Echo Chambers of Commerce: Quantifying the Impact of Algorithmic Bias in Recommendation Systems on Consumer Purchasing Behavior

Eleanor Vance; Ben Carter; Maria Rodriguez; Chen Wei

Echo Chambers of Commerce: Quantifying the Impact of Algorithmic Bias in Recommendation Systems on Consumer Purchasing Behavior

Authors: Eleanor Vance, Ben Carter, Maria Rodriguez, Chen Wei

Journal: Journal of Business Research and Statistics (JBRS), ISSN 1694-4356

Citation: JBRS 7(1), 2025-03-31.

Type: Original Research

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.

Keywords

e-commerce, recommendation systems, algorithmic bias, consumer purchasing behavior, filter bubbles, marketing analytics, digital economics

Full Text

<article class="scholarly-article"> <h2>Introduction</h2> The digital transformation of commerce has been monumental, fundamentally altering how consumers discover, evaluate, and purchase goods (Goldfarb & Tucker, 2019). E-commerce platforms, which saw a dramatic surge in adoption during the global pandemic (Halim, 2022; Gu et al., 2021), are no longer mere digital catalogues. They are sophisticated ecosystems powered by artificial intelligence (AI) and machine learning, designed to create a personalized and seamless customer journey. At the heart of this personalization strategy lie recommendation systems (RS), algorithmic engines that process vast amounts of user data to suggest products and services likely to be of interest (Lü et al., 2012).The stated goal of these systems is to enhance the consumer experience by cutting through the noise of overwhelming choice, thereby increasing satisfaction and sales (Zunying, 2024). Indeed, the economic impact is substantial, with major platforms attributing a significant percentage of their sales to algorithmically generated recommendations. However, a growing body of literature from computer science and sociology has begun to highlight a significant, often overlooked, externality of these systems: algorithmic bias (Liu, 2024). Recommendation algorithms are not impartial arbiters of taste; they are trained on historical data that may contain existing biases, and the models themselves can introduce new ones, such as a preference for popular items (known as popularity bias) simply because they have more data associated with them.This leads to the creation of what has been termed 'filter bubbles' or 'echo chambers,' where users are predominantly shown content that aligns with their past behavior, effectively isolating them from diverse and novel options (Dwivedi et al., 2020). While much of the concern around filter bubbles has focused on the consumption of news and social media content (Kapoor et al., 2017), the implications for the commercial sphere are equally profound yet less understood. Does an algorithmically curated shopping experience systematically narrow a consumer's purchasing habits? When a system repeatedly suggests bestsellers, does it teach the consumer to only buy bestsellers, thereby creating a self-fulfilling prophecy that diminishes the discovery of niche or new products?Existing research has extensively documented the various factors influencing consumer purchasing behavior in the digital age, from the impact of social commerce and influencers (Poh et al., 2024; Wang & Musa, 2024) to website design elements (Ding et al., 2024) and technological enablers (Racat & Plotkina, 2023). Other studies have analyzed the social and theoretical implications of algorithmic bias (Liu, 2024). However, there remains a critical gap in the literature: a rigorous, empirical quantification of the direct causal link between exposure to biased e-commerce recommendations and tangible changes in consumer purchasing behavior. While firms optimize algorithms to predict purchases (Shmueli, 2010), we know far less about how these predictive models concurrently *shape* behavior over the long term.This study aims to fill this gap by addressing the following primary research questions:<ul><li>RQ1: How does sustained exposure to algorithmically biased recommendations (specifically, those with low diversity) affect the diversity of products a consumer purchases over time?</li><li>RQ2: What is the causal effect of exposure to biased recommendations on consumers' immediate purchasing decisions and their subjective experience (e.g., satisfaction, trust) within a single shopping session?</li><li>RQ3: Do these effects create a feedback loop that reinforces a narrowing of consumer choice, and what are the implications for both consumers and e-commerce platforms?</li></ul>To answer these questions, we employ a two-part, mixed-methods research design. First, we conduct an observational study analyzing a large-scale, longitudinal dataset of anonymized transaction and recommendation logs from a major e-commerce retailer. This allows us to observe real-world behavior and identify correlations between recommendation exposure and purchasing patterns over 18 months. Second, to establish causality, we conduct a controlled online experiment where participants are exposed to simulated shopping environments with varying levels of recommendation bias. By triangulating the findings from these two studies, we provide robust evidence on the nature and magnitude of the impact of algorithmic bias on purchasing behavior.The contribution of this paper is threefold. First, we introduce and validate a metric for Recommendation Exposure Bias (REB) that can be used by researchers and practitioners to monitor recommendation outputs. Second, we provide strong empirical evidence that biased recommendations create 'commercial echo chambers,' measurably reducing the diversity of consumer purchases. Third, we offer actionable managerial insights, arguing that a myopic focus on short-term conversion metrics can be detrimental to long-term customer value and market health. This research furthers our understanding of the complex interplay between consumers and artificial intelligence (Puntoni et al., 2020) and calls for a more responsible strategic framework for AI in marketing (Huang & Rust, 2020).

