Every time you open YouTube, TikTok, Netflix, or Amazon, an AI has already decided what you'll see before you've typed a single character. These recommendation algorithms now drive more than 70% of what people watch on YouTube and account for roughly 35% of purchases on Amazon, according to figures the companies themselves have published. They aren't just predicting your preferences — they're actively constructing them. Understanding how these systems actually work isn't a technical exercise. It's the difference between using the internet and being used by it.
What does a recommendation algorithm actually do?
At its simplest, a recommendation algorithm answers one question: what is this specific person most likely to engage with right now? Engage, in this context, usually means click, watch, buy, or share — not enjoy, learn from, or benefit from. That distinction matters enormously, and we'll come back to it.
The systems powering these decisions are called collaborative filtering models at their core. The basic idea, developed in research labs at companies like Xerox PARC in the early 1990s, is straightforward: find users whose behaviour resembles yours, then recommend what they engaged with that you haven't seen yet. If thousands of people who watched the same three documentaries you watched also binged a fourth one, the algorithm surfaces that fourth one to you. You never asked. It just appeared.
Modern systems layer multiple approaches on top of this foundation. Content-based filtering analyses the actual attributes of items — the genre tags on a film, the keywords in an article, the tempo and key signature of a song on Spotify — and finds items with similar attributes to ones you've already shown interest in. Hybrid models combine both approaches, weighing them differently depending on how much data they have about you.
Deep learning changed the game significantly in the 2010s. Neural networks can now extract abstract patterns from raw behaviour data — how long you hovered over a thumbnail before clicking, at what second you paused a video, whether you rewatched a clip — that no human engineer explicitly programmed. The system learns these signals matter without anyone telling it they do.
What signals is the AI actually watching?
The naive assumption is that the algorithm tracks what you like. In reality, it tracks what you do — and those two things diverge in revealing ways.
The signals feeding most recommendation engines fall into a few categories:
- Explicit signals: ratings, likes, thumbs up/down, saves, follows, purchases
- Implicit signals: watch time, scroll speed, pause behaviour, cursor movement, replay frequency
- Session context: time of day, device type, whether you arrived via a search or a notification
- Social graph data: who you follow, who follows you, mutual connections
- Cross-platform data: on some systems, behaviour from other apps or websites feeds in via advertising identifiers
Research published by engineers at Netflix found that implicit signals — particularly watch time and completion rate — predict future engagement far better than explicit ratings. People click five stars on films they feel they should like, but they finish films they actually like. The algorithm learns to trust behaviour over stated preference.
This creates a quiet but significant problem. Emotional content — outrage, anxiety, awe, desire — reliably drives more engagement than calm, informative content. Not because people want to feel bad, but because high-arousal emotion extends watch time and increases sharing. The algorithm has no concept of 'this made the user feel worse.' It only sees: they kept watching.
Why your feed stops surprising you over time
There's a well-documented phenomenon in algorithmic systems called the filter bubble — a term popularised by Eli Pariser in his 2011 book of the same name. The idea is that personalisation progressively narrows your information environment. The algorithm learns what you engage with, serves you more of it, and in doing so reinforces the very preferences it's measuring. Your feed becomes a mirror rather than a window.
The mechanism is a classic feedback loop. Early engagement data shapes recommendations. Recommendations shape future engagement. Future engagement produces more data confirming the original pattern. Over weeks and months, the system becomes increasingly confident about a narrow version of you — the you who watched three videos about home renovation at 11pm on a Tuesday — and optimises heavily for that version.
Exploration versus exploitation is the technical name for this tradeoff. A system that only exploits what it already knows about you will never discover what else you might love. Good recommendation engineers try to inject deliberate randomness — serving unfamiliar content occasionally to test whether your interests have expanded. But commercial pressures push against this. Exploitation drives reliable short-term engagement. Exploration risks showing you something you don't like, which tanks session metrics.
A large-scale study of YouTube viewing behaviour, conducted by researchers at MIT and published in the early 2020s, found evidence that recommendation systems do push users toward progressively more extreme or niche content over time — not through any conspiratorial intent, but simply because specialised content has fewer competitors and the algorithm finds it easier to predict engagement within a narrow lane.
How TikTok's system broke every rule
Most recommendation algorithms require history. They need months of behaviour before they can reliably personalise your experience. TikTok's system, built around a model called interest graph ranking, changed the assumption entirely.
When you open TikTok for the first time, it knows almost nothing about you. But it still builds a uniquely compelling feed within minutes. How? It treats every new video as an experiment. Each piece of content you see is a probe: do you watch it fully, replay it, share it, or scroll past in two seconds? That micro-behaviour generates a rich signal immediately, without any prior history.
The other innovation is that TikTok's algorithm weights content features heavily, not just user features. Traditional collaborative filtering says: find users like you, recommend what they liked. TikTok's system also asks: find content with these engagement patterns, surface it to users who might respond to those patterns. The content leads the user discovery, not the other way around.
This is why complete strangers end up in the same algorithmic pocket. You might have nothing in common demographically with another TikTok user, but if you both rewatch videos at a specific moment of emotional payoff, the system routes you toward the same content cluster. The algorithm found a pattern you share that you didn't know existed.
The commercial result has been extraordinary — and the implications for attention are significant. TikTok users in the United States spent an average of 95 minutes per day on the app in 2023, according to data analytics firm eMarketer. That figure outpaces every other social platform.
Can you actually influence what the algorithm shows you?
Yes — but it requires understanding what the system actually responds to, not what it claims to respond to.
Most platforms give you explicit controls: 'Not interested', 'See less of this', 'Manage your history.' Research and engineering blog posts from teams at Spotify, Netflix, and YouTube consistently confirm that these signals do update your model, though the effect varies. Repeatedly ignoring an entire content category has more weight than a single 'not interested' click.
More powerful than explicit controls is deliberate engagement behaviour. The algorithm is measuring everything, so you can use that against it:
- Watch to completion only the content you actually want more of
- Scroll past quickly when something doesn't serve you — don't let autoplay run
- Use search intentionally to introduce new topics; this seeds the interest graph with your terms rather than inferred ones
- Follow accounts in categories you want explored; the social graph signal is strong in most systems
The harder intervention is structural. Using incognito mode or signed-out browsing breaks the personalisation loop entirely — you get a generic feed based on population-level trends rather than your individual model. Counterintuitively, this is sometimes more intellectually useful. You see what's broadly popular rather than what an increasingly narrow version of you supposedly wants.
The real insight is that you are always training the algorithm, whether you're aware of it or not. Passive consumption is not neutral. Every second of watch time is a vote for more of the same.
Recommendation algorithms are not neutral pipes delivering content you'd have found anyway. They are active systems with commercial incentives that don't fully align with your wellbeing, built on the insight that engagement and satisfaction are different things. Understanding this doesn't require you to delete your apps. It requires you to remember one thing: your behaviour is always data. The algorithm is always listening. The only question is whether you're feeding it deliberately or by default.
Originally published on SnackIQ
Top comments (0)