Personalization Engine Algorithms

Personalization has become the standard for providing customers an exceptional experience. We offer several algorithms to ensure you can produce the optimal content recommendations for every scenario.

While all of the algorithms can help provide a personalized experience, there are certain contexts where one algorithm may outperform others. To help you choose the best algorithm, you'll find below a description of each one, along with the most opportune scenarios for using it in order to produce the ideal user experience.

But how will you actually use these algorithms? Once your site's content metadata (URLs, titles, etc.) is stored in the system, you can call upon it within any email or onsite template using Zephyr--specifically, the personalize function.

For example, using this function, you could have your template pull the best 10 items for each user based on the user's recent browse or purchase history, or the 5 top trending items on your site.You can code your template to display these item names, links, and descriptions however you'd like. And you can freely reuse the template and offer fresh content with each campaign; your item selections tailored to individual users each time. You can even mix curation with personalization, pinning certain items to a certain place in your message. Or simply create a campaign that is the same for each recipient which you update for each send.

Notes:
* For information on using Lifecycle Optimizer's Predictive Entries, see Predictions in Lifecycle Optimizer.
* If there is not enough content to meet the personalization and zephyr settings, the system will fall back to recommending content using the popular algorithm, ignoring any filtering rules and cancel/assert statements
* The trending algorithm can't be used with merged feeds. The personalize() content parameter only applies to the interest and random algorithms.

 

Personalization Algorithm List

 All algorithms are available for use in email and SPM on your site. Your account must be enabled for some algorithms before you can use them in email. These are noted with an asterisk. With the exception of the interest algorithm, all algorithms will, by default, pull content directly from your Content Library, with optional support for a data feed as their source.

Name Description Method Type
Popular The most-viewed pages or, for e-commerce clients, the most-purchased items on your site.* Does not require a data feed. Ranking
Trending The items that have gained in popularity in the past week, based on their number of purchases or pageviews over the past 7 days.* Does not require a data feed.
Context Given a specific URL, the context algorithm returns the  items from your content library which are most often viewed or purchased by users who viewed or purchased the given item. This is based upon the history all views and purchases.* Does not require a data feed. Collaborative filtering (Wisdom of the crowd)This method requires 1000 page views over a 30 day span to work.
Purchased Taking into account the given user's recent purchase history, returns a set of recommendations based on other items that similar users also purchased.* Does not require a data feed.
Viewed Taking into account the given user's recent pageview history, returns a set of recommendations based on other items that similar users also viewed.* Does not require a data feed.
Interest This algorithm returns items in the selected feed with tags that match the user's interest profile. Requires a feed of content. Interest (Content tag-based filtering)
Random A randomized selection of content from a data feed. Random
Custom Use your own custom algorithm to serve recommendations to your users. Call in up to 100 specific Content Library items -- by URL, SKU, title, or content_id -- which your own algorithm may have preselected for the given user and added to their user profile as an array of content identifiers.* Does not require a data feed. Custom

*Please contact Support if you would like to enable this algorithm for email

 

Algorithm Details

Interest 

.The interest algorithm is the first algorithm developed to recommend personalized content to end consumers. It works by recording all of the content a customer has visited and registering all interest tags associated with that content to the user profile. As the customer continues to consume content, the system builds out an interest graph of all of the tags a customer is interested in. At the time of recommendation it then looks at the pool of content that's available and selects the best content by comparing the tags on a user's interest graph vs. the tags associated with the articles / products within the pool of available content. Unlike the others, this algorithm requires a feed of content. It uses the feed that is assigned to the template or an alternative content array, if specified in the personalize function--for example, the result of feed that you have filtered within your Zephyr code and stored in a new variable.

The popular algorithm recommends the content items with either the most purchases (for e-commerce clients with the purchase API configured) or the most pageviews (if purchases are not being passed), for all time. This algorithm is best used to make recommendations for customers with little to no browsing or purchase history. Popular products are usually popular for a reason! It also makes sense to use as your backup algorithm. That way if there is no data on the user, it will default to popular products which is always a safe bet. Email Examples: Welcome series / onboarding, daily newsletter, verticalized newsletter (filter on tags, ex. 'books') Onsite Examples: Landing or Category pages for unknown users, or as a fallback for other recommendation approaches across the site.