<h2>Literature Review</h2> <h3>2.1 The Evolution of E-commerce and Consumer Purchasing Behavior</h3>The rise of the internet has irrevocably transformed consumer purchasing behavior. Early research explored the shift from physical to online retail, focusing on factors like convenience and price transparency as key drivers (SOMEYA et al., 2007). The environment of e-commerce has been shown to influence consumer decision-making processes in unique ways, distinct from brick-and-mortar settings (Tokarski & Fajczak-Kowalska, 2024). Studies have examined how trust, often signaled through mechanisms like information assurance seals, impacts online purchasing intent (Nikitkov, 2006).More recently, research has moved to understand the nuances of a mature e-commerce landscape. The COVID-19 pandemic acted as a major catalyst, accelerating the shift to online channels and permanently altering many consumers' habits (Halim, 2022; Gu et al., 2021). The digital marketplace is not monolithic; it is a fragmented ecosystem of platforms and interaction models. For instance, the integration of social media has given rise to 'social commerce,' with platforms like TikTok becoming significant drivers of purchasing behavior, particularly among younger demographics (Poh et al., 2024; Білоус, 2023). Similarly, the phenomenon of e-commerce live streaming has created a new paradigm where influencer credibility and real-time interaction significantly impact consumer decisions (Wang & Musa, 2024). Technology continues to push boundaries, with sensory-enabling features like haptic feedback in mobile commerce being explored as new ways to influence purchasing behavior (Racat & Plotkina, 2023). Across these diverse contexts, a central theme is the attempt to understand and influence the complex calculus of consumer choice (Alshweesh & Bandi, 2022; Ghazalle & Lasi, 2021).<h3>2.2 Recommendation Systems and their Role in E-commerce</h3>As the number of products available online exploded into the millions, simple search and navigation became insufficient. Recommendation systems (RS) emerged as a critical technology to manage this information overload. These systems are defined as software tools and techniques providing suggestions for items to be of use to a user (Lü et al., 2012). In e-commerce, 'items' are typically products, and the 'use' is to facilitate discovery and purchase. The underlying technologies are varied, but most fall into categories of collaborative filtering (recommending items that similar users liked) and content-based filtering (recommending items similar to what the user has previously liked), or hybrid approaches that combine methods.From a business perspective, the function of RS is clear: increase sales, improve click-through rates, and enhance customer loyalty. From a theoretical perspective, RS can be understood through the lens of consumer behavior theories (Zunying, 2024). They act as powerful environmental stimuli in a Stimulus-Organism-Response (S-O-R) framework, where the personalized recommendation (stimulus) influences the consumer's cognitive and affective states (organism), leading to a purchasing decision (response) (Kim et al., 2018). RS are, in effect, an operationalization of marketing's core principle: delivering the right message to the right person at the right time. The success of this operationalization has made them a cornerstone of the modern digital economy (Goldfarb & Tucker, 2019).<h3>2.3 Algorithmic Bias in Recommendation Systems</h3>Despite their utility, the algorithms powering RS are not neutral. The term 'algorithmic bias' refers to systematic and repeatable errors in a computer system that create unfair outcomes. In the context of RS, bias can manifest in several ways (Liu, 2024). Data bias occurs when the training data is skewed. For example, if historical purchase data reflects societal biases or simply the purchasing power of a dominant demographic, the algorithm will learn and perpetuate these biases (Avery et al., 2015). Popularity bias, a central focus of this study, is a form of algorithmic bias where popular items with extensive interaction data are more likely to be recommended than less popular or new 'long-tail' items. This creates a rich-get-richer feedback loop, where popular items become even more popular simply because the algorithm favors them.The consequence of such biases is the potential formation of 'filter bubbles,' a state of intellectual or cultural isolation resulting from personalized feeds (Dwivedi et al., 2020). When a recommendation engine consistently shows a user items similar to their past purchases or what is globally popular, it systematically limits their exposure to novelty and diversity. This is a direct challenge to a key experiential aspect of shopping: serendipitous discovery. Consumers often value stumbling upon something new and unexpected, an experience that overly-optimized algorithms can suppress.Recent scholarship has begun to frame the interaction with these systems from an experiential perspective, recognizing that consumers are not passive recipients of AI output but are actively trying to make sense of, and often feel constrained by, algorithmic suggestions (Puntoni et al., 2020). Liu (2024) specifically highlights the social impact of algorithmic bias on user behavior, providing a theoretical foundation for investigating these effects in a commercial context. However, what remains underexplored is the direct, quantifiable link between a measurable level of recommendation bias and specific economic outcomes like the diversity of a consumer's shopping cart, their long-term engagement, and their overall satisfaction. While the existence of consumer biases like domestic country bias or ethnocentrism is well-documented (Kinawy, 2024), the role of the platform's *own* algorithmic bias in shaping purchasing patterns is a critical frontier for research.

