Academic papers

Heterogeneous Effects of Generative AI on Knowledge Seeking in Online Communities (with D. Gutt).
Journal of  Management Information Systems, 2025.
Last SSRN version | Published version

Abstract - How do generative artificial intelligences (GAIs) affect their own training data such as that from Q\&A platforms? The Paradox of Reuse Theory posits that GAI usage would substitute traffic to Q\&A platforms where the training data originates. However, it remains an open question whether and how this affects subsequent training data generation on these platforms. We address this question by leveraging the launch of ChatGPT and using rich StackExchange panel data. We find a decrease in the number of questions on the platform, driven by platform abandonment of casual users, but the remaining questions are the more complex ones and exhibit novelty. Further investigation illustrates that the more the number of questions decreases, the more the remaining questions increase in complexity and novelty. These findings support the development of the Paradox of Reuse Theory by illuminating a special case in which training data erosion may be beneficial to the GAI.

Content Moderation and Advertising in Social Media Platforms (with L. Madio).
Journal of  Economics and Management Strategy, 2024.
Last SSRN version  | Published version

Abstract - We study the incentive of an ad-funded social media platform to curb the presence of unsafe content that entails reputational risk to advertisers. We identify conditions for the platform not to moderate unsafe content and demonstrate how the optimal moderation policy depends on the risk the advertisers face. The platform is likely to under-moderate unsafe content relative to the socially desirable level when both advertisers and users have congruent preferences for unsafe content and to over- moderate unsafe content when advertisers have conflicting preferences for unsafe content. Finally, to mitigate negative externalities generated by unsafe content, we study the implications of a policy that mandates binding content moderation to online platforms and how the introduction of taxes on social media activity and social media platform competition can distort the platform’s moderation strategies.

The Welfare Effects of Mobile Internet Access : Evidence from Roam-Like-At-Home (with M. Godinho de Matos and C. Peukert).
The Economic Journal, 2023.
Last CESifo working paper version | Published version

Abstract - We evaluate the welfare effects of the Roam-Like-At-Home regulation, which drastically reduced the price of accessing the mobile internet for EU residents when traveling abroad in the European Economic Area. Estimates from individual-level usage data suggest that consumer surplus increased by 2.77 EUR/user/travel day. A decomposition shows the heterogeneous impact of the regulation on different user segments. We estimate that around half of the gains stem from a reduction in deadweight loss, i.e., new users accessing the mobile internet. We further show that the impact of the regulation varies with usage intensity abroad and at home, by the nature of the trip (leisure vs. business), and by content type. We discuss implications for content providers and other policy areas such as net neutrality.

Payment Instruments, Financial Privacy and Online Purchases (with Y. Balgobin, D. Bounie and P. Waelbroeck).
Review of Network Economics, 2017.
Last SSRN version | Published version

Abstract - The protection of financial personal data has become a major concern for Internet users in the digital economy. This paper investigates whether the consumers’ use of non-bank payment instruments that preserve financial privacy from banks and relatives may increase their online purchases. We analyze the purchasing decisions and the use of bank and non-bank payment instruments of a representative sample of French Internet consumers in 2015. Using two econometric methods, namely a two-step regression and a Bayesian Markov Chain Monte Carlo model to account for a potential endogeneity problem, we find evidence that the use of a non-bank payment instrument positively influences consumers’ online purchases. 

Hit the GAS: Designing Optimal Generalized Ad-supported Subscription Mechanisms  (D. Gutt, and S. Mehta).
R&R at Information System Research.
Last SSRN version 

Abstract - Digital Content Platforms (DCPs), such as Netflix and Spotify, rely on subscriptions and advertising as their primary revenue sources. Traditional revenue models, such as subscription-only and ad-only options, either limit user accessibility or compromise user experience. In response, DCPs have started to combine these two models by offering a two-tier menu: a free ad-supported tier for price-sensitive users and a paid ad-free tier for ad-sensitive users. In this paper, we introduce the Generalized Ad-Supported Subscription (GAS) mechanism- a broad class consisting of subscription-fee and ad-intensity combinations that spans the entire spectrum from fully ad-supported to fully subscription-based access. Using the principles of mechanism design, we characterize a revenue-maximizing GAS mechanism and contrast it with ad-only, subscription-only, and the two-tier menu. Further, we estimate our model parameters and empirically validate our theoretical results in the context of Video-on-Demand platforms. Our analysis shows that the optimal GAS mechanism yields a substantially higher revenue than the ad-only and subscription-only mechanisms. More importantly, this additional revenue improvement does not come at the expense of consumer surplus- this shows that the GAS mechanism improves the overall economic pie by better matching user preferences.

