If you’ve ever browsed an eCommerce site, chances are, you’ve been exposed to product recommendations. These machine learning-powered recommendations, designed to suggest relevant products to every site visitor, are used by marketers and merchandisers to drive discovery, increase monetization, and reduce exit rates. However, behind every powerful widget are algorithmic decisions, dictating various approaches to recommendations. Primarily referred to as recommendation strategies, these provide the logic behind which item is selected and presented in all digital recommendation widgets.
Choosing the proper strategy for the desired objective is essential to success, and the misapplication of strategies can have serious repercussions. Not only can it cause a loss of revenue, but it also may drive customers to lose faith in both the value of recommendations and your ability to deliver personal, engaging experiences altogether. And with customer experiences increasingly taking over as the key differentiator between online entities, brands must capitalize on the opportunities afforded to them through recommendations, leveraging them to make them a contributing factor to overall success.
Getting started with recommendation strategies
Using personalized recommendations is sometimes seen as a swing for the fences, but it doesn’t have to be. While contemplating the most appropriate strategies to apply, every marketer must first consider the context of each experience (i.e. which page to deploy it on) as well as the audience they want to target (i.e. specific personas or segments). Only then can a suitable strategy be selected and applied, in hopes they ultimately drive revenue. Additionally, with the ubiquity of data collection, different recommendation strategies can be delivered to different users and segments, presenting a great opportunity to connect with visitors on a personal level.
If a recommendation is properly matched with a user, it can create a pleasant user experience, increasing the chances of a connection forming between a site and the visitor. And with access to ample user and product data, marketers can do so at scale with ease. However, marketers must strike the right balance, ensuring they have the right data for the right strategies – such as purchase data for purchase-based recommendations – and don’t overstep any boundaries that may compromise their relationships with consumers.
Picking a recommendation strategy
When it comes to product recommendations, there are three major types of strategies users can look to when crafting their plans:
- Global recommendation strategies: ideal for new users during their first session who brands lack user data profiles for, serving them recommendations based on product popularity and trends
- Contextual recommendation strategies: tailor recommendations according to product attributes (i.e. color, style), traffic source, device, operating system, local weather, time, the frequency at which different products are purchased together, and more
- Personalized recommendation strategies: the most sophisticated strategy-type, leveraging both product data and user data to surface relevant recommendations for each user on an individual level
Examples of specific recommendation strategies that fall under these three tiers are listed in the table below.
Global recommendation strategies
Contextual recommendation strategies
Personalized recommendation strategies
A comprehensive collection of recommendation strategies
Let’s dive into the various recommendation strategies provided by the top personalized recommendation engine providers, breaking down how they work and when they should be used.
Top Items / Most Popular
One popular recommendation strategy is to display “Top Items,” or items ranked as “highest” or “best.” Items are scored based on the weighted sum of all interactions, such as purchases, adds-to-cart, and product views. The system favors recent interactions over historical ones and updates scores every time a data feed is synchronized.
As this strategy usually isn’t based on hyper-personalized user data, it’s especially useful when little-to-nothing is known about a user or when a user displays behavior communicating they are simply browsing around the site. It’s also great for promoting your hottest items, helping your business stand out against your competition. It will assist with and enhance the product discovery experience, helping you market your brand, as well as the popular merchandise it offers.
Additionally, using the “Popular” recommendation strategy, brands can surface popular products to site visitors. Similar to the “Most Popular” strategy, products are scored based on the sum of all interactions before being served to users. For example, once a brand identifies the top 20, 50, or 100 products for sale on the site, it can use this strategy to display any five products in the cohort in a homepage recommendation widget, rather than solely the five, most popular products.
One other similar strategy is the more-specific “Most Popular in Category” strategy, which not only surfaces the most popular items but only includes items from a specified category.
