With the global e-commerce market expected to bring in $5.5 trillion this year (a 17.9% increase since 2020), all eyes are on better understanding and managing consumer behavior online. Traditionally, retailers have relied on reliable and time-dependent utility maximization choice models, such as the multinomial logit model, which date back decades to evaluate their marketing policies. With this approach, it was assumed that customers had the time and patience to review all of the products on offer. These models are elegant and have been validated in different contexts as an accurate measure of customer search and purchase behavior.
But that’s outdated: existing models may no longer make sense in a world of near-limitless online choices. A more realistic reflection of customer choice behavior centers on a sequential, top-down approach in which consumers do not inspect all products. This evolution is answered by what is called the “cascading click model”. Since research published by Nick Craswell et al in 2008, this has been widely integrated into the global e-commerce industry, particularly in online advertising and product recommendation systems. Now, our research suggests that the Cascade Click model should be extended to online retail settings with potentially substantial benefits for all.
The price and ranking of a product are among the most important decisions in e-commerce. This requires a good understanding of a customer’s behavior so retailers can make the right pricing and ranking decisions. Thus, incorporating a better approximation of the actual customer’s search and purchase behavior into pricing algorithms seems essential in the era of e-commerce.
We are four academics who developed algorithms that could increase the expected revenue of online retailers by up to 13%. To do this, we believe that the “click and search” algorithms used by retailers should be rebuilt to put pricing and ranking models at their very heart.
Critical management and leadership roles
For this to happen, managers and business leaders need to understand how customers search and buy online. Using standard traditional algorithms to learn a consumer’s click and purchase behavior is likely to result in slow convergence which, in practice, results in a huge loss of potential revenue.
But managers can tailor click and search pattern parameters that are consistent with customer behavior by focusing on both the sequential approach to customer behavior and what we call the positional bias effect. If customers are looking for, say, a 12-megapixel camera, they are unlikely to sift through every one of the thousands of products on offer. Thus, the display position and ranking of products on the webpage will significantly influence the likelihood of a product being clicked and purchased.
After understanding the real behavior of customers, managers have to modify the parameters and functions of the model which, at the moment, is sometimes guided by black box learning algorithms. Business leaders must also operate with an awareness of the seasonality and short life cycles of many products. Indeed, the sooner managers realize the true parameters of the Cascade Click model, the sooner they are able to exploit their products and put an accurate price on them – before the end of the season!
Equally important is the need for managers to synchronize prices and rankings. Managers must coordinate these two sectors that are too often separated within companies. We have extended the new-look algorithm to take into account common product ranking and pricing. This helps us prove the performance of the algorithm and its benefits for retailers. In our simulations, we find that joint decisions on rank and price can contribute up to an eighth of additional revenue, highlighting the importance of managerial coordination in firms.
Dynamic pricing with limited stocks
Integrating cascading click patterns into operational settings remains a relatively new practice, and the possibilities for further study of click and search patterns are rich. For example, we went on to explore dynamic pricing with limited stocks, another vital choice facing managers, this time in the hospitality and fashion industries. In this case, we assume that each product has a limited amount of inventory, for example the number of flight seats in each class of an airliner. How do managers decide what price to charge if they use the Cascade Click model? This question applies to other finished sales, such as fashion products or hotel rooms. We also suggest that further research could be conducted on how companies can provide optimal assortment sets.
All of this underscores the importance of better understanding customer click and search actions. Behavioral pattern analysis is essential to help businesses make various operational and marketing decisions in the context of large-scale web analytics. This includes product recommendation, pricing and ranking of displays, etc. As the e-commerce market grows exponentially, we hope the research will provide a better understanding and, therefore, a smoother approach to online sales and purchases.
Sajjad Najafi is a lecturer in operations management at HEC Paris
Daniel Brown is editor-in-chief at HEC Paris