We study how cookie-consent design affects consumer welfare using a framed field experiment that randomizes consent banners. We examine three margins: visual nudges, defaults when users dismiss banners, and whether consent is set at the website or browser level. Visual changes have little effect, but deliberate obstruction such as hiding "reject all" substantially shifts choices. Because many users close banners, welfare depends on both the website default and users' beliefs about defaults. We estimate a model of consumer choice. The model estimates show that when consent is requested on a site-by-site basis, the optimal design eliminates obstructions and aligns defaults with the majority preference. Relative to this optimum, typical U.S. consent practices cost 20% of surplus through obstruction, and typical EU consent practices decrease surplus by 6%. Correcting users' beliefs about website default raises surplus by 15% compared to the optimal site-by-site consent. The largest welfare gains, however, come from replacing site-by-site consent with browser-level consent, which helps consumers avoid the time costs of making repeated consent decisions.

Markets are a promising way to coordinate AI agent activity for similar reasons to those used to justify markets more broadly. In order to effectively participate in markets, agents need to have informative signals of their own ability to successfully complete a task and the cost of doing so. We propose MarketBench, a benchmark for assessing whether AI agents have these capabilities. We use a 93-task subset of SWE-bench Lite, a software engineering benchmark, with six recently released LLMs as a demonstration. These LLMs are miscalibrated on both success probability and token usage, and auctions built from these self-reports diverge from a full-information allocation. A follow-up intervention where we add information about capabilities from prior experiments to the context improves calibration, but only modestly narrows the gap to a full-information benchmark. We also document the performance of a market-based scaffolding with these LLMs. Our results point to self-assessment as a key bottleneck for market-style coordination of AI agents.

Digital platforms offer consumers vertically integrated products alongside third-party alternatives. Using a field experiment that hides Amazon brands, we quantify the impacts of vertical integration—through both the availability of private labels and their placement in search results. In their absence, consumers choose similar products with no increase in search effort or cross-platform substitution. Yet some consumers prefer Amazon brands, and removing these products reduces consumer welfare by 5.5 percent. Contrary to concerns about self-preferencing, demoting Amazon brands in search results does not increase welfare, as heterogeneous preferences make their prominent placement valuable to a meaningful share of consumers.

We describe the emerging business-to-business market for large language model (LLM) inference and document key empirical patterns in its supply, pricing, and dynamics, using data from OpenRouter. First, supply has expanded rapidly: the number of commercially available models, model creators, and inference providers has grown sharply, driven heavily by open-source entrants. Second, the price of intelligence has fallen roughly a thousandfold, and open-source models now cost about 90 percent less than comparable closed-source ones. Third, the market is highly dynamic, with frequent turnover among leading models and creators. Fourth, we document substantial horizontal and vertical differentiation: no single model dominates across use cases, and demand for intelligence varies widely across applications. We place these patterns in historical perspective alongside earlier general-purpose technologies.

AI agents—autonomous systems that perceive, reason, and act on behalf of human principals—are poised to transform digital markets by dramatically reducing transaction costs. This chapter evaluates the economic implications of this transition, adopting a consumer-oriented view of agents as market participants that can search, negotiate, and transact directly. From the demand side, agent adoption reflects derived demand: users trade off decision quality against effort reduction, with outcomes mediated by agent capability and task context. On the supply side, firms will design, integrate, and monetize agents, with outcomes hinging on whether agents operate within or across platforms. At the market level, agents create efficiency gains from lower search, communication, and contracting costs, but also introduce frictions such as congestion and price obfuscation. By lowering the costs of preference elicitation, contract enforcement, and identity verification, agents expand the feasible set of market designs but also raise novel regulatory challenges. While the net welfare effects remain an empirical question, the rapid onset of AI-mediated transactions presents a unique opportunity for economic research to inform real-world policy and market design.

Understanding the behavior of users online is important for researchers, policymakers, and companies. But measuring behavior online and conducting experiments is difficult for independent researchers, who do not have access to the user bases or software of technology companies. We introduce Webmunk 2, an improved open-source tool designed to make conducting online studies much easier. Webmunk 2 is a substantial improvement over the original Webmunk, released in 2024. It is easier to use, cheaper to deploy, and more powerful than the original. The user-facing side of Webmunk 2 is a browser extension that can track consumer browsing behavior, capture screenshots for analysis, and experimentally modify consumers' experiences as they browse the Internet. A key innovation is that Webmunk 2 uses a standard back-end stack consisting of Jitsu for data ingestion, Firebase for user management and configuration, and BigQuery for data storage. This makes it more accessible to researchers and software developers who are familiar with these widely-used tools. Through this extension, researchers can collect a host of consumer data, from URLs to ad impressions, clicks, and user responses to in-browser surveys. The extension can also modify information and change the look of a web page, allowing researchers to implement interventions that vary across study participants. We demonstrate the power of Webmunk 2 by discussing ongoing research applications.