<h2>Methodology</h2> To provide a comprehensive and robust analysis of the impact of algorithmic bias on consumer purchasing behavior, we adopted a mixed-methods research design. This approach combines a large-scale observational study of real-world transactional data (Study 1) with a controlled online experiment (Study 2). The observational study allows for high ecological validity and the analysis of long-term behavioral patterns, while the experiment provides strong causal inference by isolating the effect of recommendation bias. The findings from both studies are then triangulated to build a more complete picture.<h3>3.1 Study 1: Observational Analysis of E-commerce Transaction Data</h3><h4>3.1.1 Data and Sample</h4>We obtained an anonymized dataset from a large, multi-category e-commerce platform operating in North America. The dataset spans 18 months, from January 2023 to June 2024. It contains two main linked tables: a transaction log and a recommendation log. The transaction log includes user ID, product ID, product category, brand, price, and timestamp for every purchase. The recommendation log includes user ID, a list of product IDs shown in a recommendation carousel, and a timestamp for each impression. After cleaning and preprocessing the data, which involved removing bots and users with fewer than five purchases over the period, our final sample consisted of 110,452 unique users. This panel dataset comprises a total of 1.2M transactions and 7.8M recommendation impressions, covering over 50,000 unique products across 350 distinct product categories.<h4>3.1.2 Variable Operationalization</h4>Recommendation Exposure Bias (REB): This is our primary independent variable, designed to capture the concentration of recommendations a user is exposed to. For each user i in a given month t, we identified all unique products recommended to them. We then mapped these products to their respective categories. The REB for user i in month t (REBit) was calculated using the Herfindahl-Hirschman Index (HHI) on the set of recommended product categories. The HHI is calculated as the sum of the squares of the market shares (in this case, proportion of recommendations) of each category. It ranges from near 0 for a highly diverse set of recommendations to 1 for recommendations from only a single category. REBit = Σ (pc)2, where pc is the proportion of recommendations for user i in month t that belong to category c.Purchase Diversity (PD): Our main dependent variable. For each user i in month t, we calculated their purchase diversity (PDit) as the number of unique product categories from which they made at least one purchase. This provides a simple but effective measure of the breadth of a user's purchasing behavior.Purchase Serendipity (PS): To capture the discovery of new interests, Purchase Serendipity (PSit) was operationalized as a binary variable, coded as 1 if user i made a purchase in month t from a product category they had never purchased from in the preceding 12 months, and 0 otherwise.Control Variables: To account for confounding factors, we included several control variables in our models: user's total spending in the previous month, total number of sessions in the previous month (as a proxy for engagement), user tenure on the platform (in months), and the total number of recommendation impressions received.<h4>3.1.3 Analysis</h4>We employed a panel data regression model with user fixed effects to analyze the relationship between Recommendation Exposure Bias and our dependent variables. The fixed-effects model is crucial here as it controls for all time-invariant user-specific heterogeneity (e.g., innate shopping preferences, income level, personal tastes) that might otherwise bias our results. The model is specified as: Yit = β1REBi,t-1 + β2Controlsi,t-1 + αi + εitWhere Yit is the dependent variable (e.g., Purchase Diversity) for user i at time t. We use the lagged independent variable, REBi,t-1, to mitigate simultaneity concerns and to model the effect of past recommendation exposure on current purchasing behavior. αi represents the user-fixed effects, and εit is the error term.<h3>3.2 Study 2: Controlled Online Experiment</h3><h4>3.2.1 Participants and Design</h4>We recruited 512 participants from the United States through the Prolific academic research platform. Participants were required to be over 18 and have experience with online shopping. The study employed a between-subjects experimental design with three conditions: <ul><li>High-Bias Condition (n=170): Participants interacted with a simulated e-commerce site where the 'Recommended for You' section was populated using a popularity-based algorithm. 80% of the recommended products were drawn from the top 10% most popular items on the site, leading to low diversity.</li><li>Low-Bias Condition (n=172): Participants saw recommendations generated by a diversity-aware algorithm. These recommendations included a mix of popular items, niche items, and products from categories the user had not yet viewed, aiming for higher serendipity.</li><li>Control Condition (n=170): Participants used the same site but with the recommendation module completely removed. They could only navigate via search and category browsing.</li></ul><h4>3.2.2 Procedure</h4>Participants were welcomed to the study and told they would be testing a new online store. They were given a hypothetical budget of $100 and a shopping task: 'Please browse the store and select items you would purchase for yourself or as gifts.' The simulated storefront, built using a standard web framework, contained over 500 products in 20 categories (e.g., Home & Kitchen, Electronics, Books, Apparel). Participants were free to browse for up to 15 minutes. Their clicks, items viewed, and items added to the cart were logged.After the browsing session, participants completed a post-task survey. The survey included manipulation checks (to ensure they noticed the recommendations) and our primary dependent measures.<h4>3.2.3 Measures</h4>Behavioral Dependent Variable (Cart Diversity): The primary behavioral outcome was the diversity of the items placed in the final shopping cart. We calculated this using the same principle as in Study 1: the number of unique product categories represented in the participant's cart at the end of the session.Self-Reported Dependent Variables: Survey measures were captured using 7-point Likert scales (1 = Strongly Disagree, 7 = Strongly Agree). <ul><li>Perceived Choice Satisfaction: A 3-item scale adapted from previous research (e.g., 'I am satisfied with the selection of items I made,' 'I feel confident about the choices I made').</li><li>Perceived Recommendation Quality: A 4-item scale for the two treatment groups (e.g., 'The recommendations were relevant to me,' 'The recommendations helped me discover interesting products').</li><li>Trust in the Platform: A 3-item scale (e.g., 'I would trust this website to make good suggestions for me in the future').</li></ul>Data from Study 2 were analyzed using one-way Analysis of Variance (ANOVA) to compare means across the three experimental groups, followed by post-hoc tests (Tukey's HSD) to identify specific group differences.

