Recommendation Strategies

Similarity	This strategy recommends similar products to the one the user is currently viewing, while taking into account the popularity of the product. Utilized fields: 1. Categories 2. Keywords Additional fields that can be used: 3. Price 4. Name These additional fields are used if the feed doesn’t contain keywords. Otherwise, all products in the feed will essentially be equally close within the viewed product’s category. A higher score is assigned to rare keywords that are common to products, while a lower score is assigned to the most common ones. The algorithm is suitable for product and cart pages, the ranking is recalculated once per day.
Viewed Together	This strategy recommends products that were viewed by other users in the same browsing session as the product provided in the context of the current page. The ranking of these suggested products is determined based on the frequency with which they were viewed together within a single session. However, if a product is usually viewed with many other products, its ranking is lowered, as it implies a weak connection between such products. The ranking is recalculated once per day. Data range: 30 days. The algorithm is suitable for product and cart pages.
Purchased Together	This strategy recommends products that have been purchased together with the product provided in the context. The ranking is determined based on the frequency of cases the products were purchased together within a single transaction. However, the ranking is lowered for products that are usually purchased with many other products. Thus, the algorithm recommends products that are closely related to each other rather than those that are just associated with a popular product. The ranking is recalculated once per day. Data range: 180 days, with greater weight given to purchases in the last 30 days. The algorithm is suitable for product and cart pages. To launch the Purchased Together strategy for clients, it is recommended to have a minimum of 30 days' worth of data in the platform. In other words, we advise to launch this strategy no earlier than 30 days after completing the basic integration.
Popularity	This strategy assesses products based on the weighted sum of all interactions with them, such as views, cart additions, and purchases. The assessment is calculated sequentially using two methods: a. by the type of action; b. by how recently the interaction occurred. The first method uses pre-set weights of actions: 1. View: 1 2. Add to Cart: 20 3. Purchase: 60 Coefficient weights for actions are as follows: 1. View: 1 2. Add to Cart: 2 3. Purchase: 4 For example, the final calculation of the actions of a user viewed a product, added it to the cart, and purchased it will be as follows: "1*1+20*2+60*4". The second method assesses based on recency. Coefficients are distributed as follows: 1. Up to 2 days - coefficient 8 2. Up to 30 days - coefficient 2 3. Up to 180 days - coefficient 1 If a user is identified on multiple devices, the calculation is based on cross-device behavior. The assessment is recalculated with each feed update. Data range: 180 days, with the last 30 days as recent interactions and real-time interactions within the last 2 days. This algorithm is suitable for all types of pages.
Recently Viewed	This strategy returns the latest products viewed by the current user (sorted from newest to oldest). Recommendations are based on data from the last 90 days. Unlike other algorithms, this algorithm will "collapse" empty slots. In other words, if 10 slots were requested but only 5 products were viewed, the strategy will only return 5 products. It may also show products that are not in stock. The algorithm updates in real-time when a new recommendation request is received. Data range: 90 days, with a limit of 100 products per user.
Recently Purchased	This strategy returns products the user recently purchased (the latest purchases appear first). The algorithm updates in real-time when a new recommendation request is received. Data range: no limit (within the general data retention period), with a limit of 400 products per user.
User Affinity	This algorithm works with personalization based on the assessment of products, user preferences, and product popularity. The algorithm generates recommendations based on a user's interactions with products (views, adding to cart, purchases) and their weighted ranking. Subsequently, the algorithm analyzes product attributes, such as brand, color, category, and others, to calculate the user's affinity profile. Attributes are taken from the product feed, and to determine the list of considered columns, you should contact the team working on your project. The strategy operates in real-time and can track changes in user preferences over time. Data range: 180 days. This algorithm is suitable for all types of pages. More about the affinity profile: Affinity Profile and How It Works Affinity feed attributes: • By default, the Affinity Score is calculated based on the Categories field (all category branches up to the top level). • To calculate the Affinity Score based on additional product properties, you need to assign them (it’s recommended to limit the number of affinity columns to between 3 and 5). Limitations and clarifications: 1. Affinity is based on the feed. Only attributes with discrete values can be used for affinity assessment. 2. Affinity data is collected as the user performs actions. 3. To launch affinity for clients, a minimum of 30 days of data in the platform is required. In other words, it is recommended to launch User Affinity no earlier than 30 days after completing the basic integration.
Viewed with Recently Viewed	This strategy displays products that are typically viewed together with those viewed by the user earlier. Up to 50 most recently viewed products are considered for each user.
Purchased with Recently Purchased	This strategy displays products that are usually purchased with items previously purchased by the user. Up to 50 most recently purchased products are considered for each user.

Fallbacks:

• User Affinity → Viewed with Recently Viewed → Popularity (shuffled results)

• Similarity → Viewed with Recently Viewed → Popularity (shuffled results)

• Purchased Together → Viewed Together (cart & product pages) → Similarity (cart & product pages)→ Popularity (shuffled results)

• Viewed Together → Similarity → Popularity (shuffled results)

• Viewed with Recently Viewed → Viewed Together (cart & product pages) → Popularity (shuffled results)

• Purchased Together → Viewed with Recently Viewed → Viewed Together (cart & product pages) → Similarity (cart & product pages) → Popularity (shuffled results)

• Purchased with Recently Purchased → Purchased Together → Similarity (cart & product pages) → Popularity (shuffled results)

• User Affinity → Popularity (shuffled results)

When moving to a fallback strategy, the same filters and rules that were set in the initial strategy remain active. If there is loosening in the filters, it will work the same way as in the initial strategy.

In accordance with the settings, when there is an insufficient response from the current strategy, first, the filter loosening is applied. If this is not adequate to meet the response criteria, then the fallback strategy is triggered. This process is repeated until there is a sufficient amount of products returned from the strategy.

Special Conditions:

• Recommendation algorithms return products by group_id, i.e. only one most suitable product from every group_id will be returned. This does not apply to the Recently Viewed, Recently Purchased, and Last Purchase strategies, as they show the SKUs viewed/purchased by the user.

• For RP (Recently Purchased) and RV (Recently Viewed) algorithms, products are returned regardless of their availability (in_stock). For others, the in_stock:true field is checked.

• For the page context CART products that are currently in the cart are always excluded from the results (SKU is taken from the context, group_id is obtained from the feed, and the platform filters by it). This applies to all algorithms available in cart page.

• The Affinity and Purchased Together algorithms work with the data collected from integration and require a critical volume of data for correct operation. It is recommended to launch them no earlier than 30 days after completing the basic integration.