Asymmetric content moderation in search markets: The case of adult websites.  (with L. Madio, M. Mitchell, and C. Reggiani).
Under Review at Journal of Political Economy.
Last SSRN version

Abstract - We study the competitive impact of content moderation by a leading online platform. We exploit an exogenous shock that led the largest online platform in the adult industry to remove 80% of its video library. Using a difference-in-differences approach, we find that the content moderation of the platform led to a 37-51% drop in visits in one month, indicating strong user preferences for removed content. However, the traffic diverted was largely captured by other firms, including unregulated and often illegal fringe websites as well as mainstream competitors, which increased their traffic by 1.7% per week. Ultimately, the fringe sites saw an increase of 68% visits over 6 months. Search engines facilitated this migration - fringe websites experienced a surge in traffic from search channels, including aggregators, as users sought alternative content sources. The results highlight how asymmetric regulation can reshape competition dynamics by pushing users towards unregulated spaces, and how consumers substitute across platforms.

The Returns to Targeting: Evidence from a Field Experiment (with P. Ferreira, M. Godinho de Matos & C. Peukert).
Submitting.

Abstract - We examine the effectiveness of data-driven targeting strategies for product recommendation by focusing on two key data phases: the training data used to build recommendation models and the inference data available at the moment of recommendation. Specifically, we compare techniques that require training data comprised of personal historical purchase information with those relying solely on non-personal data, namely detailed product characteristics. In collaboration with a video-on-demand platform, we conduct a field experiment involving 380,000 users. Participants are randomly assigned to receive targeted recommendations from models trained either on personal or non-personal data, while a control group receives random, non-targeted recommendations. Our findings show that, overall, targeted recommendations outperform random recommendations in driving user engagement. Notably, however, there is no significant difference in the effectiveness of techniques using personal versus non-personal training data. Instead, recommendation performance depends critically on the availability of inference data. Users about whom the firm has more real-time information engage more with recommendations, suggesting that personal inference data is more important for driving engagement than the type of training data used.

Regulatory Content Moderation: Evidence from Turkey’s Ban of Twitch and Kick (with Dominik Gutt).
Work in progress.

Abstract - This paper examines how sudden government‐imposed content moderation shapes content creators’ behavior and platform dynamics in live streaming markets. For our empirical investigation, we leverage the bans of the popular live streaming platforms Twitch and Kick in Turkey in February 2024. We use daily data on 2,213 top European and Turkish streamers over 42 days to implement a difference‐in‐differences design to isolate the impacts of the ban on streamers' airtime, participation, viewership (hours watched), and content diversification. We find that during the Twitch ban, Turkish streamers’ airtime fell by 68%, participation by 34%, and hours watched by 89%, with only partial rebound in participation and viewership after reinstatement. The Kick ban triggered similar immediate declines (78% airtime, 20% participation, and 98% viewership) but was followed by a substantial post‐ban surge, suggesting permanent platform migration, possibly because Twitch was still banned when the ban on Kick was lifted. Content‐diversification metrics reveal a 6–8 % increase in the number of languages used on both platforms during bans, while game‐portfolio expansion occurs only on Kick. These results illustrate streamers’ rapid, strategic adjustments—multilingual outreach and multi‐platform engagement—in response to regulatory disruptions, highlighting resilience mechanisms in two‐sided digital ecosystems.

Others

Advertising Viewability in Online Branding Campaigns (with D. Bounie and V. Morrisson)
Digiworld Economic Journal, No. 104 - 4th quarter 2016.
Last SSRN version 

Abstract - A significant number of online ads served by publishers are actually never seen by Internet users. This results in ineffective branding campaigns and a considerable waste of money for advertisers. In reaction, more and more advertisers use technologies to measure the viewability of advertising campaigns on publisher websites. This paper discusses how the adoption of such technologies could impact the economics of online advertising. 

Do You See What I See? Ad Viewability and the Economics of Online Advertising  (with D. Bounie and V. Morisson).
SSRN

Abstract - Between 40% and 50% of online ads served by publishers are actually never seen by Internet users, resulting in ineffective branding campaigns and a considerable waste of money for advertisers. In reaction, more and more advertisers use technologies to measure the viewability of advertising campaigns on publisher websites. This paper provides the first comprehensive economic analysis of the impact of the adoption of such technologies on the economics of online advertising. We construct a two-sided market model for advertising where publishers manage their website to attract Internet users and advertisers. We show that the adoption of ad viewability technology affects the number of viewable ads displayed by publishers, the price of ads and publisher profits, and user experience. We finally analyze the total welfare impact of ad viewability and examine how ad-blockers constrain publishers from both sides of the market.