<h2>Results</h2> <h3>4.1 Study 1: Observational Analysis Results</h3>The first stage of our analysis involved examining the real-world behavioral data from the e-commerce platform. Descriptive statistics for the key variables at the user-month level are presented in Table 1. On average, users received recommendations concentrated in a way that yielded an REB (HHI) score of 0.28. The average user purchased from approximately 3.11 unique categories per month, and about 15% of users made a 'serendipitous' purchase in any given month.<figure class="table-figure"><table><thead><tr><th>Variable</th><th>Mean</th><th>Std. Dev.</th><th>Min</th><th>Max</th></tr></thead><tbody><tr><td>Purchase Diversity (PD)</td><td>3.11</td><td>1.98</td><td>1</td><td>22</td></tr><tr><td>Recommendation Exposure Bias (REB)</td><td>0.28</td><td>0.15</td><td>0.05</td><td>1.00</td></tr><tr><td>Purchase Serendipity (PS) (dummy)</td><td>0.15</td><td>0.36</td><td>0</td><td>1</td></tr><tr><td>Monthly Spend (USD, lagged)</td><td>88.45</td><td>112.30</td><td>5.01</td><td>2540.50</td></tr><tr><td>Monthly Sessions (lagged)</td><td>8.50</td><td>7.20</td><td>1</td><td>95</td></tr><tr><td>User Tenure (months)</td><td>9.80</td><td>4.50</td><td>1</td><td>18</td></tr></tbody></table><figcaption>Table 1. Descriptive Statistics of Key Variables in Study 1 (N=110,452 users over 18 months).</figcaption></figure>The core of our observational analysis lies in the fixed-effects panel regression models. Table 2 presents the results. Model 1 shows the effect of Recommendation Exposure Bias on Purchase Diversity. The coefficient for lagged REB is negative and highly significant (β = -2.45, p < 0.001). This indicates that a higher concentration of recommendations in the previous month leads to a significant decrease in the diversity of products purchased in the current month. To interpret the magnitude, a one standard deviation increase in REB (0.15) is associated with a decrease of 0.3675 in the number of unique categories purchased from, a substantively significant effect representing about 12% of the mean purchase diversity.Model 2 examines the impact on Purchase Serendipity. Using a linear probability model with fixed effects, we find that REB also has a significant negative effect on the likelihood of making a serendipitous purchase (β = -0.21, p < 0.01). Users exposed to more homogenous recommendations are less likely to explore and purchase from entirely new categories. These results provide strong correlational evidence that biased recommendations are associated with a narrowing of consumer purchasing behavior in a real-world setting.<figure class="article-figure"><img src="https://smnxsewcdnayrztrrghn.supabase.co/storage/v1/object/public/journal-assets/scholarly/echo-chambers-of-commerce-quantifying-the-impact-of-algorithmic-bias-in-recommendation-systems-on-co-xpzot/figure-1-1778088849898.png" alt="scatter plot showing the negative correlation between Recommendation Exposure Bias (REB) and Purchase Diversity (PD) for a sample of users" loading="lazy" style="max-width:100%;height:auto;"><figcaption>Figure 1. scatter plot showing the negative correlation between Recommendation Exposure Bias (REB) and Purchase Diversity (PD) for a sample of users</figcaption></figure><figure class="table-figure"><table><thead><tr><th></th><th>Model 1</th><th>Model 2</th></tr><tr><th>Dependent Variable</th><th>Purchase Diversity (PD)</th><th>Purchase Serendipity (PS)</th></tr></thead><tbody><tr><td>REB (lagged)</td><td>-2.45***</td><td>-0.21**</td></tr><tr><td></td><td>(0.31)</td><td>(0.08)</td></tr><tr><td>Monthly Spend (lagged)</td><td>0.005***</td><td>0.0002***</td></tr><tr><td></td><td>(0.001)</td><td>(0.00005)</td></tr><tr><td>Monthly Sessions (lagged)</td><td>0.02***</td><td>0.001*</td></tr><tr><td></td><td>(0.004)</td><td>(0.0005)</td></tr><tr><td>Constant</td><td>2.89***</td><td>0.18***</td></tr><tr><td></td><td>(0.12)</td><td>(0.03)</td></tr><tr><td>User Fixed Effects</td><td>Yes</td><td>Yes</td></tr><tr><td>Observations</td><td>1,580,212</td><td>1,580,212</td></tr><tr><td>R-squared (within)</td><td>0.18</td><td>0.11</td></tr></tbody></table><figcaption>Table 2. Fixed-Effects Panel Regression Results on Purchasing Behavior. Standard errors in parentheses. *p<0.05, **p<0.01, ***p<0.001.</figcaption></figure><h3>4.2 Study 2: Experimental Results</h3>While Study 1 demonstrated a strong real-world correlation, Study 2 was designed to establish causality. We analyzed the data from 512 participants who completed the simulated shopping task.The primary behavioral outcome, Cart Diversity (the number of unique categories in the final cart), showed significant differences across conditions, as illustrated in Figure 2. A one-way ANOVA revealed a significant main effect of the experimental condition on Cart Diversity (F(2, 509) = 19.82, p < 0.001). Post-hoc tests using Tukey's HSD confirmed the direction of this effect. The High-Bias group had the lowest Cart Diversity (M = 1.65, SD = 0.88), which was significantly lower than both the Low-Bias group (M = 2.45, SD = 1.12; p < 0.001) and the Control group (M = 2.20, SD = 1.05; p < 0.001). Interestingly, the Cart Diversity of the Low-Bias group was slightly, but not significantly, higher than the Control group (p = 0.15), suggesting that a well-designed, diversity-aware recommendation system can encourage exploration beyond a user's un-aided browsing.<figure class="article-figure"><img src="https://smnxsewcdnayrztrrghn.supabase.co/storage/v1/object/public/journal-assets/scholarly/echo-chambers-of-commerce-quantifying-the-impact-of-algorithmic-bias-in-recommendation-systems-on-co-xpzot/figure-2-1778088856264.png" alt="bar chart comparing mean purchase diversity scores across High-Bias, Low-Bias, and Control experimental conditions" loading="lazy" style="max-width:100%;height:auto;"><figcaption>Figure 2. bar chart comparing mean purchase diversity scores across High-Bias, Low-Bias, and Control experimental conditions</figcaption></figure>The self-reported measures, summarized in Table 3, further reinforce these findings. There was a significant effect of the condition on Perceived Choice Satisfaction (F(2, 509) = 8.54, p < 0.001). Participants in the High-Bias condition reported significantly lower satisfaction with their choices (M = 4.88) compared to those in the Low-Bias (M = 5.51) and Control (M = 5.45) conditions. This suggests that even if users follow biased recommendations, they feel less confident and happy with their resulting selections.Comparing only the two treatment groups, we also found significant differences in the evaluation of the recommendations themselves. Participants in the Low-Bias condition rated the recommendation quality significantly higher than those in the High-Bias condition (M = 5.12 vs. M = 4.15; t(340) = 6.21, p < 0.001). Crucially, this translated into higher trust; the Low-Bias group reported significantly greater Trust in the Platform for future recommendations (M = 5.05 vs. M = 4.30; t(340) = 5.01, p < 0.001). This finding causally demonstrates how a biased recommendation strategy can directly erode consumer trust.<figure class="table-figure"><table><thead><tr><th>Measure</th><th>High-Bias (n=170)</th><th>Low-Bias (n=172)</th><th>Control (n=170)</th><th>F-statistic</th><th>p-value</th></tr></thead><tbody><tr><td>Cart Diversity</td><td>1.65 (0.88)</td><td>2.45 (1.12)</td><td>2.20 (1.05)</td><td>19.82</td><td> < .001</td></tr><tr><td>Perceived Choice Satisfaction</td><td>4.88 (1.21)</td><td>5.51 (1.09)</td><td>5.45 (1.15)</td><td>8.54</td><td> < .001</td></tr><tr><td>Perceived Recommendation Quality</td><td>4.15 (1.30)</td><td>5.12 (1.18)</td><td>N/A</td><td>-</td><td>-</td></tr><tr><td>Trust in Platform</td><td>4.30 (1.25)</td><td>5.05 (1.20)</td><td>N/A</td><td>-</td><td>-</td></tr></tbody></table><figcaption>Table 3. Comparison of Mean Scores for Key Variables Across Experimental Conditions (Study 2). Standard deviations in parentheses. N/A for Control group on recommendation-specific questions.</figcaption></figure>

