When and How to Implement the Product Recommendation System

Product recommendations became increasingly popular after their introduction in retail by Amazon in 2008 and could provide a great value both for retailers and their customers. The hype led to the appearance of plenty of sometimes dangerous myths associated with product recommendations (mostly created by vendor’s sales and marketing teams). I.e. you might still see the product recommendation solution being presented as a fully autonomous magic bullet, which would increase one’s profits just after being installed on the website.

In reality, things are way more complicated. The product recommendations could do both - boosting your sales and ruining the conversion. And it is challenging to measure the added value of it because most free analytical tools may mislead you here resulting in losses increase.

In this article, I’d attempt to cover the most essential topics that need to be considered by a retailer while dealing with product recommendations and would provide some specific tips & tricks on how to get the most of your product reco.

Start with workflows and campaigns, not the tools

Product recommendations are collections of items proposed to customers for purchase. Product recommendations may differ depending on where the product suggestion is displayed, for example, a sales manager’s tablet or a product card, and the businesses’ goals, for instance, an increase in the average order value or an increase in the conversion or revenue.

Before implementing product recommendations, it’s crucial to understand in which workflows they will be useful for the customer. In classic eCommerce business, recommendations look like product cards displayed on various parts of the website. For example, choosing a T-shirt on a brand’s website will entail other similar T-shirts being recommended and shown to the customer.

Offline recommendations can be shown at the POS to a store employee, a sales manager on their tablet. Product recommendations can also be useful to a call center. When a customer calls, the operator can suggest products based on the customer’s purchase history and views.

At the same time, product suggestions can always be fine-tuned to better fit specific business goals or an audience segment needs. For instance, recommendations can display only products with a discount (in order to increase the UPT, unit-per-transaction), only goods produced by a specific brand (for example, the one that generates the most profit), or only products that you have a lot of in stock.
Here are some examples of what online and offline recommendations can look like:

Online channels

• On the homepage — popular products;

• In the catalog — popular products from the category;

• In the product card — related or similar products;

• In the cart — related products or those most frequently purchased by customers.

  
  

Offline channels

• In the call center — related or frequently purchased products;

• On the sales manager’s tablet — bestsellers and product collections;

• At the POS — related or promotional offers.

  
  

What is crucial in all these use cases - your recommendations ideally must be consistent across the different touch points and communication channels. Just imagine the experience your customers may get when receiving different (or even contradictory) sets of recomended products consequently in your promotional email, eCommerce catalog and while talking with a call center representative during the order confirmation. If it is the case you might consider a solution, enabling you to centrally orchestrate your product recommendation logic across the touchpoints - i.e. a Customer Data Platform (CDP) or similar.

The same logic applies to product recommendation algorithms variants on a different stages of the customer journey - i.e. it might not make much sense to promote cheaper alternatives in the product cart or checkout (we will cover it in more detail below).

How to Evaluate the Efficiency of Product Recommendations

Different combinations of the state and composition of product recommendations affect the metrics in various ways. This is true whether you’re looking at business metrics, such as the average order value or revenue, or proxy (or intermediary) metrics, such as page depth, click-through rate, and the number of times products were added to favorites or to the customer’s cart.

Popular products with a discount on the homepage can reduce the average order value, but increase the conversion rate instead. The recommendation of similar, more expensive products in the product card can reduce conversion rate, but increase the average order value and overall profit.

Information on proxy metrics (such as views) is collected faster, however, it doesn’t always mean the business success. For instance, one pharmacy retailer, as part of the experiment we launched on their website, added product recommendations to the cart page in order to increase the number of products in each purchase. During two days of testing, the company lost $30,000 relative to the control group. It turned out that customers started abandoning the cart in order to visit the product card pages from the recommendations and forgot to complete their orders. We saw the page depth increase, but in the end, the overall revenue decreased.

