Frequently Bought Together

This item-based widget enables you to recommend a range of products that are often purchased with a specific product.

By offering complimentary product options, you not only personalize the shopper experience but also improve upsells/cross-sell opportunities for greater revenue.

General Features

Frequently Bought Together (FBT) Algorithm powers the Frequently Bought Together widget recommendations. The following are the overall algorithm features:

  • The algorithm returns up to 250 recommendations.
  • The recommendations are refreshed every 48 hours. This ensures that the product data in recommendations accurately reflect the latest changes in the catalog.

Models Overview

Here are the models we built for Frequently Bought Together recommendations:

1Standard [default for new integrations after April 1st, 2024]Standard FBT — a vector and LLM-based co-bought model — recommends complementary products based on the basket and browsing data.
2PrecisePrecise FBT — a tree-based model — only recommends those products that were seen together in the relevant session. The model generates very precise recommendations.
3Relaxed [deprecated]

Note: There is no impact on existing customers. We will continue to provide support for customers who are already using this model. We recommend that you upgrade to the Standard model for more benefits.
Relaxed FBT — a neural network-based model — returns more recommendations than the Precise FBT model.

For example, if a<=>b was a recommendation and b<=>c was a recommendation, then it recommends a<=>c. Suppose users bought pants with a shirt and purchased shoes with pants; then, users can see shoes as recommendations for a shirt search query.

The model details are specified under their respective sections. You can select one model at a time. To fine-tune the chosen model, refer to the Available Customizations section.

Standard Frequently Bought Together Model


Default model for new integrations after April 1st, 2024

This vector and LLM-based co-bought model captures intricate product relationships and displays relevant product recommendations.

How does the model work?

The model is like a super-powered recommendation engine that uses two types of data points as input:

  • Baskets: What people actually buy together (like phones and phone cases).
  • Browsing sessions: What people view together but don't necessarily buy (like a shopper viewing a phone case browses other phone cases).

By analyzing both, the model learns hidden connections between products. It creates "profiles" for each product, maps complementary products together, and suggests items likely of interest.

With dual embeddings (codes of how products connect to other products), the model handles scenarios where product relationships aren’t the same in both directions.

For instance, a screen protector can be recommended for a phone, but a phone is not a suitable recommendation for a screen protector. So, the model ensures that a screen protector is recommended for a phone but not the other way around.

Where can this model add value?

  1. Cross-category recommendations: With richer product profiles, the model can suggest cross-category co-bought product recommendations that can lead to increased conversions. Suppose for the product “shoes,” the model could recommend socks, shoe laces, or inserts.
  2. Recommendations with sparse data: The model can predict recommendations for products with little or no purchase data by learning from existing product relationship patterns. Consider the example below:
“Red shirt” → “Denim shorts”Co-bought relation
"Red shirt" → “Red hooded shirt”Similar products based on past data. “Red hooded shirt” has no purchase history.

The model uses these product associations to predict “Denim shorts” as a complementary product for the “Red hooded shirt.”

  1. Scalability: Thanks to the intelligent use of embeddings, the Standard model can efficiently produce recommendations while maintaining accuracy even with massive amounts of browsing data.


Best Practice

Kindly ensure that the Conversion pixels are integrated correctly. This helps generate the right recommendations.

Standard FBT Model FAQs

  1. Can I modify the lookback window of data?
    Yes, you can set it to any value.

  2. When does Bloomreach recommend using this model?
    While the model is effective with ample conversion data, it is also useful when purchase data is limited for certain products. It makes use of the transitive property to recommend relevant complementary products.

  3. Does the model require a secondary fallback algorithm?
    The model provides sufficient recommendations, so there is no need for a fallback.

  4. What happens if a product does not have enough purchase data nor do the similar products identified by the algorithm have purchase data?
    It is possible that some products may not have enough data, nor do similar products have any data; in such cases, we recommend using merchandising to alter the recommendations.

Standard FBT Model Access

  • Existing Customers: Customers integrated before April 1st, 2024, will have continued support for their existing model (Relaxed or Precise). Kindly contact Bloomreach Support if you wish to upgrade to the Standard FBT model.
  • New Customers: New integrations after April 1st, 2024, will support the Standard model by default.

Precise Frequently Bought Together Model

It is a tree-based algorithm that only recommends those products that were seen together in the relevant session. The algorithm generates very precise recommendations.

Relaxed Frequently Bought Together Model


Relaxed model is now deprecated and will not be supported for new integrations after April 1st, 2024. We will continue to provide support for customers who are already using this model.

This model analyzes behavioral and product data to recommend the products purchased together frequently.

Bloomreach’s Machine learning models leverage a powerful combination of Neural Networks and Association Rule mining. These deep learning models are trained on the past 90 days of behavioral and product data to recognize purchase patterns and make product predictions


Best Practice

Kindly ensure that the ATC pixels are integrated correctly. This helps generate the right recommendations.

Available Customizations

Please reach out to Bloomreach Support to customize the below configurations:

Custom SettingStandardPreciseRelaxed

[deprecated after April 1st, 2024. Only supported for existing customers]
1Customize the maximum recall size (number of recommendations per PID) generated by the algorithm. This should be between 200-250.SupportedSupportedSupported
2Set the minimum threshold for products to be frequently bought together: This refers to the minimum likelihood that if product P1 is purchased, it will be bought together with product P2. If the number of observed instances is low, this minimum threshold can be lowered.Not SupportedSupportedNot Supported
3Filter products based on Add to Cart (ATC)/Conversion activity: This is the minimum threshold for the number of times a product needs to be added to the cart/purchased in different sessions to be considered for recommendation. If the number of ATC/Conversion instances falls below this threshold, the product will be excluded from the recommendations. By default, this threshold is set to 5, meaning a product must be added to the cart/converted at least 5 times in different sessions to be considered.SupportedSupportedSupported
4Choose to combine results from the Similar Products algorithm if the algorithm doesn’t generate enough recommendations.Not SupportedSupportedSupported
5Choose to aggregate browse session data with visit session data. By default, the algorithm uses browse session data.Not SupportedSupportedNot Supported