<h2>Discussion</h2> This research set out to investigate whether algorithmic bias in e-commerce recommendation systems has a tangible impact on consumer purchasing behavior. By combining a large-scale observational study with a controlled experiment, we find a consistent and robust answer: it does. The findings from both studies converge to show that exposure to biased, low-diversity recommendations systematically narrows the scope of consumer purchases and can negatively affect customer satisfaction and trust.<h3>5.1 Summary and Interpretation of Findings</h3>Our first study, using 18 months of real-world data, demonstrated a clear negative relationship between exposure to concentrated recommendations (our REB metric) and the diversity of products a consumer buys. As users were shown a less varied slate of recommendations, they subsequently purchased from fewer product categories. This finding provides ecological validity to the long-held concern that recommendation algorithms risk creating commercial echo chambers. The effect was not trivial; a one standard deviation increase in recommendation bias was associated with a 12% drop in purchase diversity. Furthermore, we found that this bias reduces purchase serendipity, directly inhibiting product discovery—one of the supposed benefits of e-commerce.Our second study established the causal nature of this relationship. By experimentally manipulating the diversity of recommendations in a simulated shopping environment, we demonstrated that participants exposed to high-bias (popularity-driven) recommendations constructed significantly less diverse shopping carts than those exposed to low-bias recommendations or no recommendations at all. This is a critical finding, as it isolates the algorithm's effect from user predispositions. The algorithm is not just reflecting preference; it is actively shaping it. The experimental results further revealed the psychological cost of this narrowing effect: consumers in the high-bias condition were less satisfied with their choices and, critically, reported lower trust in the platform's ability to make good suggestions in the future. This suggests a perilous trade-off for firms: a strategy that maximizes clicks on popular items in the short term may erode the long-term relational asset of customer trust.<h3>5.2 Theoretical Implications</h3>These findings make several contributions to theory. First, they extend the literature on consumer behavior in digital environments (Tokarski & Fajczak-Kowalska, 2024; Halim, 2022). We empirically demonstrate that the algorithmic curation of the choice environment is a powerful, and perhaps underestimated, factor influencing purchasing patterns, on par with other studied factors like social influence (Poh et al., 2024) or live streaming (Wang & Musa, 2024). We provide a concrete mechanism through which the 'filter bubble' phenomenon (Dwivedi et al., 2020) transitions from the world of information to the world of commerce.Second, our research adds a crucial dimension to the Stimulus-Organism-Response (S-O-R) model as applied to e-commerce (Kim et al., 2018). We show that the nature of the algorithmic 'stimulus' (high-bias vs. low-bias recommendations) directly impacts not only the final 'response' (purchase) but also the consumer's cognitive and affective 'organism' state (satisfaction, trust). The negative impact on satisfaction suggests a mismatch between the algorithm's objective function (e.g., predicting clicks) and the consumer's holistic goal of a gratifying shopping experience.Third, we contribute to the emerging field of consumer experience with AI (Puntoni et al., 2020). Our work moves beyond conceptual arguments to provide quantitative evidence of a key tension: the friction between AI-driven personalization and the human desire for agency, discovery, and exploration. The finding that the 'Low-Bias' group did not report significantly lower recommendation quality than the high-bias one, but had higher satisfaction, suggests that consumers do not necessarily want perfect prediction of their existing tastes. Instead, they may value a system that gently pushes them to explore, acting as a discovery partner rather than a restrictive gatekeeper. This reframes the goal of personalization away from simple preference matching and towards preference expansion.<h3>5.3 Managerial Implications</h3>The implications of our findings for e-commerce managers and marketing strategists are significant and actionable. 1. Measure and Monitor Recommendation Diversity. The most immediate takeaway is that if you don't measure it, you can't manage it. Firms should move beyond simplistic metrics like click-through rate (CTR) and conversion rate for their recommendation modules. They should develop and track metrics that capture the diversity and novelty of the items being recommended and purchased, such as the REB and Purchase Diversity metrics used in this study. These metrics should be part of the standard dashboard for evaluating RS performance.2. Rethink the Objective Function. Our results serve as a warning against myopically optimizing for short-term engagement on popular items. This strategy, while seemingly efficient, may lead to a long-term decline in customer engagement and lifetime value by trapping customers in taste silos and reducing trust. Firms should consider incorporating diversity and serendipity directly into the algorithmic objective function. The goal should be a balanced portfolio of recommendations that includes predictable hits, surprising discoveries, and relevant niche products.3. Embrace User Control and Transparency. One way to mitigate the negative effects of bias is to empower the user. Platforms could experiment with features that allow users to control the diversity of their recommendations—for example, a 'slider' that moves between 'Show me what I usually like' and 'Surprise me with something new'. Being transparent about why an item is being recommended ('Because you bought X' vs. 'Customers like you also discovered Y') can also help manage expectations and build trust.4. Strategic Value of the Long Tail. By over-promoting popular products, platforms inadvertently harm the vendors of 'long-tail' or niche products. A healthier, more diverse marketplace ecosystem is likely more resilient and profitable in the long run. Actively promoting diversity in recommendations is not just good for the consumer; it is good for the platform's relationship with its suppliers and for the overall health of its product catalogue.<h3>5.4 Limitations and Future Research</h3>No study is without limitations. While our mixed-methods approach provides strength through triangulation, each component has its caveats. The observational data in Study 1, despite fixed-effects controls, is ultimately correlational, and unobserved time-varying confounders could exist. Study 2, while establishing causality, was conducted in a simulated environment with a hypothetical budget, which may not perfectly capture the complex motivations and constraints of real purchasing decisions.Our operationalization of bias was focused on popularity and concentration. Future research should explore other facets of algorithmic bias, such as demographic, gender, or price-point biases, and their impact on consumer behavior. For example, does the algorithm disproportionately recommend higher-margin products, and how do consumers react to that? (Avery et al., 2015). Longitudinal experiments that track users over weeks or months would be a powerful next step to observe how the feedback loops we identified play out over time. Does sustained exposure to biased recommendations permanently alter a consumer's taste profile? Can interventions promoting diversity reverse this effect? Finally, future work could explore the heterogeneity of these effects. Are some consumers more susceptible to algorithmic influence than others? Factors like shopping expertise, domain knowledge, or personality traits could moderate the relationships we found, offering opportunities for even more nuanced personalization.