These are the metrics most affected by product recommendations:

Business metrics

• Revenue
• Profit
• Conversion to an order
• Average order value
• Number of products in the order

Proxy metrics

• Visit depth
• Session duration
• Bounce rate
• Product card views
• Item added to cart, wish list, or comparison list

To simplify the task of compiling a list of product recommendation campaigns, I would suggest cloning this Miro board and brainstorming solutions that will be useful for your business. I added a few options there as an example.

How to Personalize Product Recommendations

After determining the list of product recommendations and setting target metrics, you need to think about how to optimize the quality of your product recommendations. Product recommendations should help suggest the most useful items for customers. To do this, the product suggestions must be based on the following factors:

• Business goals — revenue growth, profitability, average order value, UPT, and sales volumes;

• Customer behavior — browsing history, items added to the cart and favorites, online and offline purchase history;

• Customer behavior of other similar customers.

  
  

This way, product suggestions will be formed taking into account the interests of the business and the customer. The more data is taken into account, the more accurate the suggestions will be. If the customer bought a shirt offline, the mobile app will recommend them matching pants, because other customers who bought the same shirt loved these pants.

As it often happens, the history of the customer’s interaction with the brand and the product range are stored in several systems: offline sales in ERP software, online sales in something like Shopify, and customer actions (e.g., adding items to wishlists) on another platform. In this case, we may not know that the customer who has just returned to the website already made an offline purchase an hour ago. Crocs Eastern Europe, for example, used to send out email blasts to their entire audience on a weekly basis. Online purchase history, offline purchases, emails, SMS, and web push notifications were all stored in different systems.

As a result, their marketers did not have access to a single source of data that would enable them to send campaigns that took purchase history into account. Customers could receive emails that, for example, recommended Crocs that they’d purchased the day before. One of the golden rule of analytics sounds like “Garbade in - Garbade out”, meaning that if you’ll feed your algorithm with inconsistent or incomplete data, you’ll most likely receive unsatisfying results as well. There is no magic here (yet).

To resolve this issue, the data should be centralized in a single system. It is possible to develop a repository that will allow the accumulated data to be used in marketing activities, but this is expensive. The most recent way to solve that bottleneck is the class of technologies named Customer Data Platforms. They provide a full scope of ready-to-use solutions for different industries.

The technology allows companies to automatically upload data on customer behavior from an unlimited number of sources, cleanse and unify data, and get a complete history of customer interactions with the brand, on the basis of which you can launch marketing campaigns, including product recommendations.

The accumulated data can also be used to train machine learning algorithms. Even businesses with tens of thousands of customers accumulate enough data in 3-4 months (with approximately a million customer action records generated over this time) to benefit from machine learning. Algorithms create a profile of the customer’s interests, find similar users and, based on what they bought, recommend the customer other products they may want to buy. This is exactly how personal recommendations work in the Tom Tailor fashion brand.

Another bonus of centralization is cohesive omnichannel marketing. This is when online channels take into account the popularity of items offline, and product recommendations on the website and in campaigns are synchronized. Thus, data centralization significantly increases the quality of product recommendations.

There are many services offering product recommendations. Their capabilities differ in the number of algorithms, whether they allow you to customize product recommendations or not, the tools they offer for measuring the efficiency of recommendations, and sources available for uploading the data in order to build a recommendation matrix. Bloomreach, Insider, and Mindbox provide companies with technologies allowing them to fine-tune product recommendations with ML solutions and take into account data from online, offline, and mobile apps.

How to Evaluate the Efficiency of Product Recommendations

The performance measurement tools differ depending on the channel in which the recommendations are used. At the same time, regardless of the channel, the principle is the same. Customers are divided into two groups. One group receives recommendations, while the second does not. If the sales are higher in the group that gets recommendations, then the customers found the recommendations useful.

Online Channels

When testing online, I recommend using Google’s Optimize tool. It’s free, it takes just a few steps to set up the experiment, and it doesn’t require programming skills to create the test.

