When you are shopping online, how do you feel about the suggestions the website or app makes? Those personalized suggestions are part of what has driven the demand for Big Data in the last decade, and are arguably at least partly responsible for Amazon’s incredible success. Perhaps then, you, as a retailer, also need the magic of personalized product recommendations. But how?
What isn’t a personalized product recommendation?
Soon, we will focus in on what makes a recommendation personalized, but let’s first consider what is not personalized:
“Social proof” refers to items that are generally popular items—the best-selling products—on the online store. No data unique to the user is necessary here. Still, the social aspect is useful because there are those who are looking for items that their friends or their in-group already have or are interested in. Plus, retailers who follow the Pareto principle “know” that 80% of their revenue comes from the 20% of their product range that sells best.
Showing recommendations based on “business rules”
These are items that you need to get rid of, because you have too many in stock; or those that are low in stock, so you add a little note saying “Hurry! Only 4 items left!” (Customers are savvy to such a note being potentially untrue, however.) Another rule regarding related items: Amazon, of course, offers books by the same author, in the same genre, and on a similar topic to the one the user searches for. These related items are not personalized, but are seen as helpful.
You’ll also want to consider that if your business is new, you don’t have the data an algorithm needs to make relevant recommendations. You will, however, want to have the algorithm in place, collecting historical data and learning customer behavior as your business grows, so that it’s ready to recommend relevant products when you’ve made enough sales.
During this initial period, your recommendations will be based on the non-personalized data mentioned above, plus perhaps location. (Just don’t force customers to reveal their location, or take it without asking. Let customers browse, in the same way a shop assistant leaves shoppers alone.)
You may also use traffic sources—that is, if your business is selling water sports equipment, and customers come to your site via an ad or affiliate link from a surfing website, then offering surfboards and rashies would be expected, whereas drysuits and masks would not.
From the very first search, of course, your algorithm can start to make some useful recommendations. If the user searches for a blue top, perhaps shorts or skirts in black, white, and a different (curated!) shade of blue would be a welcome suggestion; orange or olive green might not leave your customers with such a good impression. Even if they don’t leave, your average order values may suffer.
What a personalized product recommendation is
You’ve converted a customer to a client. Now what? Personalized product recommendations are the suggestions of an algorithm that uses all sorts of shoppers’ data in determining what an active user might be interested in. If your site and overall shopping experience (plus some good deals?) brings your new customers back often enough, your algorithm can use the following data to refine its recommendations and bring them back for more:
- Their purchase history
- What they’ve put into their shopping cart—especially if they’ve abandoned it and didn’t finish the sale
- Their social behavior: did they like or share a product they found on your site?
- The customer’s demographic segment
The first two items are self-explanatory. The second two, however, need a bit of unpacking.
While Amazon enjoys the highest trust rating among customers when it comes to their data, and some other shopping portals also score well, social media comes dead last. The upshot for retailers is that you should refrain from using data sold by Meta or Twitter to make product recommendations.
That is, don’t offer Chicago Cubs merchandise just because a customer follows the team or writes posts about them. You can, however, offer the merchandise if the customer clicks the Like button that you put next to a Chicago Cubs coffee mug product detail page.
It should come as no surprise that Gen Z is much more comfortable sharing data with an online store than Gen X and Baby Boomers are—it’s just that they want to be the ones deciding what to share.
So, in order to get more personalized product recommendations, 55% of those polled by Cheetah Digital will give you their family structure, age, and even body measurements, if you’ve shown you can be trusted with that information.
You may send emails with your recommendations: Smarter HQ found that the top 3 most desired marketing tactics involved email. Skip the text messages and push notifications, though: those figure in three of the top 5 creepiest tactics!
How to build a personalized product recommendation engine
Recommendation engines have six tasks:
- Gather data from user behavior, both browsing history and purchase history
- Look for patterns in what customers bought
- Formulate the patterns found as predictions
- Calculate probabilities for various behaviors
- Compare the probabilities with what’s available in the inventory
- Present the most plausible matches
A recommendation engine will use data mining algorithms to handle the first three of these tasks. The other half of the engine is the machine learning element. Its job is to upgrade the predictions made and the probabilities found, so that the results of the mining and the learning are incorporated into the online shopper’s experience.
To return to the first step, we need to look at how customer data can be filtered. There are two types of systems commonly used to create recommendations:
- Collaborative filtering systems
- Content-based filtering systems
Collaborative systems analyze data about what other customers bought in order to figure out what items are most often purchased, whether in the same order or in various orders over time. For example, if a user viewed a certain set of bike pedals with clips, the recommendation engine might show them complementary products such as shoes that other users bought along with those pedals.
Other useful data in this case can be geographic location, because a customer in Canada would be more interested in long, warm leggings for biking in, while an otherwise similar customer in Florida would probably always wear bike shorts. The complementary products will be different.
Content-based filtering systems are ones that consider a customer’s past choices to say, “If you liked this, you might also like…” They are hyperfocused on the individual, meaning, of course, that this will be the product personalization filter that takes the most time to perfect.
Once you have enough data, though, it can be powerful in terms of keeping undesired products out of the customer’s way. If the user has always searched for a particular brand of bike gear and never clicked on a competing brand, then perhaps shoes from that other brand would not be a good recommendation.
Luckily, you can also try to collect more data, such as body measurements and family structure. If you only have size S of a particular product in stock, it’s no use including it in recommendations for a person who always buys size L; if the customer searched for that item specifically, you’ll want to offer items that are as similar as possible and that are available in their size.
Also, if the shopper who normally buys size L buys a child’s size or a pink unicorn bike bell, it would be useful for your algorithm to understand that they were buying for a family member, and to not make suggestions on the basis of those products with every visit.
For the best results, of course, personalized product recommendation engines merge these two filtering systems.
Personalized product recommendations: the data
Yes, if you have enough products and enough customers, you need a personalized product recommendation engine.
Monetate found in their study that 86% of e-tailers who see a high ROI reported that 21% or more of their marketing budget went on personalization. Significantly, those with the highest ROI were the ones focusing on customer loyalty and customer lifetime value. You will probably also find that good personalization has a positive effect on average order value… but beware: Monetate also found that it’s the companies focusing on average order value who have the lowest ROI from personalized recommendations.
If you have earned your customers’ trust, you’ve opened doors for yourself. Smarter HQ reports that 90% of those surveyed will actually share their behavioral data in return for discounts and easier shopping. Cheetah Digital’’s research, meanwhile, shows that half of all respondents had clicked through an email and gone on to buy something from the sending retailer. In other words, zero-party data is golden.
Marketing based on cookie (third-party) data, meanwhile, is considered creepy: 35% of respondents reported frustration with such marketing. As a result, we really don’t need to fear the advancing restrictions on cookies, nor the effects of Europe’s GDPR and the California Consumer Privacy Act. Earn the customers’ trust, invite/entice them to share data with you, and only you, and give them better recommendations.
Clearly, knowing your customers as much as they expect is essential. Between 30% and 50% of Cheetah Digital’s respondents reported frustration when presented with poor recommendations and mistargeted marketing messages, whether by email or on the retailer’s website.
The research seems clear: Millennials and Gen Z are willing to trade their data directly with retailers in exchange for relevant recommendations, discounts, and a smoother online shopping experience. You can build your product recommendations into not just the home page and the product pages, but also the shopping cart page and even the 404 page.
The benefits to retailers go beyond the short term, beyond higher average order values—personalized recommendations bring you stronger loyalty.