Affinity-based strategies allow marketers to make compelling product recommendations when and to whom they matter most. As users browse a site, interacting with various products, they are exposed to a number of product attributes, such as color, brand, style, and more. Recommender systems then use these interactions to identify and infer user affinities and preferences, building rich, user affinity profiles for each site visitor.
When using this strategy, recommendations are personalized to each individual user. Affinity profiles feature a weighted score based on the correlation between user interactions (number of views, add-to-carts, purchases, etc.) and the attributes of products they’ve interacted with. The system then bases its recommendations on these scores and can work in real-time, detecting any preference changes. Often taking form as a “Recommended for You” widget, this strategy is suitable for all page types.
As its name would imply, widgets powered by a “Similarity” strategy display products that resemble the item (or group of items) currently in view, factoring in the item’s popularity. Complex algorithms are designed to ascertain the metrics – using categories and keywords provided from the data feed – resulting in a similarity “score” for each item. Then, the products with the highest similarity scores are displayed to the user.
While there are many applications for this strategy, one of the most efficient is to place a recommendation widget powered by the Similarity strategy on a product detail page (PDP). Doing so will not only expose the site visitor to additional products but to products they are likely interested in. Surfacing items similar to a product in view will potentially communicate a business’ ability to understand what the user is looking for, increasing the likelihood of driving a successful sale.
Under this strategy, widgets display items that are frequently purchased together with the item currently in view. If your goal is to pad a visitor’s cart with additional products, using this strategy presents upselling and cross-selling opportunities (i.e. a car charger to go with a new smartphone, sandals for a pair of shorts, socks to go with a fresh pair of sneakers).
The system scores products based on the number of times they have been purchased together in the same transaction, demoting products that are typically purchased with many other items. Additionally, it recommends products that are strongly linked to one another rather than products that have an arbitrary connection to a popular product. Typically, these recommendations are based on data from the past six months, with scores recalculating every 12 hours.
One popular way marketers employ this strategy is for a “Complete Your Look” widget, which showcases a collection of products that complement one another. Additionally, when a visitor adds a product to their cart, using a “Bought Together” strategy to display recommendations can upsell the customer, encouraging them to make another, well-suited purchase decision.
This recommendation strategy is dependent on what product a user is currently viewing. The system scores other items based on the number of times it has been viewed with the item in view in a single session. When an item is typically viewed with many different items, the system deems the connection weak, decreasing the likelihood of it being served as a recommendation. The system recalculates these scores every 12 hours, and this strategy is suitable for any product page.
When using this strategy, the system recommends the last items viewed by the current user, with the most recently viewed items appearing first. These recommendations are typically based on data from the last 30 days.
Additionally, a similar approach is the “Viewed with Recently Viewed” strategy, which displays items that are typically viewed in the same session with the last items viewed by the current user.
One of the most popular recommendation strategies, “Collaborative Filtering,” bases recommendations on similarities between different users. The system analyzes the behavior of users similar to a current user – the items they’ve viewed, the products they’ve purchased, the items they’ve added to their carts – and recommends these products to other users displaying similar preferences and behavior, on any page type.
Purchase-based recommendation strategies
A number of unique purchase-based recommendation strategies exist, each dependent on products a user has bought in the past. The system uses transaction data to identify which products users are likely to buy based on the preferences they have displayed through their purchase behavior.
This recommendation strategy specifically looks at the most recent purchase(s) a user has conducted, typically within the last year. Ideal for encouraging repeat purchase behavior, it surfaces products a user has purchased that may need to be replenished (i.e. pet food, makeup).
Purchased Together Offline
If a retailer onboards offline transaction data, they can leverage it to power smarter recommendations. With the “Purchased Together Offline” strategy, the system recommends products that have been purchased together offline with the item currently in view on the eCommerce site. The products displayed in these widgets are scored based on the number of times they have been purchased together in the same transaction, recommending products that are strongly linked to one another and demoting products that are typically purchased with many other items.
