How Canva Personalized for 95M Users Without Private Data
TRIGGER
Teams need to personalize AI-generated content at scale but cannot access users' private data due to privacy policies—yet still need to understand user preferences and behaviors to create relevant experiences.
APPROACH
Canva's team couldn't access users' private designs to understand their style preferences, so they used public template metadata as a proxy. When users created designs from templates, they logged which templates were used. Since templates have public style/theme tags, they could infer user preferences from aggregated template usage. Input: user's template usage history + public template metadata (style/theme tags). Output: inferred user preferences (top styles, themes, design trends). They matched 95 million users to design trends and personalities using this approach.
PATTERN
“Users interacting with public templates and libraries leave preference signals in metadata without exposing private content. Infer from what they choose, not what they create.”
✓ WORKS WHEN
- Users frequently interact with a library of public, tagged resources (templates, presets, components)
- Public resources have rich metadata that correlates with user intent or style
- Aggregate behavior patterns are sufficient—you don't need to analyze individual creations
- Privacy policies restrict access to user-generated content but allow behavioral analytics
✗ FAILS WHEN
- Users primarily create from scratch without using public templates or resources
- Public resources lack meaningful metadata or tags
- Personalization requires understanding the actual content of user creations (not just which templates they chose)
- User base has highly unique needs that aren't represented in public resource taxonomy