What makes recommendations work?
From Black-box AI to Communicable AI
By focusing on labelling and choice architecture, Communicable AI brings better results than any Black-box AI engine ever will.
The pitfall of the “Black-box AI”
If you haven’t ever integrated any product recommendations on your e-shop you might get lured into promises of vendors like “our AI-based recommendations can increase 20% to 30% of your revenue”. Every one who tried a recommendation engine, including the ones very satisfied with it, knows that turning it off would not reduce their turnover of such a percentage.
However, the idea of a very powerful AI generating the perfect recommendations comes back again, and again, always with the same reasoning “this time it works because our AI is much smarter”.
Before dismissing the core claim of this “All-powerful AI” argument, let’s give the devil its due:
AI is indeed a very important component of well functioning recommendations
(manual and rule-based approach will almost always underperform compared to a well tuned AI engine)Recent advances in deep learning (the most performant field of Machine Learning) show that applications in collaborative filtering are well suited to the field of recommendations and show better results than their predecessors
Managing product recommendations manually well (and even semi-manually) is not possible for almost any e-shop due to obvious resources limitations, so relying on recommendation algorithm is a must.
Therefore AI is definitely an important component of the story. We are not disputing this fact.
What is less obvious though is the idea that an “all-powerful AI” figuring out “what are the perfect product recommendations” for “every single visitor of your e-shop” is where the key opportunity for your e-shop is.
Our opinion on this matter is quite clear. We believe it is not.
The “what” versus the “how”, the “where” and the “why”
Until recently, all the focus of recommendations has been put on the “what”: what products should be recommended?
What is unfortunately not discussed is the diminishing returns on more and more sophisticated approaches. For example, a simple algorithm recommending what is most often bought together (arguably not even an AI algorithm and which most e-shop can easily compute for free) tend to perform at 70% to 80% of the level most advanced cutting-edge cross-selling machine learning algorithms.
One could argue, 30% (and possibly even a bit more) is quite substantial. But consider now the following aspects:
the WHERE
Displaying these recommendations in an engaging and visually dominant way at the right moment of the visitor’s journey (for example, as an overlay after the “add-to-basket” button is clicked) can increase the impact of these recommendations 2x to 5x compared to their traditional position in the basket page
the WHY
If you are a Netflix user, you know that their personalization system is heavily oriented towards grouping their recommendations in many sections, each with a clear and engaging label. It is not only telling you what you might be interested in watching, it tells you why these recommendations are presented to you.
Such communication can sometimes be more important that the recommendations themselves. Imagine a wine-shop giving you recommendations of the “best matching wines”. If you are not a wine expert, how will you judge if they are any good? why should you engage with them? They just look like any other wines. But telling you “casual drinkers who like Chardonnay loved these wines” would give you a totally different perspective on why they are good matches.
Simply generating the label of your recommendations in such a way can increase 2x to 3x their impact.
the HOW
When scientists like Richard Thaler built the foundation of behavioral economics, in what resulted in the (relatively) simple idea of a “nudge”, their approach rooted in something deeper called “choice architecture”.
One of the user benefits of recommendations is their convenience: I don’t have to look for these products myself, they are presented to me directly. This is even more true when adding the fact the user might not even have thought of looking for such products at all.
But let’s think about “choice architecture” when optimizing recommendations. It comes quite quickly that, while sometimes, the choice is best suited between specific products (which the traditional product recommendations focus on) it is often the case that the real choices are between different groups of products (if I buy a sofa, I might want a table as well, but I also might want some cushions or some covers, these are not different products, but groups of products).
Providing several groups of suggestions with appropriate explanations in their label and letting the visitors “see more” of each group by going to a dedicated listing page can further increase 1.5x to 2.5x the impact (on top of the prior section’s effects).
The power of “Communicable AI”
We summarize this capacity to understand where to display recommendations visually (the WHERE), communicating why the user sees such recommendations (the WHY) and the choice architecture of what types of recommendation groups should appear where and for whom (the HOW) under the term “Communicable AI”.
Boxalino platform integrates many different AI algorithm to generate state-of-the-art product recommendations (the WHAT) and we optimize our platform to tune and test such algorithm in highly effective and impactful way.
But what makes us truly innovative is our Communicable AI which will make a real difference in your e-shop.
On top of the key advantages listed above, Communicable AI comes with many additional advantages:
Better Self-learning
In addition to knowing what product people click and buy to improve the recommendations further (as all good AI providers do), Communicable AI enables Boxalino to learn much faster and better by identifying on what labels people engage with and which dedicated listing page are most visited and used. All these feed-back information give highly valuable information about what works or not and why.Better Personalization
By letting the user rate the suggestions and their groups (labels) by giving thumbs up/down feed-backs or adding such labels in their wish-list, Communicable AI can profile your customers not only based on what product they click on, but what motivating reasons are most suited for their needs.Better Reporting
How should a product recommendations report look like? This is a tough one, as you can’t analyze each context / profile performance for each product (they are too many of them) and grouping by categories or brands might hide the key information of what works or not. However, reporting on the performance of the different labels both in the context of cross/up selling recommendation, but also in the context of the visit and impact of their dedicated listing page is much more effective and useful as a report.Better Marketing
If the dedicated listing page related to any of the product recommendation labels drives many clicks and has a good conversion rate you might want to use such pages as landing page for your marketing activities?
Boxalino Winning Interactions platform goes even beyond this point, fully automating not only the generation of such “topic” landing pages, but also automating the generation of ads bringing you visitors to your web-site on such pages.