If you already work with Google Analytics, Optimize can use e-commerce data to evaluate experiment efficiency. One variant of the website will be the original version without any changes, whereas the other variant will display product recommendations for customers. For each individual widget, you need to set up your experiment, making sure that the data doesn’t mix, otherwise you won’t know which specific recommendation widget helps and which doesn’t. Incanto, a lingerie and swimwear store, received a 5.5% increase in revenue after the implementation of product recommendations.

Experiments in email campaigns are configured in a similar way. One portion of recipients receives emails with recommendations, and the other without them. A/B tests are available in pretty much every email marketing platform.

Offline Channels

An offline A/B test with a control group works based on the same principle as an online test. This can be done when testing product recommendations within a call center. When a call is made, the operator’s software sends a request to the CDP, where the audience is already divided into two groups. In half of the cases, the platform displays product recommendations on the operator’s screen. In other cases, no recommendations are provided. Then, based on the CDP’s built-in report, the behavior of both groups is compared to determine where customers made more purchases. The same approach can be applied with POS software when the cashier sees recommendations to prompt the customer only half the time.

When Product Recommendations Are Not Useful

Product recommendations as a tool were invented by the Amazon Corporation to introduce customers to a product range. Automated product recommendations aren't useful for brands with a product range consisting of less than 100 products. In this case, algorithms will simply not have enough products to choose from. Therefore, the recommendations can be configured one time manually in the CMS. This will not only be cheaper as there is no need to buy third-party technology but also easier since once the recommendations are configured, there won’t be any need to support them.

However, the situation differs when it comes to recommendations that have not been tested on a control group. Using the example of the pharmacy mentioned above (which implemented recommendations and lost $30,000 in two days) it would be fair to say that if recommendations are not tested, it could lead to the company losing money. If we take the example of a clothing store for kids under the age of ten, it took me three iterations of the experiment in a period of two months to achieve a revenue growth of +25% relative to the control group.

How to Launch Product Recommendations

If you decide to try product recommendations in your business, I suggest you follow these steps:

Create a campaign list. Open your own website and try to understand in which situations recommendations could be useful to your customers. At this stage, I recommend making a kind of “wish list” of everything you think may work, starting from simple solutions, such as popular products on the homepage, to pop-ups with personal recommendations when a customer wants to leave the website. Arrange the resulting set of hypotheses according to the reach. The more people that see the recommendations, the faster you will get a statistically significant result in the tests. To create a list of campaigns, use the Miro Mindmap template I provided above.

Define metrics. Metrics will help you understand which products you want to recommend and determine the criteria for success. My advice is not to overcomplicate this task in the beginning. Look at the revenue and page depth (also known as the “fast proxy metric”). For offline recommendations, it could be revenue and an average order value.

Show this “wish list” to the developers or representatives of the product recommendation service. Developers will be able to tell you how long the implementation will take, and the service representatives will tell you how to quickly set up the desired campaigns. In services like Bloomreach, Klaviyo, or Mindbox, the most popular campaigns are provided out of the box. Feedback from your colleagues will also allow you to adjust the launch plan — certain stages can be implemented faster.

Ensure manageability and customizability. Check that you’ll be able to coordinate your recommendations logic across different touchpoints and customize it to better suit some specific customer’s segment - i.e. brand lovers, heavy purchases and so on.

Upload the history of the customer’s interaction with the brand and the product range to the product recommendation service. Data from online and offline channels as well as mobile apps will allow you to generate better suggestions and guarantee uniform marketing in all touchpoints. If you use a CDP, you can use the accumulated data for other marketing campaigns as well.

Set up a testing tool. For example, you can use Google Optimize for online channels, and a control group for offline channels. The distribution of the main and control groups can be 50/50, while efficiency can be assessed by revenue.

Monitor the progress of the experiment and adjust the product recommendations if they don’t work as expected. It may take up to 2-3 months before you receive the first successful results, and some widgets may cause a decrease in revenue. However, once everything’s up and running, you’ll see an increase of 5.5% in revenue, as the Incanto online store did.