequences attached.

Whether it changes much in practice depends on enforcement, and enforcement of algorithmic systems is genuinely hard. Auditing a system that runs billions of inference cycles daily, that no single engineer fully understands, that produces different outputs for different users at different times, is not a solved problem. The law exists. The tools to implement it are still being built.

taking back some control

The controls are there. Most platforms now offer some version of them: pause watch history, mark content as “not interested,” tell the algorithm you’ve seen enough of a particular topic. On YouTube, you can delete individual videos from your history and watch the feed shift within days. On TikTok, you can reset your For You page entirely, though it rebuilds from whatever you watch next.

How much that matters is a different question. Research on how recommendation algorithms weigh different signals finds that explicit negative feedback, clicking “not interested,” carries less weight than the passive signals you generate automatically: completion rates, rewatches, the half-second you spend hovering before scrolling past.

The practical upshot is less satisfying than a listicle headline. Use the controls, because they are not nothing. Be deliberate about what you finish watching, not just what you start. The system reads your behaviour more fluently than it reads your instructions. You can shift what it serves you. It does not offer an exit.

Closing / key takeaways

Understanding how recommendation algorithms work doesn’t hand you control. It hands you context. That’s worth having, even when it doesn’t produce a clean answer.

A few things worth keeping:

• Completion rate is the signal that counts most. The platform notices what you watch to the end more than what you rate or tell it to hide.
• Filter bubbles are real, but the research picture is more complicated than either camp admits. Algorithmic feeds often surface content outside your usual preferences.
• Australia now regulates recommendation as a distinct platform feature. Other governments are watching.
• The controls are real and imperfect in roughly equal measure. Using them shifts what you see.