Abstract
Purpose: As e-commerce platforms increasingly rely on algorithmic recommendation systems (RS) to personalize user experience and drive sales, concerns about inherent biases and their potential to create 'filter bubbles' have grown. This study investigates the tangible impact of algorithmic bias, specifically popularity bias, within e-commerce RS on consumer purchasing behavior. Methodology: This research employs a mixed-methods approach. First, an observational study (Study 1) analyzes 18 months of anonymized transaction and recommendation data from a large e-commerce platform (N=110,452 users). We developed a Recommendation Exposure Bias (REB) metric and used panel data regression to assess its relationship with purchase diversity, serendipity, and platform engagement. Second, a controlled online experiment (Study 2; N=512) with a between-subjects design (High-Bias vs. Low-Bias vs. Control recommendations) was conducted to establish causality between biased recommendations and purchasing decisions in a simulated shopping environment. Findings: The observational data reveals a statistically significant negative correlation between exposure to biased recommendations and the diversity of products purchased over time. Higher REB is associated with a contraction in the number of unique product categories a consumer explores and buys from. The experimental results corroborate this, demonstrating that participants in the High-Bias condition selected a significantly less diverse basket of goods compared to those in the Low-Bias condition. Furthermore, the High-Bias group reported lower perceived choice satisfaction and reduced trust in the recommendations. Conclusions and Implications: Algorithmic bias in recommendation systems is not a benign technical artifact; it actively shapes consumer purchasing behavior by creating economic echo chambers that narrow consumer tastes and limit product discovery. While potentially optimized for short-term engagement with popular items, this practice risks long-term customer dissatisfaction and reduced lifetime value. Our findings provide a quantitative basis for managerial action, urging firms to adopt more sophisticated metrics that balance personalization with exploration and to design more transparent and controllable recommendation engines.