How does the Crobox AI work?

AI is a system that is taught a set of rules to autonomously predict patterns. Our AI works by choosing which products on a webshop to promote with a Dynamic Message as well as the best copy to feature.

In short, the Crobox AI uses contextual (in-session) data to: 

1) Select the product(s) to be highlighted with a Dynamic Message. 

2) Select the message most likely to resonate with the shopper on a specific product.

Machine Learning is applied to ensure that only the most relevant products are chosen for Dynamic Messages. By leveraging Machine Learning for this process, the shopper’s decision-making process is eased by highlighting products that are relevant to their needs without overwhelming them with messages.

It should be emphasized that the Crobox AI works fully in-session, allowing us to collect more relevant data and optimize the Dynamic Messages in real-time (whilst training the algorithm faster). This approach ensures that we stay within GDPR regulations because we do not process any data that is collected outside of the Crobox client’s webshop. 

In other words, Crobox does not process a shopper’s PII, meaning we do not use one’s browsing history, social media data, or any other data that is out of session. 

Product Ranking Algorithm 

The Product Ranking Algorithm is specifically relevant to Dynamic Badge campaigns on the Product Listing Page. Our AI is built off an implicit recommendation engine. This works much like a recommender algorithm. Except, instead of showing product recommendations on the product detail page explicitly, the AI will use the recommended products to showcase a Dynamic Message implicitly (in the form of a product badge) on the product listing page. Technically speaking, these messages are called promotions. Promotions are a piece of content - in our case a Dynamic Message - that enriches a specific product. 

In this way, shoppers still see their recommended products based on their on-site behavior, but they are seeing it in a more natural environment. This approach helps ensure that the badges they see are aligned with their shopping goals. 

If a product has been enriched by a promotion (i.e., a badge) on one landing page, we want to offer consistency. Therefore, our algorithm ensures that visitors will see the same promotion on the same products across different pages in their session.

Message Placeholders


The jacket above is shown along with three different placeholders, each capable of listing one promotion (attribute, behavioral, or custom badge). Placeholders are fueled by so-called providers, i.e., product-driven campaigns that take the form of Behavioral Nudge or Product Attribute Dynamic Badge or even a custom icon. 

Whenever the provider is asked to fill a position for this placeholder, it first evaluates all its triggers to filter ‘active’ badges. It will then select the best one matching the current visitors’ context. 

‘Bidding’ System

To select the product to show a Dynamic Badge on, the system looks first at the contextual data profile of the shopper, for example, taking into consideration:

  • Geo-location
  • Device
  • Medium
  • Products viewed
  • Sorting options
  • Time on page
  • Returning vs. new visitor

Using this information, the AI will predict the most relevant product and message for the individual. We call this a ‘bidding system’ because our AI will determine which promotions should be filled in the placeholders by process of making bids on the available slots.  

If a placeholder only has one slot available for a promotion, then it’s up to the providers to make a bid. The ‘winner’ of the bid will be the promotion with the most relevant providers. For example, imagine a product is both “Recycled Down” and part of a “Cyber Week” campaign. But there is only one slot available for one of these promotions. It’s up to the providers to make a bid on the slots, and the AI to choose the winning bid. The best bid will be the one that will most likely drive behavior based on the visitor’s contextual profile. 

Random Forest 

In order to get a deeper understanding of the algorithm we use to train our AI (which is called Random Forest) for product promotions, let me give you a scenario. 

Let’s call our AI Bob. Bob is having lunch with Monica and wants to know what to order her in advance. The menu options are chicken, fish, and vegetables: These options are our ‘promotions’.  

In order to determine what menu Monica will like best, Bob asks Monica’s ten friends what her preferred menu would be. Most friends make a bid on the vegetable course (based on Monica’s contextual information, e.g., that she’s a vegetarian, ordered it before, or even had a vegetable starter). A few other friends make a bid for the chicken course (because Monica spent more time looking at Chicken-based dishes). Since Bob can only choose one dish as a promotion, he decides to go with the vegetable course. 

This is generalized data - using information from different sources. Bob takes all the predictions of Monica’s friends and establishes which prediction, on average, Monica would most likely go for. This is how a Random Forest algorithm works, by taking into account which promotion will most likely drive click-behavior. 

In this scenario, Monica is the webshop visitor, and the menus are the different promotions that would appeal to her as she navigates the site. Her friends are the data points our AI can collect on Monica as a user, and then it’s up to the AI (Bob) to learn from the data in order to place a promotion that will most likely drive behavior.


Process of ‘promotion’ selection

Still need help? Contact Us Contact Us