Purchased Together Online or Offline
Similar to the offline recommendation strategy, this strategy – using offline transaction data – recommends products that have been purchased together either offline or online with the item currently being displayed. Using the same scoring mechanism as the offline strategy, scores are recalculated every 12 hours. Additionally, this strategy is best suited for product pages.
Purchased with Recently Purchased
When using this strategy, the recommender system looks at the last items purchased by the current user, suggesting items that are usually purchased together with these recent purchases. For example, when deployed on the homepage, this strategy can show all visitors who have bought an item in the past week complementary items while they are still relevant.
Purchased with Last Purchase
Akin to the “Purchased with Recently Purchased” strategy, this approach ensures all recommendations are based solely on the most recent purchase completed by the current user.
A simple and straightforward strategy, when deployed, it displays the most recently purchased product by a user in a recommendation widget.
Applying a recommendation strategy
After you’ve selected a recommendation strategy you want to use for a specific campaign, there are a few more steps to run through during the setup phase. Below is a list of the settings, rules, and steps every marketer should cover to before setting a campaign live:
- Name: Name each strategy within your testing platform to distinguish differences between campaigns
- Strategy: Select the desired strategy for each recommendation test
- Shuffle results: Option to allow items displayed in widgets to change with every page refresh
- Filters: Exclusion and inclusion rules marketers can set up for any given campaign
- Custom Rules: Custom exclusion and inclusion rules marketers can set up based on product properties. A few examples include:
- Only include: only recommend products with a specific property (e.g. items that cost more than $30)
- Exclude: never recommend products with a specific property (e.g. never recommend items that are on sale)
- Pin: always recommend a specific product first, regardless of the algorithm
When using predefined filters, the need to apply custom “include” or “exclude” targeting rules is eliminated. These dimensions are more flexible, as they do not require property values to be explicitly specified. Instead, they must match / not match the current property value of the item in view. A breakdown of how to apply these strategies is as follows:
- With the same property: ensures recommendations have certain properties in common with the item being displayed
- With different properties: ensures recommendations do not have certain, specified properties as the item being displayed
- With the same category: ensures the recommended items match the same category / parent category of the item being displayed
If these dimensions are insufficient, marketers can create custom filter rules, choosing which items to include, exclude, or pin in a recommendation widget.
Additionally, marketers are not limited to a single strategy per recommendation widget; instead, they can use multiple strategies in a single widget to improve performance. For example, on a PDP, marketers can direct half of the slots in a widget to rely on a Popularity strategy and the other half to rely on a Similarity strategy. Combining these strategies can notably increase conversion rates, identifying the highest yield specific to each product category and maximizing revenue.
The power of fallbacks
In case a specific strategy returns fewer items than the number of slots available in a given widget (i.e. only three pairs of sneakers on a site match a user’s affinity profile), the system is able to “loosen” any constraints imposed by the algorithm or filters. If this doesn’t solve the issue, the recommendation will put a respective fallback algorithm in place. For example, a “Similarity” strategy can have “Popularity” as its fallback, and a “Purchased with Last Purchase” strategy can have “Purchased Together” as its first fallback and “Similarity” as its second fallback.
Fallbacks respect filtering rules put in place. For example, if you set up a strategy and put an “exclude items under $25” rule in place, the system will respect this constraint, even if there are not enough products to display in the available slots in a widget.
Picking a strategy according to the audience
Not every strategy will work well for all site visitors. Marketers must be able to discern differences between segments of users and tailor experiences according to their behavior and expressed interests. For example, site visitors that browse 5+ products on a retail site each week should be treated differently than first-time visitors that click on an SEM ad.
When picking which strategy to use for a particular site visitor, use customer data, context, behavior, and more to identify which strategy will best suit their needs. Global strategies, like “Most Popular,” are suitable for anonymous or first-time visitors, whereas personalized strategies will help tailor recommendations for your most loyal customers, who have robust data profiles available.
This chart provides a few examples of the different recommendation strategies that suit each type of shopper and scenario.