Recommendation strategies provide the logic behind the item selection for your web and mobile recommendations. The strategy determines which items to display from your feed based on an algorithm that you select, and any additional filters you define to include or exclude specific items.
- For websites, access strategies at Assets › Strategies.
- For mobile sites, access strategies at App Personalization › Recommendations.
Creating a strategy
Setting | Description |
---|---|
Name, Notes, and Labels | These settings help you organize and label your strategies, but do not affect the strategy performance or behavior in any way. For more details, see Notes and Labels. |
Algorithm | Algorithms determine which items to recommend (such as the most popular items, items the user viewed in the past, and so on). For details, see Algorithms. |
Shuffle Results | For better recommendation results, you can set the widget to change items upon every pageview. Retrieve more products than requested and select a subset from these products randomly (4x the requested number of items, but a maximum of 50, are selected and shuffled). For example, if the widget has 4 slots, instead of showing the same top 4 products every pageview, show 4 popular products out of the top 16. The selection is random, with weights based on the product score. This is available for the popularity and Collaborative Filter algorithms. |
Filters | A set of exclusion/inclusion rules:
Note: Categories are selected only in the designated filter and not as part of the properties filter. For more details, see Recommendation Filters |
Activate Using API |
Use a client-side API to get the recommendations from Dynamic Yield, and render the widget with your CMS. This option also lets you use real-time filters to filter results based on data obtained within the session (for example, show products priced higher than the currently-viewed product, or present products based on the visitor's explicit selection). For more details, see Recommendations API – Client-Side. |
Custom Filter Rules |
Additional filters you can apply based on product properties, for all widgets or to the sum of them:
For more details, see Recommendation Custom Filter Rules |
Algorithms
Dynamic Yield offers a wide range of algorithms, suitable for various use cases.
Algorithm | Description |
---|---|
Empathic recommendation |
An adaptive automatic algorithm that leverages the optimal recommendation algorithm and strategy based on the user’s state of mind, journey stage, and website location. |
Automatic (deprecated) (Remains available in sections already using this algorithm) |
Let Experience OS select the best algorithm for you based on best practices for the page type you select. Algorithms per page type:
|
Keyword Similarity | Recommends items that are similar to the item currently displayed, factoring in item popularity. The algorithm uses categories and keywords from the data feed to score similarities between items, assigning higher scores to rare keywords shared between a pair of items and lower scores to more common terms. This algorithm is suitable for product/post pages. The similarity scores are recalculated whenever the product feed is updated. |
VisualML (Formerly Visual Similarity) |
A deep learning recommendation algorithm designed to identify and recommend items that are visually similar to the item currently displayed, matching things that the user can’t describe and that the marketers didn't tag into the metadata. Learn more about installing and using this extension. |
Viewed Together |
Recommends items that have been viewed in the same session as the item currently displayed. It scores items based on the number of occurrences they have been viewed together in the same session but demotes items that are typically viewed with many other items, implying a weaker connection. The scores are recalculated every 24 hours. This algorithm is suitable for product/post pages. |
Purchased Together |
Recommends products that have been purchased together with the item currently displayed. It scores items based on the number of times they have been purchased together in the same transaction while demoting products that are typically purchased with many other items. Consequently, it recommends products that are strongly linked to one another rather than products that have an arbitrary connection to a popular product. Recommendations are based on data from the past 6 months. The scores are recalculated every 24 hours. This algorithm is suitable for product and cart pages, and is not available for media or financial institution sections. |
Purchased Together Offline |
Recommends products that have been purchased offline together with the item currently displayed. It scores items based on the number of times they have been purchased together in the same transaction while demoting products that are typically purchased with many other items. Consequently, it recommends products that are strongly linked to one another rather than products that have an arbitrary connection to a popular product. The scores are recalculated every 24 hours. This algorithm is suitable for product pages and is not available for media or financial institution sections. This algorithm requires the import of offline purchase data. For more details, contact your Customer Success Manager. |
Purchased Together Offline or Online |
Recommends products that have been purchased together offline or online with the item currently displayed. It scores items based on the number of times they have been purchased together in the same transaction while demoting products that are typically purchased with many other items. Consequently, it recommends products that are strongly linked to one another rather than products that have an arbitrary connection to a popular product. The scores are recalculated every 24 hours. This algorithm is suitable for product pages and is not available for media or financial institution sections. This algorithm requires the import of offline purchase data. For more details, contact your Customer Success Manager. |
