‹ back to all posts
Product recommendations vs. Product discovery

Product recommendations vs. Product discovery

The concept of "product recommendations" in e-commerce is not something new. A decade ago, Amazon introduced products displayed in a carousel under the tagline: "Customers who viewed this item also viewed these other items". Following Amazon’s success, third-party vendors started to offer this feature as a service to retailers. It became so pervasive that the mere words "product recommendations" became synonymous with a carousel of similar products and branded this term as such. The caveat is that “product recommendations” as a branded term does not actually deliver personalized experiences that fresh and relevant.

To better define personalized fresh and relevant e-commerce, we have found that "personalized content ranking" is the better term. While a personalized set of products in a marketing campaign may include some product recommendations, selecting customer-specific merchandise for recurring customer experiences like promotional and lifecycle emails and website home pages is a completely different problem from finding similar products to a given SKU. A completely different type of technology is required to solve this upper funnel problem.

What is the current state for product recommendations?

While vendors like to use words like “machine learning” to describe the technology behind standard "Customers who viewed X also viewed Y" recommendations, the technology behind them could be explained in very simple terms. The idea is to find for any item X other similar items like X. The intuition behind the solution is that items similar to X should be frequently viewed together in the product catalog. Indeed, if I’m looking to buy a pair of running shoes, I’ll probably browse many similar pairs of running shoes in the same session. So to find shoes similar to X, I need to aggregate all other viewed items with X in the same sessions. The most frequent items should naturally be the most similar ones to X.

Product recommendation vendors commercialized this technology and offered “widgets” for retailers to put into product pages: items similar to the last browsed item, items similar to the last purchased item, items similar to the last shopping cart item, etc. Many other simple variations were added for the contexts other than the product pages: products sorted by sales, views, date, discount, or price. It’s basically just sorting mechanisms, but the vendors would frequently tout them as “algorithms” claiming to offer hundreds of different “algorithms”. In reality, there is only one algorithm behind all this—the algorithm of finding products similar to X, designed to close the funnel once the user zoomed in on a specific item.

Traditional recommendations

What is Personalized Content Ranking and how is it different?

Algorithmic discovery was first pioneered by Groupon (who was also our first customer). Groupon recognized early on that email has the opportunity to be more than a re-engagement channel but also a primary channel for engagement and discovery. While traditional retailers would wait for the consumers to come to their site to discover what to buy (transactional commerce), Groupon empowered consumers to engage and discover products through email. It’s not your standard “customers who viewed X also viewed Y” recommendations, but rather a fresh set of products personalized to each user for every experience. Not surprisingly, largely thanks to this more effective utilization of the email channel, Groupon was able to transition to the mobile era more effectively than other e-tailers, both effectively converting its large base of web users to mobile (110M out of 260M subscribers) as well as driving an increasingly higher share of transactions through mobile (54% in Q1 2014).

What is Adaptive Product Discovery?

The natural question is whether other retailers can replicate Groupon’s strategy with respect to email’s role in engagement and discovery. If so, then how? Here, we enumerate the key differences of personalized content ranking from the standard product recommendations.

1. No SKU Context

The first thing to realize is that 80% of email subscribers haven’t been to the retailer's site in weeks. Product recommendation widgets will surely offer you many different “algorithms” to personalize to one of the user’s last actions and the specific SKU in the context of the action: last product view, last shopping cart item, last purchase. But all of these actions and items individually will no longer be relevant. The user is no longer interested in whatever item was put in the shopping cart three months ago. Instead, you want to help the user discover new items while matching the general attributes drawn from a user's overall profile, without over-optimizing towards any individual item.

2. Repetitiveness: learning from the user’s passive behavior and iterating

Product discovery is about engagement and exploration and cannot get repetitive. The worst thing you can do for your subscriber engagement is keep showing the same items from “customers who bought X also bought Y” category or the same set of best sellers over and over again. If you pay attention to Groupon emails, you’ll see that the featured deals are very different from one email to another. Delivering a fresh and relevant experience requires tracking what content / products are displayed for each user in each position quickly learning from the user’s passive behavior about the products a user ignores.

3. Inventory performance

Personalization and product performance have historically been two separate dimensions for traditional product recommendations. Vendors will often split “most viewed items” or “best selling items” as separate algorithms from the traditional “customer who bought X also bought Y” configurations. However, product performance is instrumental to content ranking. Most products have a seasonality or trends and algorithms should pick it up automatically in the context of personalization. More importantly, as new items come in stock, the system should automatically prioritize the items that perform better vs. the ones that underperform.

4. Merchandising needs to go in hand with campaign’s business requirements

Email campaigns almost always feature some sort of a sale event, promotion, or a coupon: “10% off if you spend $100 or more”, “20% off all accessories”. It’s not that you need a new algorithm for each promotion, but rather you may want to limit the scope of featured products depending on the promotion. Meanwhile, you expect the system to implicitly apply all the other optimizations: personalization (trying to match the general attributes of products user has been shopping for), non-repetitiveness, dynamic optimization based on inventory performance, etc. Again, this type of behavior is very clunky to accomplish with traditional recommendation systems even though they may offer you a hundred different options.

Why is Personalized Content Ranking Important?

The next question on your mind could be whether personalized content ranking is something that you should be spending your resources on. Email drives on average 20-25% of online retail sales. The bulk of these sales comes from promotional email (the daily/weekly email campaigns make up more than 90% of retail email). Fresh and relevant experiences (enabled by adaptive content ranking) can drive up to 70% of incremental sales per each unit of email. Now if you do a simple multiplication, 20 x 70% represents 14% of incremental revenue for your business.

If this is not enough, consider the trends. While 66% of email opens happens on mobile devices, click-through and click-to-open rates on mobile are declining: 40% lower than on desktop. While lack of responsive design makes up for some of the decline, the ultimate issue is that the mobile users don’t want to bother searching/browsing for products on their phones and would rather see the relevant products directly in the email. As a result, unless retailers adopt the merchandised format powered by the algorithmic technology, they will continue to generate less revenue from email in the coming years as more and more users are shifting to mobile. 

A case in point is Groupon. The key to the company’s success on mobile in the last couple of years is not just its investments in the mobile app but the format of their email and the technology behind it. Product discovery on mobile is a perfect fit for email as traditional browsing continues to be clunky while search is making a poor transition to mobile (driving less than 30% of retail orders on mobile phones). The time is ticking as more and more users are shifting to mobile: email merchandising needs to be the cornerstone of your mobile strategy.