Attentive's Product Recommendations feature is a powerful engine designed to deliver personalized product discovery to every shopper. By showing the right product to the right person at the right time, you can drive revenue and loyalty across all channels.
This guide explains how our recommendation engine works behind the scenes and how you can use it to power your campaigns.
How the Recommendation Engine Works Attentive uses a three-stage machine learning architecture to power product discovery rather than relying on a single, static algorithm. The system processes behavioral signals across your store—such as browsing, carting, and purchasing—and pairs them with a deep semantic understanding of your product catalog.
The 3-Stage Recommendation Architecture
- Stage 1: Candidate Generation The engine gathers a diverse pool of eligible products from multiple sources simultaneously:
- Personalized Candidates: Selected by a "two-tower" deep learning neural network. One tower maps shopper behavior into a mathematical embedding, while the other maps product attributes. The model pairs shoppers with items that match their predicted interest space.
- Best-Sellers: Top-performing products brand-wide.
- Trending & New Arrivals: High-momentum products and newly added catalog items.
- Stage 2: Ranking A learned ranking model evaluates the candidate pool to determine the optimal product selection and display order for each shopper. It dynamically balances personalized relevance, storewide trends, and historical popularity. For new subscribers with minimal history ("cold start"), the ranker weighs trending items and recent pre-signup activity heavily, automatically shifting toward 1:1 personalization as engagement grows.
- Stage 3: Scoring & Filtering Real-time business logic and live catalog state are applied to the final candidate lineup. The system automatically removes out-of-stock items, applies custom feed exclusions, filters out recent purchases, and enforces diversity rules before message delivery.
Under the Hood: Data Signals & Privacy Standards
- Shopper Signals: Aggregates product views, add-to-cart events, and purchase history across flexible lookback windows (ranging from 1 day to over a year), combined with location attributes and recent interaction sequences.
- Product Signals & Text Fingerprinting: Analyzes engagement trends, pricing, brand taxonomy, and product text. A language model generates a compact numerical "fingerprint" of item descriptions, allowing the engine to recognize semantic relationships—such as connecting a "merino crew-neck sweater" to a "lambswool pullover"—even for brand-new or unviewed catalog items.
- Platform Scale: Ingests billions of daily events across millions of learned signals to detect nuanced intent (such as recognizing shoppers who prefer technical outerwear) without manual rule configurations.
- Catalog Impact: Recommendation accuracy directly scales with catalog health; well-structured product names and detailed descriptions maximize predictive performance.
- Data Privacy: The model learns broad, statistical patterns of shopper behavior and does not reveal any individual brand's raw data, so your specific products, catalog, and customer information are never exposed to other brands. Any use of your data is governed by your agreement with Attentive, and the recommendations your shoppers receive are generated from your own first-party data.
Choosing a Recommendation Strategy
You can control the foundation of your recommendations by selecting a strategy that aligns with your campaign goals. If a specialized strategy returns too few products, the system automatically falls back to Best Sellers.
| Strategy | How it Works | Best For |
|---|---|---|
| Personalized Recommendations | Uses our two-tower deep learning and ranking models to select products tailored to the individual shopper's behavior and tastes. | The default for most sends; 1:1 personalization at scale. |
| Best Sellers | Highlights top-selling products (sorted by number of purchases) brand-wide. | Merchandising and catalog-driven sends; a reliable baseline. |
| New Arrivals | Shows your most recently added catalog products. | Launches, "just dropped," and freshness-driven campaigns. |
| Most Viewed / Most Added to Cart | Features products capturing the most attention or high-intent demand. | Showcasing high-interest or high-demand items. |
| Trending | Highlights products gaining momentum over a recent window. | Timely, "what's hot right now" messaging. |
Controlling What Gets Recommended
While the AI models handle the heavy lifting on relevance, you retain control over the final output using built-in filters:
- Product Feeds: Restrict recommendations to a specific subset of your catalog (e.g., a "Spring" collection or a specific price floor).
- Diversity Controls: Cap the number of recommended products that share the same category, collection, or tag, ensuring visual variety in your messages.
- Automatic Exclusions: For triggered messages, the system automatically excludes a shopper's most recent purchases and the product that triggered the message itself.
Data and Continuous Learning
To keep recommendations fresh and accurate, the core personalization models are automatically retrained multiple times per week on the latest behavioral data. Recommendations for triggered messages are computed at send time using live data, helping shoppers see up-to-date availability and options based on their very latest interactions.
To use Product Recommendations, you must have a healthy product catalog (10+ valid products), recent purchase activity, and a correctly configured Attentive tag.