How Algorithms Actually Decide What You See

In Part 1, I explained how platforms score and rank content based on your engagement signals — every pause, every like, every second of watch time feeding a prediction about what you’ll engage with next. But ranking existing content is only half the system. The other half is the recommendation engine: the part that decides what to show you that you never asked for.

This is where the infrastructure gets genuinely consequential.

The Difference Between Search and Recommendation

When you search for something, you’re expressing intent. You typed the words. You made the request. The platform is responding to you.

Recommendation is the opposite. The platform is making a decision on your behalf, without a request, based on a prediction about what will keep you engaged. It is proactive, not reactive. And it operates at enormous scale.

YouTube has publicly stated that over 70% of the content people watch on the platform is driven by its recommendation algorithm, not by search or direct navigation. Netflix has reported that more than 80% of viewing comes from its recommendation system. On TikTok, the For You page is the product — almost everything you see is algorithmically placed, with no search required at all.

These numbers are not incidental. They represent a fundamental shift in how people discover information, news, opinion, and entertainment. The platform is no longer a library you browse. It is a curator that has decided, based on your behavioral data, what you should see next.

How Recommendation Engines Actually Work

Recommendation engines use a technique called collaborative filtering. In simple terms: the system looks at what people like you have watched, read, or engaged with, and uses that to predict what you’ll want next. If thousands of people who watched Video A also went on to watch Video B, the system infers a relationship between those two pieces of content and will recommend B to the next person who watches A.

This sounds useful, and often it is. It’s how you end up discovering a documentary you’d never have found on your own, or a Substack you didn’t know existed. The problem is not that recommendation engines exist. The problem is what they’re optimized for.

The recommendation engine is not optimizing for your satisfaction in any deep sense. It’s optimizing for a proxy metric: continued engagement. Watch time. Click-through rate. Session length. The underlying assumption is that if you keep watching, you must be satisfied. That assumption is frequently wrong.

Anyone who has spent two hours on a platform feeling vaguely unsettled and unable to stop understands the gap between engagement and satisfaction. The algorithm cannot tell the difference between content that genuinely enriched your life and content that simply held your attention through anxiety or compulsion. Both look identical in the data.

The Rabbit Hole Question

The most discussed concern about recommendation engines is the rabbit hole effect: the idea that platforms systematically push users toward progressively more extreme or sensational content, because extreme content tends to drive higher engagement.

The research here is genuinely contested, and it’s worth being precise about what the evidence actually shows.

A 2023 study published in PNAS, which audited YouTube using 100,000 automated accounts, found that the algorithm does recommend ideologically congenial content to partisan users, and that recommendations become more congenial — and more likely to come from problematic channels — the deeper you go in a recommendation trail. The effect was most pronounced for right-leaning users. Other studies have found more limited effects, and the research overall does not support a simple story of inevitable radicalization.

What the evidence does consistently support is something narrower but still significant: recommendation engines are not neutral. They make choices. They amplify some content over other content based on engagement signals. And because emotional and sensational content tends to drive higher engagement, those choices are structurally biased toward content that activates strong responses, regardless of whether that content is accurate, healthy, or representative of the information environment as a whole.

The Personalization Trap

There is a second effect that gets less attention than rabbit holes but may be more pervasive: the gradual narrowing of what you see.

Recommendation engines, by design, show you more of what you’ve already engaged with. They are reinforcing machines. Over time, this means the range of content you encounter shrinks, not because the platform has less content, but because the algorithm has learned a profile of you and optimizes against it. Niche interests become dominant. Familiar voices crowd out unfamiliar ones. Your information diet becomes less varied without you noticing or choosing it.

This is the filter bubble in practice — not a sealed chamber where you only see one viewpoint, but a gradual drift toward content that fits the model the algorithm has built of who you are. The model is based on your past behavior. It has no way of knowing who you want to become, what you’d benefit from encountering, or what you’re missing.

Why This Matters Beyond Individual Experience

The scale of recommendation systems means that their aggregate effects on what information circulates matter at a societal level, not just an individual one. When platforms systematically surface certain kinds of content over others, they shape what seems prominent, credible, and worth discussing. They influence which voices are amplified and which are marginalized. They affect what gets treated as news.

This is not a claim that platforms are running deliberate influence operations. It is a structural observation: systems optimized for engagement, operating at the scale of billions of daily interactions, have effects on collective perception that go well beyond what any individual user experiences or chooses.

Understanding that the content you see is not neutral, not random, and not a reflection of what’s most important is the starting point for thinking clearly about the information environment you’re operating in.

If this series is useful to you, share it with someone who needs to understand how the system works. Subscribe below to follow the rest of The Blueprint.