Lilly Tech Systems
Data Signals for Personalization
Effective personalization requires rich, real-time data from multiple sources. Learn how to collect, unify, and activate behavioral, contextual, and preference signals to power intelligent personalization decisions.
Signal Categories
| Signal Type | Examples | Latency |
|---|---|---|
| Behavioral | Page views, clicks, searches, purchases, cart additions | Real-time |
| Contextual | Device, location, time of day, weather, referral source | Real-time |
| Transactional | Purchase history, order value, frequency, recency | Near real-time |
| Preference | Explicit ratings, saved items, communication preferences | On-demand |
| Demographic | Age, location, language, account type | Batch |
Building a Unified Customer Profile
Identity Resolution
Link anonymous sessions to known profiles using deterministic (email, login) and probabilistic (device fingerprint) matching across touchpoints.
Customer Data Platform
Centralize data from web, mobile, email, CRM, and point-of-sale into a unified profile that updates in real time as new events arrive.
Feature Store
Pre-compute and serve ML features (engagement scores, purchase propensity, content affinity) at low latency for real-time personalization decisions.
Event Streaming
Use Kafka or similar streaming platforms to process behavioral events in real time, updating user profiles and triggering personalization within milliseconds.
Data Quality for Personalization
- Freshness: Stale data leads to irrelevant personalization. Ensure behavioral data updates in real time and profile data refreshes daily
- Completeness: Handle sparse profiles gracefully with fallback strategies (popular items, segment defaults) for new or anonymous users
- Consent: Track consent status per data type and respect user preferences. Only personalize with data the user has explicitly or implicitly consented to share
- Deduplication: Merged profiles prevent showing the same recommendation or offer to a user who interacts across multiple devices