<h2>Conclusion</h2> This study provides robust, multi-method evidence that algorithmic bias in e-commerce recommendation systems is a powerful force that actively shapes consumer purchasing behavior. We demonstrated through both observational data and a controlled experiment that a reliance on popularity-biased algorithms, while perhaps efficient in the short-term, systematically reduces the diversity of products consumers purchase. This creation of commercial 'echo chambers' not only limits consumer discovery but also correlates with lower satisfaction and eroded trust.The findings challenge a simplistic view of recommendation systems as passive servants that merely fulfill pre-existing preferences. Instead, they operate in a complex feedback loop with the consumer, both predicting and shaping behavior. The managerial imperative is clear: firms must evolve their strategies from a narrow focus on conversion prediction to a more holistic goal of fostering a healthy, diverse, and satisfying customer experience. This involves measuring and managing recommendation diversity, rethinking algorithmic objectives, and empowering users with more control and transparency.As artificial intelligence becomes more deeply embedded in the fabric of commerce and daily life (Puntoni et al., 2020), understanding its behavioral consequences is paramount. Our research underscores the need for a responsible, human-centric approach to designing marketing AI (Huang & Rust, 2020). The ultimate goal should not be to build a perfect cage of personalized predictions, but to provide a window into a wider world of possibilities, enriching the consumer journey rather than constraining it. For e-commerce platforms, fostering discovery may prove to be a more sustainable path to long-term profitability than simply reinforcing the popular.

