An exploratory look at supermarket shopping paths

Marketers often assume a stereotypical supermarket shopping pattern: shoppers systematically traverse aisles, deliberating and selecting items. However, this assumption lacks empirical support. This study uses a unique dataset from PathTracker®, which utilizes RFID tags on shopping carts to track shopper movement within a supermarket with unprecedented detail. The research aims to explore and identify typical in-store travel behaviors, challenging existing assumptions and providing a foundation for future research. The study acknowledges the limitations of using cart paths as a proxy for shopper movements but notes that the methodology remains applicable as more advanced tracking methods become available. Existing methods like principal component analysis have been used to analyze behavioral curves; however, this study focuses on clustering shoppers into "types" to describe prototypical paths, a task made complex by the numerous spatial constraints of the physical store layout. This necessitates the development and application of a novel clustering algorithm.

Prior research on shopper travel behavior in supermarkets is limited. Farley and Ring (1966) used a stochastic model to study zone-to-zone transitions. Mackay and Olshavsky (1975) studied consumer perceptions of store space, and Park, Iyer, and Smith (1989) examined the impact of store knowledge and time constraints on purchasing behavior. Underhill's (1999) anthropological work offers insights into shopper behavior but lacks the quantitative rigor of this study. Research on pedestrian movement in other settings, such as museums and malls (Batty, 2003), and work in environmental psychology (Winkel & Sasanoff, 1966) on pedestrian traffic flow offers relevant but not directly applicable methodologies. This study fills a gap by providing a detailed analysis of complete shopper paths, incorporating both spatial and temporal aspects.

The study uses a dataset of 27,000 shopper paths, each comprising a series of location coordinates (x, y) recorded every 5 seconds. Due to the dataset's size and the "ragged array" nature of the data (paths differing in length), a systematic sample of 8751 paths was used, excluding paths over 2 hours. To compare paths of varying lengths, each was recoded as a 100-percentile path, representing location at 1%, 2%, ..., 100% of the total path distance. Standard clustering techniques were unsuitable due to the spatial constraints of the store layout. Therefore, a *k*-medoids clustering algorithm was employed. This algorithm, adapted to handle spatial constraints, uses actual observed paths as cluster centroids, unlike *k*-means which can produce infeasible centroids. Euclidean distance, rather than travel distance, was used to measure path similarity, justified by high correlation between the two measures (Apparicio et al., 2003) and the simplification it provides. The store was divided into six zones: Racetrack, Aisles, Produce, Convenience Store, Checkout, and Extremity. To account for the time dimension, the 8751 paths were divided into three groups based on duration: low (2-10 minutes), medium (10-17 minutes), and high (17 minutes to 2 hours). The *k*-medoids algorithm was then applied separately to each group, using scree plots and the KL statistic to determine the optimal number of clusters. Cross-validation was performed on an additional third of the data to assess the stability of the clustering results.

The analysis revealed distinct shopping path types within each time group. For the low time group (2-10 min), two clusters emerged: one following a "default" path along the racetrack and the other deviating for quicker access to desired items. The medium time group (10-17 min) showed four clusters, with variations in racetrack usage, aisle visits (often short excursions), and time spent at the checkout. The high time group (17 min-2 hours) exhibited eight clusters, demonstrating greater heterogeneity in path types. Some clusters showed extensive racetrack use, others focused on specific aisles, and some demonstrated significant backtracking. The analysis revealed that the common assumption of systematic aisle-by-aisle shopping was not supported. Instead, most shoppers focused on select aisles or took short excursions into them. The perimeter of the store served as a main thoroughfare, with aisles accessed from it rather than a systematic back and forth pattern. Cross-validation confirmed the stability of the clustering results across different subsets of the data. Profiling the clusters based on zone percentages of the path provided additional insights into the distinct characteristics of each cluster, but it was noted that this method is subject to information loss compared to the more nuanced k-medoid approach.

The findings challenge common assumptions about shopper behavior in supermarkets. The dominant shopping pattern is not the systematic traversal of aisles; shoppers prioritize the perimeter and make selective, often short, visits into specific aisles. This has significant implications for product placement, with end-caps likely more effective than center aisle locations. The "racetrack" functions as a main thoroughfare, and shorter trips tend to focus on perimeter areas. The study highlights the importance of considering both spatial and temporal dimensions of shopper behavior. The methodology demonstrated the value of a spatially-constrained *k*-medoids clustering technique over simpler approaches that rely on summary statistics (e.g., zone-based profiling) in capturing the details of shopping path heterogeneity. The results offer valuable insights into the relationship between store layout, shopper behavior, and sales optimization.

This study provides a detailed, data-driven analysis of supermarket shopping paths using a novel dataset and methodology. The findings challenge common assumptions, revealing that systematic aisle traversal is less common than previously believed. The perimeter and selected aisles are key areas. The *k*-medoids clustering algorithm proved effective in handling spatial constraints and revealing subtle differences in shopper behavior. Future research could investigate the link between travel patterns and purchase decisions, develop more formal models of shopper movement, and extend the analysis to other retail environments.

The study uses cart paths as a proxy for shopper behavior, which may not perfectly capture individual shopper actions. The findings are specific to a single supermarket and may not generalize perfectly to other stores. The time spent shopping was categorized broadly into three groups; more granular time-based segmentation could potentially further refine the analysis.