We study how information that may simultaneously signal the degree of competition and vacancy quality affects job search. To do so, we conduct three experiments on a large online job platform in which the treatment varies what information is shown to job seekers. Information about the number of prior applicants to a vacancy increases the number of applications and redirects them to vacancies with few prior applications. Information about vacancy age increases application rates, especially to new vacancies. To further investigate the causal mechanisms, we conduct and analyze a survey choice experiment. We conclude that job seekers prefer to avoid competition rather than using the popularity of a vacancy as a signal of quality.

We study the demand and supply implications of occupational licensing using transaction-level data from a large online platform for home improvement services. We find that demand is more responsive to a professional's reviews than to the professional's platform-verified licensing status. We show some evidence that consumers view licenses and reviews as substitutes. We confirm the generality of our findings off the platform in an independent consumer survey. Combining state-level licensing regulation data with platform micro-data, we find that more stringent requirements are associated with less competition, higher prices, and no increase in demand or consumer satisfaction.

Mergers among digital platforms are increasingly receiving public and regulatory attention. These mergers may benefit users if network effects from a combined platform are large enough or may hurt users if the two platforms are differentiated and one of the platforms is shut down. We study the net effect of this trade-off in the context of the merger between the two largest platforms for pet-sitting services. We exploit geographic variation in pre-merger market shares and a difference-in-differences approach to causally estimate network effects. We find that users of the acquiring platform benefited from the merger because of network effects. However, users of the acquired platform were more likely to exit the market, for reasons including switching costs, coordination failures, and disintermediation. Network effects and attrition offset each other such that at the market level consumers are, on average, not substantially better off with a single combined platform than with two separate and competing platforms. Our results highlight the importance of platform differentiation even when platforms enjoy network effects, which has important implications for antitrust authorities and platform owners.

Many online reputation systems operate by asking volunteers to write reviews for free. As a result, a large share of buyers do not review, and those who do review are self-selected. This can cause the reputation system to miss important information about seller quality. We study the extent to which a platform can improve market outcomes by attempting to increase the amount and quality of information collected by its reputation system. We do so by analyzing a randomized experiment conducted by Airbnb. In the treatment, buyers were offered a coupon to review listings that had no prior reviews. In the control, buyers were not offered any incentive to review. We find that although the treatment induced additional reviews that were more negative on average, these reviews did not affect the number of nights sold or total revenue. Furthermore, we find that, contrary to the treatment's intended effect, Airbnb's incentivized program caused transaction quality for treated sellers to fall. We examine how the quality of the induced reviews, market conditions, and the design of Airbnb's reputation system can explain our findings.

We study whether Amazon engages in self-preferencing on its marketplace by favoring its own brands (e.g., Amazon Basics) in search. To address this question, we collect new micro-level consumer search data using a custom browser extension installed by a panel of study participants. Using this methodology, we observe search positions, search behavior, and product characteristics. We find that Amazon branded products are indeed ranked higher than observably similar products in consumer search results. The prominence given to Amazon brands is 30% to 60% of the prominence granted to sponsored products.

We study the effects of enabling peer supply through Airbnb in the accommodation industry. We present a model of competition between flexible and dedicated sellers - peer hosts and hotels - who provide differentiated products. We estimate this model using data from major US cities and quantify the welfare effects of Airbnb on travelers, hosts, and hotels. The welfare gains are concentrated in locations (New York) and times (New Years Eve) when hotels are capacity constrained. This occurs because peer hosts are responsive to market conditions, expand supply as hotels fill up, and keep hotel prices down as a result.

Reputation systems are used by nearly every digital marketplace, but designs vary and the effects of these designs are not well understood. We use a large-scale experiment on Airbnb to study the causal effects of one particular design choice — the timing with which feedback by one user about another is revealed on the platform. Feedback was hidden until both parties submitted a review in the treatment group and was revealed immediately after submission in the control group. The treatment stimulated more reviewing in total. This is due to users' curiosity about what their counterparty wrote and/or the desire to have feedback visible to other users. We also show that the treatment reduced retaliation and reciprocation in feedback and led to lower ratings as a result. The effects of the policy on feedback did not translate into reduced adverse selection on the platform.

We study the role of shocks to parental income in determining the labor market outcomes of children entering the labor market. We find that a child whose parent loses a job prior to the child’s labor market entry is, on average, induced to work 9 percent more in the 3 years following labor market entry than a child whose parents lose a job after the child’s entry. This effect is concentrated on the extensive margin and decreases in magnitude over time. We find no evidence that parental support affects the quality of the initial job that entrants find.

Job search is a key choice variable in theories of labor markets but is difficult to measure directly. We develop a job search activity index based on Google search data, the Google Job Search Index (GJSI). We validate the GJSI with both survey- and web-based measures of job search. Unlike those measures, the GJSI is high-frequency, geographically precise, and available in real time.We demonstrate the GJSI’s utility by using it to study the effects of changes in the unemployment insurance (UI) system between 2008 and 2014 on job search, finding no evidence of economically meaningful decreases in aggregate search activity.

