Auditing the Ranking Strategy of a Marketplace 's Algorithm in the Frame of Competition Law Commitments with Surrogate Models: The Amazon 's Buy Box Case

In a global context where competition authorities are investigating and sanctioning Amazon's marketplace for practices of self-preferencing at the expense of their business users and consumers (Italian AGCM 2021, EU Commission 2022, UK CMA 2024, US FTC on-going since 2023), we observe a trend of imposing remedies on dominant players in digital markets. In addition, the Digital Market Act, shifting from an ex-post enforcement approach to ex-ante obligations on designated gatekeepers, is strengthening auditing power over these gatekeepers, which risk heavier penalties in the event of non-compliance. Therefore, competition authorities and regulators need tools to audit the compliance of these dominant players in the e-commerce sector over the obligations and remedies they are imposing on dynamic, and personalized algorithms. Most of these algorithms embed Machine-Learning components, introducing opacity and potentially biases in the decision-making process. The aim of the paper is to explore the benefits of using black-box auditing techniques to provide insights into the behavior of these online algorithms. We anchor our research in the literature of product prominence from vertically integrated players, of choice ranking, and of the specific literature related to Amazon search ranking, automatic pricing and Buy Box 's algorithms. Through a study of the pricing and ranking of several thousand products on Amazon, from 2017 to 2023, we illustrate the potential of surrogate models. While our dataset only covers some categories on Amazon.fr, the large number of competitions allowed us to demonstrate, with a 94% accuracy, that the variable is Amazon, or variables correlated to it, had a positive effect on winning Buy Box before mid-2022, and that this positive effect has decreased after mid-2022. In our research, the machine learnings models revealed a significantly higher degree of accuracy and sensitivity compared to a logistic regression, opening the discussion on the added value and role of surrogate models based on machine learning techniques in guiding the auditor, as well as raising the question of their probative value in the regulatory context.