The trending algorithm recommends items that have surged in popularity over the past 7 days. For example, if you release a brand new product or content item that gathers a significant number of purchases or pageviews, that would out-rank items that are consistently popular. This algorithm is all about finding products or pages that have performed really well, recently. As far as use cases, this algorithm name speaks for itself! This is best used in onsite and email sections for Trending products and content items. It also works well as a fallback algorithm where you're working with limited user or contextual data, because it does not require this context, instead using overall behavioral trends to make recommendations. Email Examples: Welcome series / onboarding, daily newsletter, verticalized newsletter (filter on tags, ex. 'books') Onsite Examples: Landing or Category pages for unknown users, or as a fallback for other recommendation approaches across the site. 

Context

The Context algorithm is used to recommend items that are typically viewed together, or as replacement options that are similar to a specific product that is being viewed or has been purchased. This type of recommendation is commonly found on sites such as Amazon and Netflix. This does not leverage user-specific data for recommendation purposes, so it is appropriate for unknown users. It works great for abandoned-cart reminders to display similar items. Example use cases for e-commerce sites: Onsite, the Context algorithm is well-suited for cases where the site visitor has signaled their interest by clicking into a specific product but has not selected a specific item to purchase. During the browsing phase of the conversion process, the Context algorithm makes replacement and complementary recommendations that will more quickly and reliably lead the shopper to find an appropriate product to purchase. The product detail page is the ideal place to leverage the Context algorithm In email, this algorithm works extremely well to provide transaction-specific recommendations. Some examples are listed below:

  • Purchase confirmation: Beneath the order details, put a section that shows items that the shopper might like based upon their recent purchase
  • Post-purchase drip: "Hope you're enjoying your item! Other shoppers also liked these similar items"
  • Wish-list reminder: "You added this item to your wish list! Receive a 10% discount if you pair it with one of the following items:"
  • Abandoned cart email: "You left this item in your cart! Check out now and try pairing it with this { } for a 10% discount"
  • Browse abandon: "Here are some recommendations based on items you recently viewed:"

Purchased

Taking into account the given user's recent purchase history, the Purchased algorithm returns a set of recommendations based on other items that similar users also purchased. This recommendation is great for commerce messaging to your paying customers. By focusing solely on products that they've bought in the past, there's a stronger signal (vs. browsing) which leads to recommending products based on items that you know they're interested in.  This algorithm is best used to make recommendations for customers with established purchase history. This makes it great for cutting through the cruft of what they may have browsed (but weren't interested in) and focusing solely on what they ended up purchasing. It leads to high quality recommendations that can be used in just about any scenario. In email, it is well-suited for welcome series' or onboarding, a daily newsletter, a verticalized newsletter (filtering on tags, ex. 'books'), or triggered transactionals (post-purchase drip, winback strategies, etc.) Onsite, it can help produce a personalized shopping section, engaging the customer with the types of products they're known to love.

Viewed

The Viewed algorithm returns a set of recommendations based on an individual user's pageview history. This recommendation is great for discovery. Given that we take account the user's most recent pageviews, we can predict content that they would be interested in based on the browse behavior of other similar users. Whereas our interest algorithm is based on tags of items the user has already consumed, viewed uses other's history to find new content to recommend.  This algorithm is best used to make recommendations for customers with at least two pageviews. With the algorithm using the history of other similar users, it can be great for discovery emails. Help your customers find products that they might not have found on their own but still add in a layer of personalization. Email Examples: Welcome series / onboarding, daily newsletter, verticalized newsletter (filter on tags, ex. 'books'), triggered transactionals (post-purchase drip, winback etc.)  Onsite Examples: Pages where the site visitor has not signaled their intent in the visit, such as landing or category pages.

Custom

Use your own custom algorithm to serve recommendations to your users. Call in up to 100 specific Content Library items -- by URL, SKU, title, or content_id -- which your own algorithm may have preselected for the given user and added to their user profile as an array of content identifiers. Your personalize function will pass this information as a custom_key_type and array of custom_keys. For example, store an array of URLs or SKUs that you have recommended for each user on their profile under the custom profile field (a.k.a. var) "recommendations", then, in your personalize function, pass the custom_key value "profile.vars.recommendations". The system will look up this custom array of content identifiers and return all content metadata for display.