The Effects of Recommendation Algorithms' Specialization on Consumers' Outcomes

While shopping online, consumers may often rely on the support of recommendation agents (hereinafter, RAs) embedded in e-commerce platforms and designed to reduce search costs, increase the chance to find products and services that match consumers’ needs and preferences (Xie et al., 2017) and thus reducing the risk of making an unsatisfying decision in relatively large, assorted platforms (Henning-Thurau et al., 2012). Typical instances of how recommender agents work are statements like “you may also like...” or "People who like this also like” that online buyers often encounter after having made a purchase. Companies such as Spotify, Netflix, Amazon, TripAdvisor and eBay are heavily investing in such systems for an estimated global spending of $5.9 billion in 2019 (Worldwide Spending, 2019). It is estimated that about 92% of online searches originate from a recommender agent and a vast majority of individuals are exposed to RAs on a daily basis as sources of advice about products and services on the Internet (Gai & Klesse, 2019). With technological developments that have nowadays made RAs highly accurate from a computational standpoint (Gai & Klesse, 2019), RAs commonly work as substitutes of gatekeepers, experts, and possibly even decision makers as key actors involved in the purchase decision process (Lee & Choi, 2017), due to their ability to convey promotional messages that are tailored to a specific consumer’s interests. RAs are particularly useful when individuals do not have a preferred option, delegate the shopping process to someone else and are unaware of the entire assortment (Lee & Choi, 2017). A plethora of studies have investigated the effects of RAs, demonstrating that these agents reduce search costs supporting consumers in the identification of items with a high fit with users’ interest in a vast collection of products, influence users’ consideration set by diminishing the number of product considered (Häubl & Trifts, 2000), increase the decision quality and reduce decision efforts in large assorted platforms (Häubl & Trifts, 2000; Panniello et al., 2014), influence user’s opinions and choices (Senecal et al., 2005; Senecal & Nantel, 2004). Past work also showed that the impact of RAs varies according to the methods used to recommend (Dzyabura & Hauser, 2019), the product categories (Senecal & Nantel, 2004) and the form of explanation involved to suggest the product (Schumann et al., 2014). Overall, such previous studies have focused on the functional effects of RAs on consumer-decision making processes. However, while designed to help consumers in their purchase decision processes, RAs ultimately lead to a reinforcement of people’s existing preferences as they typically recommend consumers to buy products that are in line with their inferred preferences, thus limiting the possibility that a consumers might sometimes consider options that do not match their interests (Matz, 2021; Shen & Ball, 2011). This reinforcement exerted by RAs on users’ preferences is called overspecialization problem and originates from the everlasting improvement of RAs accuracy. It refers to a situation where an algorithm becomes too specific or tailored to particular users, preventing consumers from discovering new products that RAs consider as far from users’ preferences (Kim et al., 2021; Banker & Khetani, 2019). When an algorithm is overly specialized, it may be associated to the implicit risk of selecting alternatives that are lowly satisfying as they are chosen only because they fit with users’ preferences (Banker & Khetani, 2019). Scant evidence, however, seems to have been produced on the underlying effects triggered by exposure and use of specialized or overly specialized RAs, as few studies have investigated whether, when and why RAs might affect consumers’ choice outcomes (e.g., Banker & Khetani, 2019). In our investigation, we move from the idea that higher degrees of RAs knowledge confine users within their preferences and could predict a decrease in terms of choice outcomes. In an between-subject experimental design and a sample of 120 participants (60 women; Mage= 27 years, SD= 3.24) online users, we investigate the effects of overspecialization on users’ choice outcomes – such as the perceived usefulness and benefit of the recommendation set, the willingness to accept the recommendation, users’ satisfaction, and enjoyment with the product selection, as well as their antecedents and a potential solution to the issue. The study investigates how increasing levels of RAs’ specialization reduce the users’ outcomes. - such as perceived usefulness and benefit of the recommendation set, the willingness to accept the recommendation, users’ satisfaction and enjoyment. We have reproduced a Recommendation Algorithm able to learn from users’ preferences and adjust the recommendation set accordingly. The programmed RA had the ability to reach three different degrees of learning (generalist vs. specialized vs. overspecialized). In accordance with our prediction, the experiment proved that over-specialized RAs with high levels of knowledge about users lead to lower choice outcomes, meaning that users prefer even less accurate RAs that are not a mere reflection of their preferences. Importantly, accurate RAs are seen as a way to bound users’ preferences to existing ones and limit the chances to form new preferences. The findings contribute to the advice-taking literature proving that a decrease of the levels of recommendation accuracy results in more positive users’ outcomes. It implies that higher levels of accuracy (i.e., overspecialization) undermine the evolution of users’ preferences and that RAs should be able to create tailored experiences while offering the chance to encounter products far from extant preferences. This study provides contrasting evidence for theories that support the precision of RSs and shows that generalized information elicit people to think that the RSs is associated with a higher outcome. From a managerial perspective, this research highlights that RAs should leverage on a combination between previously rated and novel products rather than only on rated products.