<h2>References</h2> <ol class="references"> <li>Liu, L. (2024). Algorithmic Bias in Recommendation Systems and Its Social Impact on User Behavior. International Theory and Practice in Humanities and Social Sciences, 1(1), 290. https://doi.org/10.70693/itphss.v1i1.204</li> <li>Ding, N., Yun, K., Chen, M. (2024). Study on the Impact of Packaging Design Elements and Perceived Value of Sugar-Free Tea Beverages on Consumer Purchasing Behavior. Korean Academy Of International Commerce, 39(2), 229-254. https://doi.org/10.18104/kalc.2024.39.2.229</li> <li>Poh, S., Hasan, D. G., Sudiyono, K. A. (2024). The power of social commerce: TikTok's impact on Gen Z consumer purchasing behavior. Manajemen dan Bisnis, 23(2), 501. https://doi.org/10.24123/mabis.v23i2.835</li> <li>Wang, Q., Binti Musa, R. (2024). The Impact of E-commerce Live Streaming on Consumer Purchasing Behavior in the Influencer Economy. Frontiers in Business, Economics and Management, 16(3), 102-105. https://doi.org/10.54097/nadkz803</li> <li>Tokarski, D., Fajczak-Kowalska, A. (2024). Optimization of consumer decisions and the impact of selected factors on purchasing behavior in polish e-commerce. Economics and Environment, 88(1), 728. https://doi.org/10.34659/eis.2024.88.1.728</li> <li>Nikitkov, A. (2006). Information Assurance Seals: How They Impact Consumer Purchasing Behavior. Journal of Information Systems, 20(1), 1-17. https://doi.org/10.2308/jis.2006.20.1.1</li> <li>Halim, M. A. (2022). The Impact of E-commerce on Consumer Purchasing Behavior for the Coronavirus Disease (COVID-19). Journal of Sustainable Business and Economics, 5(1). https://doi.org/10.30564/jsbe.v5i1.4283</li> <li>Zunying, X. (2024). Analysis of Intelligent Recommendation Systems and Consumer Behavior Theories on E-Commerce Platforms. Philosophy and Social Science, 1(6), 10-15. https://doi.org/10.62381/p243602</li> <li>Avery, D. R., McKay, P. F., Volpone, S. D., Malka, A. (2015). Are companies beholden to bias? The impact of leader race on consumer purchasing behavior. Organizational Behavior and Human Decision Processes, 127, 85-102. https://doi.org/10.1016/j.obhdp.2015.01.004</li> <li>Halim, M. A. (2022). The Impact of E-commerce on Consumer Purchasing Behavior for the Coronavirus Disease (COVID-19). Journal of Sustainable Business and Economics, 5(1), 20-28. https://doi.org/10.30564/jsbe.v5i1.3</li> <li>Gu, S., Ślusarczyk, B., Hajizada, S., Kovalyova, I., Sakhbieva, A. (2021). Impact of the COVID-19 Pandemic on Online Consumer Purchasing Behavior. Journal of Theoretical and Applied Electronic Commerce Research, 16(6), 2263-2281. https://doi.org/10.3390/jtaer16060125</li> <li>Alshweesh, R., Bandi, D. S. (2022). The Impact of E-Commerce on Consumer Purchasing Behavior: The Mediating Role of Financial Technology. International Journal of Research and Review, 9(2), 479-499. https://doi.org/10.52403/ijrr.20220261</li> <li>Li, H., Pan, Y. (2023). Impact of Interaction Effects between Visual and Auditory Signs on Consumer Purchasing Behavior Based on the AISAS Model. Journal of Theoretical and Applied Electronic Commerce Research, 18(3), 1548-1559. https://doi.org/10.3390/jtaer18030078</li> <li>Racat, M., Plotkina, D. (2023). Sensory-enabling Technology in M-commerce: The Effect of Haptic Stimulation on Consumer Purchasing Behavior. International Journal of Electronic Commerce, 27(3), 354-384. https://doi.org/10.1080/10864415.2023.2226900</li> <li>V, D. M., KM, R. (2024). Examining the impact of green branding on consumer behavior: A study on awareness, purchasing patterns, and gender-specific awareness