Machine learning at scale 🤖
← Back to book index

Frequency Capping

Limiting how often users see the same ads to improve user experience and campaign efficiency.

Why Repeated Exposure Hurts Everyone

User Experience

  • Ad fatigue: Users become annoyed by repeated ads
  • Decreasing engagement: CTR drops with repeated exposure
  • Negative perception: Hurts brand and platform reputation

Advertiser Efficiency

  • Diminishing returns: Additional impressions provide less value
  • Wasted budget: Better to reach new users
  • Lower ROI: Frequency capping improves campaign efficiency

Platform Health

  • User retention: Poor experience drives users away
  • Long-term revenue: Short-term revenue gains hurt long-term value
  • Market efficiency: Better allocation of impressions across users

User-Level vs. Household-Level vs. Device-Level Caps

User-Level Caps

Limit impressions per user ID:

  • Most precise: Tracks individual users
  • Requires login: Only works for logged-in users
  • Privacy considerations: Requires user identification

Household-Level Caps

Limit impressions per household:

  • Cross-device: Works across user's devices
  • More complex: Requires household identification
  • Better for brand campaigns: Prevents over-exposure within household

Device-Level Caps

Limit impressions per device:

  • Simplest: Works without user identification
  • Less precise: Multiple users may share device
  • Cookie-based: Relies on browser cookies (privacy concerns)

Choosing the Right Level

  • Campaign type: Brand awareness vs. direct response
  • User identification: What data is available?
  • Privacy regulations: GDPR, CCPA compliance
  • Technical constraints: What can be implemented at scale?

Implementing Caps at Scale: Approximate Counting

The Challenge

Tracking frequency for billions of users and millions of ads:

  • Memory: Exact counting requires too much storage
  • Latency: Must check caps in <10ms
  • Accuracy: Tradeoff between precision and efficiency

Approximate Counting Techniques

Bloom Filters

  • Space-efficient: Constant memory per ad
  • False positives: May incorrectly say cap reached (safe error)
  • No false negatives: Never incorrectly says cap not reached
  • Use case: First-stage filtering, can have false positives

Count-Min Sketch

  • Approximate counts: Estimates frequency with bounded error
  • Space-efficient: Much less memory than exact counting
  • Probabilistic: Small chance of overcounting
  • Use case: When approximate counts are acceptable

HyperLogLog

  • Cardinality estimation: Counts unique users, not frequency
  • Very space-efficient: Constant memory regardless of users
  • Use case: Counting unique users who saw ad (not frequency)

Exact Counting (When Needed)

  • In-memory cache: Hot data in fast storage
  • Distributed systems: Shard by user ID
  • Use case: When accuracy is critical and scale allows

Cross-Channel Frequency Management

The Challenge

Users see ads across multiple channels:

  • Web: Desktop and mobile browsers
  • Mobile apps: Native apps
  • Video: YouTube, streaming platforms
  • Social: Facebook, Instagram, etc.

Coordination

  • Unified tracking: Identify same user across channels
  • Shared caps: Apply caps across all channels
  • Channel-specific: Some caps may be channel-specific

Technical Challenges

  • Identity resolution: Matching users across channels
  • Real-time coordination: Update caps across systems
  • Privacy: Respecting user privacy while tracking

The Tension Between Frequency Caps and Delivery Goals

The Conflict

  • Frequency caps: Limit impressions per user
  • Delivery goals: Advertisers want to reach target audience
  • Budget constraints: Advertisers have limited budgets

Scenarios

High Frequency Cap

  • Easy delivery: Can show ads to same users repeatedly
  • Poor efficiency: Diminishing returns, wasted budget
  • User experience: May annoy users

Low Frequency Cap

  • Better efficiency: Reach more unique users
  • Harder delivery: May not reach all target users
  • User experience: Less ad fatigue

Optimization

  • Campaign objectives: Brand awareness vs. direct response
  • Audience size: Large audiences can have lower caps
  • Performance data: Use historical data to optimize caps
  • A/B testing: Experiment with different cap levels

Advanced Strategies

  • Dynamic caps: Adjust based on user engagement
  • Creative rotation: Show different creatives to same user
  • Sequential messaging: Show related ads in sequence

Content to be expanded...