Default contribution rates for 401(k) pension plans powerfully influence workers’ choices. Potential causes include opt-out costs, procrastination, inattention, and psychological anchoring. We examine the welfare implications of defaults under each of these theories. We show how the optimal default, the magnitude of the welfare effects, and the degree of normative ambiguity depend on the behavioral model, the scope of the choice domain deemed welfare-relevant, the use of penalties for passive choice, and other 401(k) plan features. Depending on which theory and welfare perspective one adopts, virtually any default contribution rate may be optimal. Still, our analysis provides reasonably robust justifications for setting the default either at the highest contribution rate matched by the employer or – contrary to common wisdom – at zero. We also identify the types of empirical evidence needed to determine which case is applicable.

Artificial Intelligence (AI) can generate text virtually indistinguishable from text written by humans. A key question, then, is whether people believe news generated by AI as much as news generated by humans. AI is viewed as lacking human motives and emotions, suggesting that people might view news written by AI as more accurate. By contrast, two pre-registered experiments on representative U.S. samples (N=4,034) showed that people rated news written by AI as less accurate than news written by humans. People were more likely to incorrectly rate news written by AI (vs. a human) as inaccurate when it was actually true, and more likely to correctly rate it as inaccurate when it was indeed false. Our findings are important given the increasing adoption of AI in news generation, and the associated ethical and governance pressures to disclose it use and address standards of transparency and accountability.

Reviews and other evaluations are used by consumers to decide what goods to buy and by firms to choose whom to trade with, hire, or promote. However, because potential reviewers are not compensated for submitting reviews and may have reasons to omit relevant information in their reviews, reviews may be biased. We use the setting of Airbnb to study the determinants of reviewing behavior, the extent to which reviews are biased, and whether changes in the design of reputation systems can reduce that bias. We find that reviews on Airbnb are generally informative and 97% of guests privately report having positive experiences. Using two field experiments intended to reduce bias, we show that non-reviewers tend to have worse experiences than reviewers and that strategic reviewing behavior occurred on the site, although the aggregate effect of the strategic behavior was relatively small. We use a quantitative exercise to show that the mechanisms for bias that we document decrease the rate of reviews with negative text and a non-recommendation by just .86 percentage points. Lastly, we discuss how online marketplaces can design more informative review systems.

Commentary on reputation design, with reference to the paper "Do Incentives to Review Help the Market?".

Summary of our Dog Eat Dog paper.

This article provided a practitioner overview of recent work on reputation system design.

This article provides an overview of the economics of digital marketplaces.

This article discusess how privacy and the design of digital marketplaces may affect inequality.

This paper documents three stylized facts about the demand for Large Language Models (LLMs) using data from OpenRouter, a prominent LLM marketplace. First, new models experience rapid initial adoption that stabilizes within weeks. Second, model releases differ substantially in whether they primarily attract new users or substitute demand from competing models. Third, multi-homing—using multiple models simultaneously—is common among apps. These findings suggest significant horizontal and vertical differentiation in the LLM market, implying opportunities for providers to maintain demand and pricing power despite rapid technological advances.

Two-sided marketplaces are distinguished by the fact that both sides have preferences regarding each others' non-price characteristics. This paper studies how digital platform design affects transaction costs and volume in these markets by analyzing the decisions of guests and hosts to search and match with each other on Airbnb. I show that the two-sided nature of the market is important. Through 2014, rejections of guests by hosts occur for 42% of inquiries regarding booking and these rejections causally decrease the rate at which guests eventually book on the platform by 43% to 70%. Rejections are primarily caused by stale vacancies and the screening of guests by hosts. I use data on search and communication to estimate a model of guest and host choices. I apply this model to study the effects of search engine design and find that, by tracking listing availability, Airbnb reduces rejections by 59%. I then show that incorporating host preferences into rankings can further increase match rates and discuss how Airbnb's subsequent innovations reflect these findings.

Digital matching marketplaces are characterized by user heterogeneity, limited capacity, and dynamic market clearing. These features create spillovers between users. For example, an Airbnb listing booked by one guest cannot be booked by another guest for the same night. Spillovers limit the applicability of many experimental and observational methods for evaluating the effects of marketplace policies. In this paper, I show how to use marketplace simulations as an input into the design of user acquisition strategies and ranking algorithms. I calibrate a marketplace simulation using data on searches and transactions from Airbnb and use it to address three topics: the returns to scale in matching, the heterogeneity in returns to user acquisition, and the size of bias in experimental designs. I find that returns to scale are initially increasing due to market thickness effects and then decreasing due to availability frictions in search. Furthermore, heterogeneity in the value of listings to the platform is large – the effect of acquiring 25% more listings on bookings varies between -4.1% and 5.4% depending on the quartile of listing quality. I then measure the extent of bias in experimental treatment effects due to spillovers. The treatment effect of a better ranking algorithm on conversion rates is overstated by 53% when a quarter of users are randomized into treatment.