in the FMCG sector in Bengaluru. Asian Journal of Management and Commerce, 5(1), 389-393. https://doi.org/10.22271/27084515.2024.v5.i1e.284</li> <li>Mirabi, V., Fathi, F., Fotouhi-Ardakani, M., Avorgani, R. K. (2021). The Impact of Celebrity Endorsement of Sporting Goods on Consumer Purchasing Behavior. CBR - Consumer Behavior Review, 5(3), 373. https://doi.org/10.51359/2526-7884.2021.251296</li> <li>SOMEYA, H., OTSUKA, T., MITOMO, H. (2007). Impact of E-Commerce Diffusion on Consumer Purchasing Behavior : An Empirical Analysis of Substitutability of ICT for Physical Travel for Book Purchasing. Studies in Regional Science, 37(4), 1157-1172. https://doi.org/10.2457/srs.37.1157</li> <li>Syahiman Ghazalle, M., Abdul Lasi, M. (2021). DETERMINANT SUCCESS FACTORS ON CUSTOMER PURCHASING BEHAVIOR TOWARDS CONSUMER PURCHASING INTENTION: A STUDY ON STUDENT PERSPECTIVE IN PUBLIC INSTITUTIONS. MALAYSIAN E COMMERCE JOURNAL, 5(1), 36-41. https://doi.org/10.26480/mecj.01.2021.36.41</li> <li>Kinawy, R. N. (2024). Unraveling consumer behavior: Exploring the influence of consumer ethnocentrism, domestic country bias, brand trust, and purchasing intentions. Strategic Change, 34(2), 137-150. https://doi.org/10.1002/jsc.2607</li> <li>Білоус, А. (2023). THE IMPACT OF SOCIAL MEDIA ON CONSUMER BEHAVIOR AND PURCHASING DECISIONS. SWorldJournal(35-03), 333-345. https://doi.org/10.30888/2663-5712.2026-35-03-121</li> <li>Olbrich, R., Holsing, C. (2011). Modeling Consumer Purchasing Behavior in Social Shopping Communities with Clickstream Data. International Journal of Electronic Commerce, 16(2), 15-40. https://doi.org/10.2753/jec1086-4415160202</li> <li>Dwivedi, Y. K., Ismagilova, E., Hughes, D. L., Carlson, J., Filieri, R., Jacobson, J. (2020). Setting the future of digital and social media marketing research: Perspectives and research propositions. International Journal of Information Management, 59, 102168-102168. https://doi.org/10.1016/j.ijinfomgt.2020.102168</li> <li>Leskovec, J., Adamic, L. A., Huberman, B. A. (2007). The dynamics of viral marketing. ACM Transactions on the Web, 1(1), 5-5. https://doi.org/10.1145/1232722.1232727</li> <li>Goldfarb, A., Tucker, C. E. (2019). Digital Economics. Journal of Economic Literature, 57(1), 3-43. https://doi.org/10.1257/jel.20171452</li> <li>Shmueli, G. (2010). To Explain or to Predict?. Statistical Science, 25(3). https://doi.org/10.1214/10-sts330</li> <li>Huang, M., Rust, R. T. (2020). A strategic framework for artificial intelligence in marketing. Journal of the Academy of Marketing Science, 49(1), 30-50. https://doi.org/10.1007/s11747-020-00749-9</li> <li>Kapoor, K. K., Tamilmani, K., Rana, N. P., Patil, P. P., Dwivedi, Y. K., Nerur, S. (2017). Advances in Social Media Research: Past, Present and Future. Information Systems Frontiers, 20(3), 531-558. https://doi.org/10.1007/s10796-017-9810-y</li> <li>Kim, M. J., Lee, C., Jung, T. (2018). Exploring Consumer Behavior in Virtual Reality Tourism Using an Extended Stimulus-Organism-Response Model. Journal of Travel Research, 59(1), 69-89. https://doi.org/10.1177/0047287518818915</li> <li>Lü, L., Medo, M., Yeung, C. H., Zhang, Y., Zhang, Z., Zhou, T. (2012). Recommender systems. Physics Reports, 519(1), 1-49. https://doi.org/10.1016/j.physrep.2012.02.006</li> <li>Puntoni, S., Reczek, R. W., Giesler, M., Botti, S. (2020). Consumers and Artificial Intelligence: An Experiential Perspective. Journal of Marketing, 85(1), 131-151. https://doi.org/10.1177/0022242920953847</li> </ol> </article>

Published by Academic Ink Review Journal. Open Access under CC BY 4.0.