The autonomous operation and adaptability of artificial intelligence (AI)-driven pricing algorithms have made them an increasingly valuable tool for firms seeking to optimise pricing strategies in fluid and competitive markets. These systems can dynamically adjust prices in response to real-time market signals, including demand fluctuations, consumer behaviour, and rivals’ pricing strategies. Their promise lies in enhancing efficiency and revenue optimisation. However, their growing prevalence has raised concerns among scholars and regulators alike. A central issue is that specific pricing algorithms have demonstrated the capacity to learn tacitly collusive behaviours—coordinating pricing in ways that suppress competition without explicit agreement. This can result in inflated prices that ultimately harm consumer welfare, prompting calls for stricter oversight and more thoughtful algorithmic design to ensure competitive outcomes.
A recent study published in Marketing Science by researchers at Carnegie Mellon University investigates how the structure of product ranking systems on e-commerce platforms influences the pricing behaviours of AI algorithms. While using personalised product rankings—those tailored to individual consumer profiles—is generally perceived as enhancing the shopping experience by reducing search time and improving product relevance, the researchers raise a compelling question: could such personalisation inadvertently enable firms to charge higher prices, thereby reducing consumer welfare? Notably, the study focuses not on traditional price discrimination but on whether personalisation in the ranking can distort market dynamics, even when identical prices are shown to all.
Param Vir Singh, Carnegie Bosch Professor of Business Technologies and Marketing at the Tepper School of Business, explains that the team compared two extreme scenarios in product ranking design. The first involved personalised rankings, where algorithms use detailed consumer data to predict and prioritise products according to expected utility for each individual. The second employed unpersonalised rankings, where products are ordered based on aggregate preferences without tailoring to any specific user. These systems are standard in digital marketplaces such as Amazon and Expedia, which serve as search intermediaries, helping consumers navigate a growing sea of third-party listings. By focusing on these two ranking types, the researchers could isolate the effects of personalisation on pricing outcomes.
Central to the study was a consumer search model in which users examine product listings sequentially, incurring a small cost with each viewed item. Consumers are assumed to behave optimally—searching until the expected utility gain no longer outweighs the cost of continuing. The ranking system, therefore, plays a pivotal role in determining the order in which products are considered. In this framework, the researchers explored how reinforcement learning (RL) algorithms, often employed for pricing decisions, adapt to these ranking conditions. The assumption is that if a ranking system consistently pushes high-utility (and potentially high-priced) products to the top, pricing algorithms will learn they can raise prices without significantly dampening demand.
Indeed, the study found that personalised ranking systems tended to diminish the price sensitivity of consumer demand. When products most aligned with an individual’s preferences appear at the top of a list, the consumer is likelier to purchase them without continuing the search. This reduces the pressure on firms to maintain competitive pricing. As a result, AI pricing algorithms operating in such an environment learn that they can charge higher prices while still achieving strong sales performance. The reduced price elasticity leads to a general upward shift in pricing, even though no explicit collusion or discriminatory pricing occurs. Conversely, unpersonalised ranking systems, which do not cater specifically to individual preferences, maintain higher search incentives and encourage broader price comparison, leading to lower overall prices and greater consumer welfare.
Their consistency across multiple experimental conditions strengthens the credibility of these findings. The researchers tested various reinforcement learning algorithm types, adjusted learning parameters, included different valuations of outside options, and simulated scenarios involving several competing firms. Across all configurations, the core outcome remained the same: personalised rankings enabled higher prices and reduced consumer welfare, while unpersonalised rankings resulted in more competitive pricing. Liying Qiu, a doctoral student who led the study, highlighted the challenge of modelling these interactions due to the complexity of dynamic learning behaviours. Nevertheless, by constructing a controlled and replicable simulation environment, the team could empirically observe how AI pricing algorithms evolve in response to different ranking inputs.
These findings have significant implications for policymakers, platform designers, and regulators. The study underscores that ranking systems, which may appear neutral or beneficial at first glance, play an active role in shaping market outcomes. Personalisation, while helpful in reducing consumer search costs, can be weaponised by algorithms optimising for profit rather than consumer welfare. Focusing solely on price transparency or algorithmic fairness in isolation is not enough. Regulators must also consider how platform design choices—particularly product visibility and ranking—interact with pricing algorithms to affect competitive dynamics. The study suggests that limiting personalisation, or at least making its influence more transparent, may be necessary to safeguard consumer interests.
Finally, the research prompts a re-examination of the widespread belief that greater data sharing by consumers leads to improved market efficiency. While more data can improve product matching, it allows firms to tailor experiences in ways that ultimately erode consumer surplus subtly. Even without overt price discrimination, the information asymmetry introduced by personalisation can empower algorithms to manipulate demand patterns. As Professor Kannan Srinivasan, another co-author of the study, points out, the value of personalisation must be weighed carefully against its broader systemic effects. This study provides a cautionary roadmap for aligning technological advancement with public interest and competitive fairness for digital marketplaces increasingly reliant on AI and personal data.
More information: Liying Qiu et al, Personalization, Consumer Search, and Algorithmic Pricing, Marketing Science. DOI: 10.1287/mksc.2023.0455
Journal information: Marketing Science Provided by Carnegie Mellon University
