The conversion funnel is a model describing the stages consumers go through in their journey towards a purchase. This journey often lasts several days to weeks and can include multiple visits to a seller’s website. A large body of literature has focused on using observable search patterns to identify consumers’ hidden purchasing stages and to estimate their likelihood of conversion. We propose a novel set of measures to better unveil the consumer’s hidden stage in the funnel. These measures are based on the diversity of the searches that a customer engages in while browsing an e-commerce website, and they include not only the number of different products that are searched for, but also measures that rely on unobserved similarities among products, captured in a product network (in which products are assumed to be “similar” if they are frequently co-searched). We operationalize and evaluate our proposed measures using a large-scale dataset from a medium-sized tourism website used for comparing and booking flights. We estimate a hidden Markov model to show that our proposed diversity measures are associated with progress in the funnel and consumers' conversion likelihood. Specifically, we show that consumers go through different distinguishable stages (states) in their journey, characterized by different values of our proposed diversity measures. To demonstrate the managerial and business implications of our theory, we show that incorporating search-diversity measures into a baseline prediction model significantly improves the model’s performance in predicting purchase likelihood and churn.
We compare mobile and desktop user progress through the conversion funnel. Using detailed log-files of an online flight search engine, we analyze consumer search behavior and model the stages consumers go through along the funnel in both mobile and desktop platforms. We ask: Do mobile and desktop users go through similar funnel stages? Are the funnel stages of both platforms characterized differently? To analyze the progress through the funnel, we use hidden Markov models (HMM) that capture latent funnel stages based on observable search behavior variables indicative of convergence towards a purchase. We find that while similar stochastic processes characterize funnel stages on both platforms, the distribution of visits across stages in the two platforms is significantly different: a larger percentage of mobile visits resides in more advanced stages, indicating that mobile consumers use the website when they are more advanced in the funnel, where less cognitive effort is required.
We investigate and compare online consumer behavior on an e-retailer website in mobile versus PC devices, through the application of a web usage mining approach on clickstream data recorded in server-side log files. Online consumer behavior is characterized through both engagement measures and the discovery of common sequences of navigation patterns, using an innovative approach that combines footstep graph visualization with sequential association rule mining. We find that sessions conducted through mobile devices are more likely to consist of task-oriented behavior whereas sessions conducted through PC devices are characterized by a more exploration-oriented browsing behavior. Moreover, we find that certain sequence rules are associated with an increased likelihood of purchase in both mobile and PC sessions. The results demonstrate the value of our approach in analyzing online browsing behavior, across platforms, in the context of electronic retailing.
E-commerce websites sell thousands of different products and require tools that support product comparisons and intelligent search capabilities. The ability to measure the similarity between products is pivotal for such tools. Over the years, many similarity measuring techniques have been suggested, user-based and content-based. Still, most techniques require previously collected data or prior domain-knowledge. This paper proposes an innovative approach that requires no prior knowledge of the inner relations among products, in our case, flight-destinations. Our approach builds on concepts of image-recognition and natural language processing (NLP) to extract hidden aspects of the products. As a testbed, we use data acquired from a tourism website where products are flight-destinations. We analyze similarity metrics based on state-of-the-art methods for image recognition, NLP, and product–network analysis, and compare them to those obtained by human subjects. We find that there is no one method that dominates the others in every aspect. Thus, we provide a hybrid approach, that brings the best of each.