Beyond Search: LLM Adoption and Web Traffic Concentration

Best Student Paper Award, WISE 2025

with Samira Gholami, Cristobal Cheyre, and Alessandro Acquisti

PDF

This paper studies how the diffusion of large language models (LLMs) as consumer-facing information intermediaries reshapes web traffic allocation and market concentration. Using large-scale behavioral data, we document substitution patterns between traditional search engines, LLM-based interfaces, and downstream content providers, and show that LLM adoption alters referral flows and concentrates attention away from the open web, with heterogeneous effects across content categories. The results speak to ongoing antitrust and platform-governance debates around AI-driven intermediaries.

Advertising, Personalization and Search Costs

Accepted, ISMS Marketing Science Conference 2026

with Cristobal Cheyre and Alessandro Acquisti

Using a large-scale online field experiment with detailed browsing and ad-exposure data, this project studies how personalized advertising affects consumer search costs, consideration sets, and downstream purchases. We quantify how targeted and contextual ads change search intensity and navigation behavior, shedding light on the dual role of advertising as both an informational device and a potential distortion in consumer decision-making, with implications for industrial organization, consumer protection, and advertising regulation.

Cursed by the Algorithm: Cue-Triggered Consumption and Platform Recommendation

Solo-authored

SSRN

I study an intermediary that learns what to offer an agent from what she takes up, when take-up is sometimes cue-triggered rather than deliberate. The leading case is an engagement-maximizing recommender serving users who may enter a "hot" state in which one content category is transiently overvalued. The platform observes consumption but not the cognitive state, so compulsion and genuine preference are observationally equivalent at the margin on which it acts, and its cursed learning concentrates triggering content on the users it harms. Standard users converge to the first best. Vulnerable users spiral into disengagement; from small initial vulnerability the spiral ignites if and only if the trigger's genuine weight in preferences is high enough, independently of the strength of the craving, and the welfare loss is hump-shaped in that weight, so the users harmed most have a mild genuine interest in the triggering content. The equilibrium is Pareto-dominated: an informed platform would hold back and earn more. Content caps and minimum-diversity floors on feed composition, applied uniformly, raise welfare without identifying who is vulnerable. When take-up carries a price, as in targeted advertising, the price disciplines the number of impulse purchases but not the overspend within them, and it sharpens the incentive to target.

Media coverage: Featured in an interview on Tech Policy Press.

An Experimental Infrastructure for Ecologically Valid Studies of Online Advertising, Tracking, and Targeting

with Cristobal Cheyre, Li Jiang, Florian Schaub, Zijun Ding, Yucheng Li, and Alessandro Acquisti

This paper introduces an experimental methodology for studying the impacts of online advertising, tracking, and targeting on users. The infrastructure deploys a client–server architecture that enables randomized field experiments by assigning participants to ad-blocking, anti-tracking, or full ad-exposure conditions, capturing rich longitudinal data across browsers, email, and mobile devices while preserving ecological validity. By combining fine-grained behavioral measures with repeated surveys, it provides a rigorous framework for causal inference on consumer welfare, information-seeking behavior, and privacy in digital advertising markets.

SafeFeeds

with Miguel Risco

We study the impact of education and algorithm reset on adolescents' well-being. Social-media recommender systems learn from engagement and can concentrate attention on a narrow set of content, with particular risks for younger users. In a field experiment, we test two low-cost interventions: a digital-literacy component that helps teenagers understand how their feeds are curated, and a reset of the recommendation algorithm that clears its learned profile. We measure how each intervention affects feed composition, time on platform, and self-reported well-being, and whether such uniformly applied tools can improve adolescent outcomes without identifying which users are most at risk.

Digital Disclosure, Information Avoidance, and Consumer Choice in Online Grocery Markets — An Online Experiment

with Mario Martinez Jimenez

An online grocery experiment that overlays Yuka-style nutritional scores on products as consumers shop, studying how digital disclosure — and consumers' tendency to avoid it — shapes their food choices.

Consent Without Choice

with Francesco Decarolis, Alessandro Acquisti, and Cristobal Cheyre

Examines how consent banners and choice architecture shape users' privacy decisions online under the GDPR, and what these patterns imply for the design of consent and data-protection rules.

Making Data Protection Fit for the Age of AI

CERRE — Centre on Regulation in Europe, 2026

with Marco Bassini

PDF

This CERRE report analyzes the EU's proposed Digital Omnibus and the reform of the GDPR in the age of artificial intelligence, combining legal and economic perspectives. It argues that data protection and AI innovation are not inherently in conflict: in many settings the interests of users and developers are aligned, and the challenge is a framework that does not distinguish the cases where they converge from those where they genuinely diverge. The report assesses key provisions — from the notion of personal data to legitimate interest for AI training and consent management — and proposes a phased path toward a more proportionate, risk-based regime.

Media coverage: Il Sole 24 Ore and MediaLaws.

Unlocking Growth: The Economic Effects of GDPR in Europe

Clifford Chance, 2025

with Francesco Decarolis

This report studies the economic effects of the General Data Protection Regulation (GDPR) on firm behavior, investment, and innovation across European countries. Bringing together cross-country evidence and institutional variation in enforcement and compliance, it documents the costs GDPR has imposed on firms — including lower profits and reduced investment — and argues for a more proportionate, risk-based approach to data regulation.

Media coverage: About Amazon and MLex.

The Digital Omnibus and the Economic Effects of Centralising Consent

with Francesco Decarolis

Studies the economic effects of centralising consent under the EU's proposed Digital Omnibus, analyzing how moving consent management to a single, browser- or system-level signal would affect users, firms, and competition in digital markets.