Popularity |
Scores items based on the weighted sum of all interactions – such as online and offline purchases, add to cart, and product view – favoring recent interactions over historical ones. The scores are re-calculated every time the product feed is updated. This algorithm is suitable for all page types. |
Most Popular in Category |
Returns items by popularity, but only includes items in the category specified in the page context of the category page. The scores are updated every time the product feed is updated. This algorithm is only suitable for category pages in media sites. It is not available for financial institution sections. |
Popular in location |
Assign scores to items by calculating a weighted sum of interactions that occurred within the previous 30 days in a specific location. Actions include purchases, adding to cart, and viewing products. “Location” refers to a sub-division within a country’s IP address, which can correspond to various administrative regions such as provinces or states, depending on the country’s regional structure (you can contact Support to get a list of locations for your countries). This algorithm functions effectively even in situations where user consent for cookies has not been granted. |
Recently Viewed | Recommends the last items viewed by the current user (most recent appears first). Recommendations are based on data from the last 90 days. Unlike other algorithms, this algorithm serves out-of-stock items. The algorithm is updated in real-time, whenever there is a new recommendation request. |
Viewed with Recently Viewed | Items that are usually viewed in the same session as the last items viewed by the current user. |
Recently Purchased |
The last items purchased by the current user (most recent appears first). The results are based on data from the past year. The algorithm is updated in real-time, whenever there is a new recommendation request. This algorithm is not available for media or financial institution sections. |
Recently Submitted |
The last applications submitted by the current user (most recent appears first). The results are based on data from the past year. The algorithm is updated in real time, whenever there is a new recommendation request. This algorithm is not available for media or e-commerce sites. |
Purchased with Recently Purchased |
Items that are usually purchased together with the last items purchased by the current user. This algorithm is not available for media or financial institution sections. |
Last Purchase |
Presents the cart content of the most recent purchase by the current user. The algorithm is updated in real time, whenever there is a new recommendation request. This algorithm is not available for media or financial institution sections. |
Last Submitted |
Presents the most recent application submitted by the current user. The algorithm is updated in real time, whenever there is a new recommendation request. This algorithm is not available for media or e-commerce sites. |
Purchased with Last Purchase |
Items that are usually purchased together with items in the most recent purchase by the current user. This algorithm is not available for media or financial institution sections. |
User Affinity |
Personalized for each individual user and scores items based on derived user preference and item popularity. The algorithm bases its recommendation on the visitor’s interactions with the item (views, add to cart, online and offline purchases) and their weighted scores. Then, it analyzes the item's attributes (brand, color, style, category, and so on) and calculates the user’s affinity profile. The strategy works in real time and can detect preference changes over time. This algorithm is suitable for all page types. |
Collaborative Filtering |
Items viewed, purchased, or added to the cart by users similar to the current user. This algorithm is suitable for all page types but is only supported for small feeds (fewer than 50,000 products). The scores are recalculated every 24 hours. This algorithm is not available for media sites. It is available to select customers using e-commerce sites upon request to your Customer Success Manager. |
NextML |
Deep learning algorithm that predicts the next item that a user is most likely to interact with, based on similar engagement patterns of your users in the same locale. The algorithm is recalculated in real time, whenever there's a new recommendation request. This algorithm is available to customers with e-commerce sections, on the following page types: Homepage, Category, and Any. To enable it, contact your Customer Success Manager. |
Some of the algorithms are relevant for specific page types (for example, Keyword Similarity is not relevant for homepage recommendations), so the list updates after selecting a page type.
Notes:
- Algorithms select one “leader” to display from each group-id set, which has the highest score based on the algorithm. This is not relevant for the Recently Viewed or Recently Purchased algorithms that show the SKU viewed/purchased by the user, even if they are out of stock. Nor is it relevant for pinned products.
- All algorithms serve a maximum of 50 products (except Recently Viewed (30)). If you require more than this, contact your account team.
Fallbacks
If a strategy returns fewer items than the number of widget slots, the "only include items with similar properties" and "exclude items with similar properties" are gradually ignored.
For example, if the strategy is to serve only items of the same color, and there are no items with the same color – the filter is ignored, and the widget fills all slots. If the filter includes multiple attributes (say, recommend only items of the same color and category), the attributes are gradually ignored – first ignoring the category (as it is the last attribute in the list), and only then, if there still aren't enough items to serve, the color filtering is ignored.
