AI summary by TwinMind
The Algorithm... Explained by YouTube Employee! (SECRETS REVEALED)
Think Media
12m 12s
Todd's Background and Algorithm Purpose
- Todd = Head of Discovery at YouTube, creator of YouTube algorithm
- Algorithm has two main goals:
- Show each viewer videos they'll love watching
- Make viewers return to YouTube
- Algorithm complexity broken down into two key parts: personalization and performance
Personalization Component
- Algorithm creates digital avatar/persona for each viewer based on watch history
- First priority: lean into personalization by showing new videos from previously watched channels
- Also analyzes topics/video types from viewing history
- Homepage shows many videos from unsubscribed channels based on topic preferences
- Example: watching basketball videos yesterday → more basketball recommendations next day
- Example: podcast clips viewing → continued podcast recommendations despite no subscriptions
Video-Based Recommendations
- Algorithm analyzes individual videos beyond just viewer history
- For each video, creates ranked list of best other videos audiences tend to watch
- Process: looks at all other videos watched by people who viewed specific video
- Generates hundreds/thousands of video suggestions per user
- Suggested videos under current video calculated by:
- Analyzing current video viewers
- Finding common videos those viewers also enjoyed
- Recommending based on similar viewer avatars/interests
Deep Learning Brain
- Uses deep learning modeled after human brain function
- Creates neurons and connections between videos
- Fed data about viewer watch patterns to find video connections
- Maps user data through "brain" to determine closest video matches
- Algorithm trained to think like human brain, not just focus on video length/tags/titles
- Goal: replicate human decision-making for video enjoyment
Performance Component
- Real-time data collection as videos are shown to viewers
- Testing process:
- Start with regular channel subscribers
- Analyze viewer history and similar user reactions
- Distribute to people with similar viewing patterns
- Performance metrics tracked:
- Click-through rates
- Watch duration
- Likes/dislikes
- "Not interested" clicks
- Scroll-past behavior
Satisfaction Priority
- YouTube prioritizes "satisfaction" over raw watch time
- Not all watch time equally valuable - quality matters more than quantity
- Examples of satisfaction:
- Life-changing content leading to subscriptions
- Inspiring content making viewers return daily
- Content providing exact answers without wasting time
- Measurement through millions of daily survey responses
- Surveys ask viewers about yesterday's watched videos
- Responses fed into deep learning brain
- Helps differentiate satisfying vs. unsatisfying content
Video Testing Stages
- Stage 1: Show to regular subscribers, measure enjoyment
- Stage 2: Expand to new audience if subscribers enjoyed
- Stage 3: Progressively larger audiences if performance maintains
- Additional diversification algorithms prevent single channel domination
- Viewers want variety on homepage, not just one channel (exception: Mr. Beast)
Algorithm Benefits for New Creators
- Algorithm designed to show new channels and videos to interested viewers
- Not subscription-dependent - based on video quality and viewer enjoyment
- Small/new channels can succeed if they create engaging content
- YouTube finds right audience for good videos regardless of subscriber count
- Focus on making videos people love watching vs. gaming technical aspects
