Recommendations
NoteThis is a Standard tier feature powered by Loomi AI.
Loomi AI-powered product recommendations personalize your website experience by analyzing customer behavior and product details. Using statistics, rules, or machine learning, the system suggests relevant products to each customer as they shop. These personalized suggestions help you:
- Boost customer engagement.
- Strengthen brand loyalty.
- Encourage customers to purchase additional or related products.
Example of recommendations on a website showing a panel with related products.
Key terms
This section defines the terms used throughout the recommendations documentation.
| Term | Definition |
|---|---|
| Engine | The deployed recommender system as a service, identified by a unique ID. This ID is the last part of the URL when you view the recommender system model in Bloomreach. For more details, see Where to find the ID campaign_id. |
| Fallback | A non-personalized recommendation strategy used when the main engine fails — for example, when the engine is unavailable or a new user appears with no prior data. |
| Interaction | The connection between a user and an item, represented as {user, item, time, action}. For example: user_123 viewed item_45 at 2019-01-01 12:00:00. |
| Item | A product that users can purchase. Used interchangeably with "product." |
| Personalization | Customizing experiences to treat each user individually. This includes using Jinja for design elements like text and color, or recommending items based on user preferences. |
| Placement | The web page locations where recommendations appear, such as the homepage, product detail page, or basket page. |
| Recommender system | A method for selecting items for users, usually personalized. Also referred to as the "recommendation model" or "recommender system (RS)." |
| Similarity score | A number between 0 and 1 indicating how similar two items are to each other. |
| Top N items | A list of the top N items ordered by relevance, using the criteria or model employed for recommendations. |
| User | A visitor, potential customer, or current customer of your web page who generates events. |
How recommendation systems work
In ecommerce, the recommendation problem can be defined as: "Given all data about users and items (metadata, interactions), find the best items for a given user that will maximize a metric relevant to the business."
Most recommendation algorithms rely on either catalog metadata (such as descriptions or categories) or collected user-item interactions (events like purchase_item or view_item).
Recommendation systems fall into 4 categories based on methodology and personalization level.
Rule-based recommendations
Rule-based recommendations use simple heuristics to suggest products.
Non-personalized
- Use simple statistical methods to identify items suitable for all users.
- Examples: top-selling or brand-new items.
Personalized
- Apply statistical methods to analyze each customer individually.
- Examples: last-seen items and recent interactions.
AI and ML model recommendations
AI and ML model recommendations use trained models to suggest products.
Non-personalized
- Train AI models primarily on catalog metadata.
- Examples: attributes matching, text models, image models, and items bought together.
Personalized
- Use user-item interactions, usually in combination with other recommendation systems.
- Examples: collaborative filtering and factorization models.
NoteIn all ML and AI-based models, items are ordered based on their calculated relevancy.
Updated 14 days ago