Research

Abstract: In customer effort goods—online education, fitness, enterprise software—customers realize value not at purchase but only through their own costly effort afterward, and the firm cannot readily observe whether that effort produced value. Firms therefore manage these products with engagement metrics as a proxy for value, but the engagement-maximizing design need not be value-maximizing. This paper develops a structural framework for value-centric post-adoption design. The model has two layers: (i) a motivation layer, in which the firm's lever shapes the customer's costly, forward-looking effort; and (ii) a value layer, in which that effort accumulates into the outcome the customer is after. The key tension is that the unit that motivates effort need not be the unit that produces value. We apply the framework to an online learning platform, where teachers partition content into goals and students' completed work builds the knowledge that determines test performance: goals motivate, but lessons teach, so smaller goals raise engagement while fragmenting the coherence that produces learning. Given the complexity of estimation, we adapt adversarial inverse reinforcement learning to recover the model's primitives, anchoring them to a measured value outcome, and identify two latent segments—high-grit and low-grit—that differ in their cost of effort. Neither uniform design maximizes learning across both segments and all content: lesson-aligned goals help high-grit students, while sub-lesson goals are better for low-grit students, especially on longer and more difficult modules. The value-maximizing design thus varies with the student and the content in ways the model makes precise and in ways a firm can act on.

Abstract: Digital platforms strategically choose what information to provide to their users. Using a uniquely constructed dataset of over 121,000 YouTube videos collected in 2023, we examine the causal impact of engagement graphs—a novel form of within-content engagement information that displays moment-by-moment prior user engagement levels throughout a video. Our identification strategy exploits YouTube's batch processing approach to engagement graph activation: we reverse-engineer the platform's criteria, revealing that activation depends on a common view threshold and an eligibility date determined solely by upload timing, with batch processing creating quasi-random variation in treatment timing across otherwise similar videos. We leverage this variation to implement a difference-in-differences strategy with matched samples. Our analysis finds that engagement graphs increase content viewing, liking, and commenting in post-treatment periods, while leaving comment sentiment and depth unchanged. To deepen our understanding of these effects, we use large language models to construct transcript-based measures of each video's informational content and narrative structure, and characterize how the treatment effect varies along these dimensions. The graph's impact is higher among videos whose value is harder for viewers to assess as they watch, for example, videos that deliver their payoffs later in the runtime, whereas narrative structure plays no moderating role. On the supply side, exploiting the feature's platform-wide rollout, we find no evidence that creators adjust their content production volume or characteristics. Our findings speak to platform design, content discovery, and the broader question of how making collective attention visible shapes consumption in digital markets.

Abstract: When expressing interest is costless, claims of special interest are cheap talk: senders who genuinely value a particular partner cannot distinguish themselves from those making opportunistic attempts. This problem is pervasive in matching markets, where participants hold private information about match-specific preferences that potential partners would value but cannot verify. 

This paper examines whether costly signaling can resolve this problem. We study a novel U.S.-based online dating platform where users must create personalized drawings—rather than simply swiping—to initiate match requests. We develop a conceptual model in which senders choose drawing effort as a costly signal of match-specific valuation, then construct a computer vision-based measure of the effort signal and apply it to over 577,000 match requests. We document three findings. First, senders vary effort strategically: effort signals vary primarily within rather than between senders, increasing with receiver desirability, geographic proximity, and age similarity, consistent with effort reflecting match-specific valuation rather than fixed sender characteristics. Second, receivers respond to effort signals: within a receiver, higher-effort requests are significantly more likely to be accepted, and this relationship holds beyond what observable sender and receiver characteristics as well as match-specific characteristics can explain. Third, conditional on match, effort signals predict downstream outcomes: within a sender, higher-effort matches are more likely to result in post-match conversation, consistent with effort reflecting genuine interest rather than mere attempts to secure a match. 

Together, these findings provide evidence that costly signaling through drawings enables senders to credibly transmit information regarding their match-specific preferences in this market. Our results have important implications for platform design: reducing the cost of expressing interest lowers friction but may destroy a valuable information channel.