How Do Recommendation Algorithms Work?

The hidden engine behind TikTok, YouTube, and Instagram — how platforms decide what you see next.

↓

Interactive

Be the Algorithm: What Do You Recommend?

You are the recommender system. Given what the user engaged with, what video do you suggest next? 4 rounds.

How It Works

6 Key Concepts Behind the Feed

How platforms decide what to show you.

📋

1. Collaborative Filtering

Tap to flip

"Users who liked X also liked Y." The algorithm finds patterns across millions of users. If User A and User B both liked videos 1, 2, and 3, but User A also liked video 4, the system recommends video 4 to User B. This is how Amazon started in the early 2000s. ^[1]

🎯

2. Content-Based Filtering

Tap to flip

The algorithm analyses the content itself — tags, captions, audio, visual features. If you watched a video tagged “vegan recipes,” the system looks for other videos with similar tags. TikTok uses this heavily alongside collaborative filtering. ^[2]

⏰

3. Watch Time & Engagement

Tap to flip

The strongest signal for video platforms is how long you watched. A 30-second watch on a 10-minute video says “somewhat interested.” Watching 90% says “give me more.” Likes, shares, comments, and replays are also weighted. TikTok’s algorithm mainly optimises for watch time.

🔍

4. Candidate Generation + Ranking

Tap to flip

YouTube uses two neural networks in sequence. The first (candidate generation) picks hundreds of videos from millions using collaborative filtering. The second (ranking) scores and reorders them by predicted watch time and relevance. This two-stage approach balances speed and accuracy. ^[3]

🔴

5. Exploration vs Exploitation

Tap to flip

The algorithm faces a trade-off: show what you already like (exploit) or try something new (explore). Too much exploitation creates filter bubbles. Too much exploration frustrates users. Platforms reserve about 10–20% of the feed for exploration.

🛡️

6. Filter Bubbles & Echo Chambers

Tap to flip

Algorithms that only show you what you agree with create filter bubbles — personalised information ecosystems where opposing views are invisible. Coined by Eli Pariser in 2011. This can polarise opinions and spread misinformation. ^[4]

By Platform

TikTok vs YouTube vs Instagram

Each platform optimises for different outcomes.

🎵

TikTok: For You Page

Tap to flip

Optimised for watch time and completion rate. Shows 30 diverse videos to new users, then learns from every action: full watches, replays, skips, shares. Uses a blend of collaborative and content-based filtering. Engagement overrides follower connections — you see what the algorithm picks, not who you follow.

▶️

YouTube: Watch Next

Tap to flip

Optimised for session watch time — keeping you on YouTube as long as possible. Uses a deep neural network for candidate generation (broad screening) and another for ranking (precision ordering). Considers subscription history, search, demographics, and video age. ^[1]

📸

Instagram: Explore & Feed

Tap to flip

Optimised for engagement and connection. The main feed shows posts from followed accounts + algorithmically inserted recommended posts. Explore tab is entirely algorithm-driven based on what you liked, saved, and shared. Stronger emphasis on social signals (friends’ likes, saves).

⚡

What They All Share

Tap to flip

All three platforms use: user embeddings (vector profiles of your taste), item embeddings (vector profiles of content), real-time feedback loops, and A/B testing to tune the algorithm. They all optimise for engagement metrics, not content quality or user well-being.

💰

Who Pays the Bills

Tap to flip

Recommendation algorithms are optimised for ad revenue, not your happiness. More time on platform = more ads shown. This creates a fundamental misalignment: the algorithm wants you addicted; you want to be informed or entertained without losing control.

🔌

You Can Influence It

Tap to flip

Every action trains the algorithm: like, share, save, follow, comment, skip, report, "not interested." Using “Not Interested" and muting keywords helps retrain the algorithm away from content you do not want. The algorithm learns from all signals, not just likes.

Quiz

Do You Understand the Feed?

6 scenarios. How does the algorithm interpret this?

The algorithm sees you scroll past a video in 1 second:

This is a negative signal — the algorithm should show less of this type of content.

A user watches a full 45-minute documentary on YouTube:

This is a strong positive signal regardless of whether they liked or commented.

Recommendation algorithms are designed to show you accurate and balanced information:

The algorithm prioritises content quality and factual accuracy over engagement.

A new user with no history opens TikTok:

TikTok shows them popular viral videos from diverse categories to start learning their preferences.

If you only watch content from one political viewpoint:

The algorithm will ensure you see opposing viewpoints for balance.

The "Not Interested" button or hiding posts:

Using "Not Interested" helps retrain the algorithm away from content you do not want to see.

Feed Depth

How Deep Is Your Feed?

From casual browsing to algorithmic rabbit hole — slide to see how your feed evolves.

050100

Slide the dial

Take Control

How to Take Control of Your Feed

The algorithm follows your signals. Give it better ones.

Use "Not Interested" and Hide

Every platform has this option. Use it aggressively on content you do not want. It is the single most effective way to retrain the algorithm.

Skip fast, do not linger

If you do not like a video, scroll past quickly. Lingering even to read the caption signals interest to the algorithm. A 1-second skip is a clear negative signal.

Actively like and save good content

Liking, saving, and sharing are strong positive signals. Use them deliberately to tell the algorithm what you want more of. Quality in = quality out.

Clear watch history occasionally

Most platforms let you clear watch/search history. Doing this resets the algorithm's learning and can break you out of a filter bubble. Do this every few months.

Follow diverse creators

Diversity in your follows = diversity in your feed. The algorithm balances what you follow against what you watch. A varied follow list gives it more room to explore.

Turn off autoplay

Autoplay is designed to keep you watching without conscious choice. Turning it off creates a friction point — you decide before each video starts.

Remember: you are the product

Your attention is what platforms sell to advertisers. The algorithm is not designed for your well-being. Take control deliberately. Every minute you scroll is revenue for them.