Also, some recommendations can result in fallback for reasons that are not related to the filter. For example, User Affinity for users with no past behavior, or Viewed Together for items that were never viewed with other items in the same session. For these cases, this is the fallback logic:
- User Affinity → Viewed with Recently Viewed → Popularity (shuffled results)
- Collaborative Filtering → User Affinity → Viewed with Recently Viewed → Popularity (shuffled results)
- Keyword Similarity → Viewed with Recently Viewed → Popularity (shuffled results)
- Purchased Together → Viewed Together (cart & product pages) → Keyword Similarity (cart & product pages)→ Popularity (shuffled results)
- Viewed Together → Keyword Similarity → Popularity (shuffled results)
- Viewed with Recently Viewed → Viewed Together (cart & product pages) → Popularity (shuffled results)
- Purchased with Last Purchase → Purchased Together → Viewed with Recently Viewed → Viewed Together (cart & product pages) → Keyword Similarity (cart & product pages) → Popularity (shuffled results)
- Purchased with Recently Purchased → Purchased Together → Keyword Similarity (cart & product pages) → Popularity (shuffled results)
- NextML → User Affinity → Viewed with Recently Viewed → Popularity (shuffled results)
- VisualML → Keyword Similarity → Viewed with Recently Viewed → Popularity (shuffled results)
Notes:
- The following strategies have no fallback: Recently Viewed, Last Purchase, and Recently Purchased.
- Fallbacks do not override a product properties filter. For example, if you exclude products with a different gender property than the currently viewed product, we'll respect the constraint even when there are not enough products to display in all available slots. However, if there are no items based on the filter, we will remove the last filter and ignore it to prevent the widget from returning empty.
Property and category dimensions
Recommendation widgets that load on Product or Post pages can have dimensions applied, which dictate that the recommendations must have or not have specific properties in common with the item currently viewed on the page. These dimensions eliminate the need to apply "include" or "exclude" rules, and are more flexible because the property values don’t need to be explicitly specified, but rather, must match or not match the current property value of the displayed item. If we don't find enough products matching this property, we remove the last dimension in the given list and ignore it to prevent empty widgets.
Dimension can be applied to all strategies as follows:
- With the same property: Recommendations must have certain properties in common with the displayed item.
- With different properties: Recommendations must not have certain specified properties in common with the displayed item.
-
With the same category: Recommended items must match the same categories or parent category of the displayed item.
- Same category: Only products with the exact same categories in the product feed are recommended. So, for example, if the viewed product categories are "men|tops|shirts", only products matching "men|top|shirts" are displayed. A product with the categories "men|tops|jackets" will not match.
- Same parent category: Recommended products have the same categories as the viewed product, except for the last category in the list. So, for example, for a viewed product with the category "men|tops|shirts", products with the category "men|tops|jackets" can be recommended.
Copying strategies between sites
You can copy a strategy to a different site in the same account. The copied strategy is not connected to the original. Any recommendation campaigns using the original strategy are not copied or affected in any way. To copy recommendation campaigns to a different site, see Copying Campaigns Between Sites.
- Go to Assets › Strategies and click the additional action icon. Select Copy to Another Site.
- Check the target site(s). Sites sharing the same vertical and platform as the current site are available for copy.
- Click Copy to Site.
FAQ
Why does strategy data in the Strategy Report look different from variation data in Experiment Reports, for variations serving the same strategy?
Strategy reports and experiment reports offer two different perspectives on recommendation data, and can sometimes look inconsistent.
This can be explained by three differences:
-
Strategies can be utilized by multiple variations
The Strategy Report tracks strategy interactions and outcomes, independently of which campaign, experience or variation might be serving them. If the same strategy is utilized by variations in different experiments, each Experiment Report will only data relevant to that variation, while the Strategy Report will consider the strategy as a whole, disregarding which variation served it. -
The two reports track different entities
The strategy reports track renders of strategies, irrespective of the variations serving them. The experiment reports track renders of variations. -
The two reports are based on different attribution logic
Strategy reports attribute revenue to a strategy only after clicking on a recommended item (see above for the definitions of Direct and Assisted Revenue). Experiment reports attribute revenue based on the attribution settings you defined.
This table summarizes how the same metrics are computed for the two reports:
Strategy Report | Experiment Results | |
---|---|---|
Impressions | Counts renders of the strategy | Counts renders of the variation |
Clicks | Counts clicks on the items recommended by the strategy | Counts clicks on any part of the injected html |
Revenue | Based on item revenue realized after a click on the recommended item. | Based on the attribution setting on the experiment. Ex: with attribution starting after "Variation is served", all purchases and their revenue will be attributed to the variation, irrespective of a click on the recommended item. |
What happens if I pin an item that is out of stock? Or, if I pin 2 slots with SKUs from the same groupID?
We display the items even though they are out of stock, or have SKUs from the same groupID.
What if I create a filter that only includes items that are out of stock?
No items will be displayed.
Is there a limit to the number of Recently Viewed items that a strategy will return?
Our Recently Viewed strategy will only return a maximum of 30 products, due to the endpoint that Recently Viewed relies on only storing the most recent 30 products that a user's ID has viewed.