WhisperProduct recommendations are a common feature of many commerce websites and commerce systems. While simple product recommenders are an important technology, the recommendations themselves are not related to the individual customer. People often think product recommenders and personalization engines are one and the same. There are very distinct differences between the two and how they work. Let’s take a look at few of the ways that simple product recommenders masquerade as personalization systems.

First, some background on the topic. The basic technology behind many product-to-product recommendations is collaborative filtering. This technique, made famous by Amazon, uses an aggregate of browsing behaviors of shoppers visiting the site to figure out things like “people who viewed this also viewed these other items” or “people that bought this item also bought these other items”. The result? The website can showcase a limited set of items that are likely to be interesting to a customer who is looking at a particular item or has placed items in their cart.

It may seem obvious, but simple product-to-product recommendations are not personalized to the individual customer. Every customer that comes to a particular product page will see the same items on that page. The product recommendations are contextual to the product on the page or in the cart. This is not a bad thing. The customer’s intent is usually aligned with the products they are looking at, so this is a good and simple way to show items a customer is likely to be interested in. This is why these simple product-to-product recommenders are so successful and why these techniques are some of the key building blocks of modern personalization engines.

Secret #1 –Simple product-to-product recommendations seem personalized because they rely on recently viewed or purchased items from a customer’s history; then use that item to make “personalized” recommendations. This is commonly seen on Amazon where recently viewed or purchased items are shown and then similar or related items are displayed next to them. Again, this is a great technique when you have tons of data like Amazon. This technique is frequently used for homepage recommendations to get people re-engaged into their shopping journey or in emails to highlight products that are related to a recent purchase. This technique does NOT take in to account anything about the customer; including their preferences, location or other attributes to refine recommended items to align with the individual customer.

Secret #2 – Simple product-to-product recommenders are not that hard to develop. Today’s computer science courses teach about these recommender systems. The tools and collaborative filtering techniques are widely available as open source. For this reason, many commerce platforms and marketing clouds (and even commerce brands) have built simple product recommenders of their own. These vendors then sell the simple recommendations as a personalization engine. Unfortunately, their clients are looking for a feature and assume that what the vendor has to offer will be as good as any other. For the commerce brands that invest in product recommendations, they feel they can get a competitive advantage or cut costs by building their own systems. More on this in my next post.

Secret #3 – Simple product-to-product recommenders can give you pretty bad results especially on low traffic areas. Many marketers assume that they can deploy automated product recommendations and that they will perform equally well on all pages on their site. For sites with low traffic, large catalogs or lots of long tail products there can be serious issues with techniques that rely only on large volumes of browsing behavior alone. Not everyone can be Amazon. If not enough people browse or buy a product within the time that you are computing your models, there will be a very weak or no affinity between products. This is dangerous. While the system comes back with results, the items displayed are probably going to look at best silly or worst inappropriate. You can see this on many sites that have implemented simple recommendations as you browse around the site. You will see things that just don’t make sense and this is one reason why.

Secret #4 – Cart recommendations based on the last product placed in the cart. Simple product-to-product recommenders have a limitation in that they can only provide recommendations for a single product at a time. For example in the shopping cart, when a customer adds more than one item, these product recommenders will simply make recommendations on the most recently added item rather than on the set of items. Therefore, the recommendations can be hit or miss. If the customer is adding a single item or random items, this might work. But if the customer is building an outfit, the recommender can miss the mark by only looking at the latest item instead of the entire group of items. Again, the recommender would benefit from knowing more about the customer like what types of items they prefer, what brands they like or what items they already have in their wardrobe. But being a simple product-to-product recommender, this information is not part of the equation.

Product-to-product recommenders are a critical stepping stone for sophisticated modern personalization engines. What these simple recommenders lack is a true connection to the individual customer. Today’s personalization engines take recommendations to the next level by combing each customer’s browsing and purchase history to understand their preferences for things like categories, brands, colors, fabrics or styles. Personalization engines also examine customer similarity to determine items that particular customer might like based on what other people in their segments or location are interested in. This technique combined with an understanding of product-to-product affinity provides the true one-to-one personalized recommendations that many retailers and customer have been hoping for.