YouTube Algorithm: How It Hooks You (and How to Fight Back)

I noticed the pattern before I understood it: I'd open YouTube to watch one specific thing and close the tab forty minutes later having watched four things I never planned to. The algorithm had not malfunctioned. It had worked exactly as designed.

How the recommendation engine is built

YouTube narrows hundreds of millions of videos down to a handful of candidates by matching your watch history, the behaviour of viewers with similar tastes, and your subscriptions. That shortlist is then ranked by a single dominant signal: predicted watch time. Not relevance to what you were looking for. Watch time.

Supporting signals include click-through rate (does the thumbnail make you click?), engagement metrics like comments and shares, freshness, and session continuity (videos that keep you on YouTube, rather than sending you elsewhere, rank higher). The system is coherent and internally rational. Its goal and your goal are just not the same.

The feedback loop worth knowing about

Each watch signals preference. The next batch of recommendations reflects that preference, but pushed a degree further toward novelty or intensity, because pure repetition produces diminishing engagement. You click again, because the topic feels familiar but the video is new. The loop tightens.

This is not malice, and it is not quite manipulation either. It is an optimisation function doing its job. The gap is between what the algorithm measures (time on platform) and what you might actually want from a session (a specific answer, a bit of relaxation, something genuinely new). Those can overlap. Often they diverge.

A few habits help. Disabling autoplay removes the single biggest source of unintended continuation. Using the Subscriptions tab rather than the Home feed reduces algorithmic injection into your queue. Clearing watch history occasionally resets the model's picture of you, which can break stubborn loops.

Tracking whether you're drifting

Knowing the mechanism is useful. Knowing whether it is working on you, in practice, on a given week, is more useful. I built Gazenest partly because I wanted that second kind of knowing: a Self-Control score that reflects how many sessions started with a declared intention versus how many were pure drift, and a Diversity score that flags when my content range has narrowed without my noticing.

The tool does not block YouTube, and it would not solve the underlying pull even if it could. There are still evenings when I open the app with no intention at all and the score takes the hit accordingly. But seeing the number is usually enough to close the tab.