As Internet users, we receive many offers for multiple products every day. Robots send them to us through various communication channels. How do advertisers choose which products to show and decide how to reach us?
Photo: Het Nieuwe Instituut (license Creative Commons)
We generally call these mechanisms “recommendation engines” and, of course, they are based on data!
For end users, the promise of this technology is that “if you liked this, you may like that," but it is actually quite a bit more complex than that. A good recommendation engine is often a mix of different algorithms, corresponding to several basic principles of purchasing behavior. Read on, or check out our how-to guidebook for a complete look at what it takes to build a recommendation engine from the ground up.
Oblivion (Remarketing)
Remarketing is the idea of pushing some content to the user, specifically content that they have viewed but not yet purchased. They might have forgotten about it, still be hesitating, and so on. Remarketing generally delivers good performance, but is limited in the number of users for whom it is relevant.
Content Similarity (Recommendation by Similar Characteristics)
Recommending by similar content is about pushing products that are similar to other ones that the user has liked or bought, i.e. that share similar characteristics. For example, if you previously liked a red sweater, you could get suggestions for other red sweaters.
Content similarity is generally the first thing that comes to mind when talking about recommendations. While it is very useful for users who are looking for alternatives, overusing it can lead to always showing the same content to the user instead of expanding their exposure to your products.
Profile Similarity (Recommendation by Proximity)
Recommending by profile similarity focuses on the actual behavior of users (“who bought what?”) by comparing their purchase history. We say that products are similar, not if they share the same features, but if they are actually bought by the same people. Therefore, users are “close” if they bought or were interested in the same products. This type of recommendation is very effective for products without well-defined features, such as cultural goods.
Complementarity (Recommendation by Short-Term Sequence of Purchase)
Recommending by short-term purchasing sequence is when you study which products are frequently bought together (same cart, same day) by many different users. Using it, we can find complementary products (a TV and its cable, a console and its games, a pair of earrings and a necklace). This type of recommendation is generally used when adding items to a shopping cart, to suggest additional related products.
Evolution (Recommendation by Long-Term Sequence of Purchase)
In this kind of recommendation, we look at purchasing sequences over several weeks or months. It involves pushing content corresponding to a change in life: having a baby, moving, switching jobs, etc. This type of recommendation is particularly interesting for generating repetition in sales for users with the same demographic.
Popularity
Last but not least, recommendation by popularity consists of pushing products that are most likely to be globally purchased (by brand, trend, etc.).
The Need for a Good Mix
Like an audience acquisition campaign, which requires a marketing mix, a good recommendation strategy should make good use of the principles mentioned above.
Other criteria should also be taken into account when designing your recommendation:
- Location: On the homepage, in a newsletter, an email reminder, on the shopping cart page, the product page, in an ad, etc.
- Relation to the customer: A regular visitor who has not bought anything yet, a recent buyer, a former buyer... It's likely that they don’t all respond to the same solicitations.
- Editorial consistency: You may wish to only feature popular or highly visual products on your main page, or, on the contrary, to only show surprising recommendations.
A good strategy is to use two or three of the principles mentioned above depending on well-understood contexts. For example:
- Content similarity and popularity for the homepage
- Complementarity and remarketing for currently purchasing visitors.
- Profile similarity for cultural products
- Evolution and popularity on email campaigns for former but recurrent buyers
"Recommendation" has multiple meanings, but let's not forget common sense: the customer is not easily fooled. Visitors on a website are more and more aware of the solicitations. Some users already seem tired of suggestions and start rejecting any advertising presence or are afraid of being tracked.
Once again, technology is bound to evolve. One day, hopefully, recommendation engines may behave like your favorite store salesman, understanding when it is not the right time to bother you and letting you stroll in peace.