Research

Firms, especially early-stage innovators, often need to decide which innovation direction to pursue, a choice usually made before the market reveals the innovation's value. When its own demand-side information is thin, the firm may infer that value from the supply side by observing what competitors do. We focus on one competitive action in particular: imitation. Imitation costs real effort, serving as credible evidence that the innovation direction is valuable. Measuring how firms respond to competitor imitation causally is difficult, however, because competitors act endogenously. To address this challenge, we run a field experiment on GitHub, where we randomize competitor imitation and measure how developers respond to this supply-side action. We find that imitation on a single project leads developers to reduce development across their broader portfolio, with no gain on the imitated project. We propose a model showing that the firm rationally stays committed to its current direction even when it recognizes that imitation may reflect the competitor's strategic incentives. To show the main effect holds across different markets and to extend the result, we conduct two additional studies: first, a tagline experiment shows verbal compliments alone have no effect, and second, observational music data reveals this impact persists in the long run.

Price haggling is widely viewed as a source of friction in retail markets: customers bear bargaining costs, and firms face operational complexity. This logic has motivated some firms to adopt fixed pricing. We show that in markets organized around personal selling, eliminating negotiation can backfire. Using lead and transaction data from a major automaker in China that introduced a one-price policy for one model while retaining negotiation for all others at the same dealerships, we find that the policy reduced sales of the targeted model and shifted demand toward models where negotiation remained available. Notably, although the fixed price was below the prior average negotiated price at most dealerships, the pattern persists within this subgroup. We argue that the process of negotiation serves as an effective deal-closing tool for the salesperson. When this instrument is removed for one product, salespeople become less equipped to convert leads into sales and may redirect customers toward products for which negotiation remains available. Consistent with this mechanism, demand switching is larger among salespeople who relied more heavily on price discussion prior to the policy.

Online reviews are central to consumer markets, yet many consumers stay silent and posted reviews can be unrepresentative. We study how generative AI affects review contribution, composition, and downstream consumer response. Using a custom-built video platform, we test two LLM-powered assistance formats: one-click generation from user-provided keywords or prompts, and an interactive chatbot for collaborative drafting. Four findings emerge. First, AI assistance increases comment rates relative to control, but by a magnitude broadly comparable to the industry-standard Guided Prompt condition. Its distinctive effect appears in the interactive chat format, which disproportionately increases posting among users typically reluctant to review online. Second, AI-assisted comments exhibit positive sentiment bias. Third, they more closely align with users' later unaided expressions, reflecting AI's role in both amplifying user articulation and elaborating user input with video-relevant content. Fourth, AI-assisted reviews increase downstream choice probability, especially among consumers who value broader topical coverage and longer reviews. These findings highlight that AI's contribution is not captured by review volume alone: chat-style AI can make the silent majority more vocal, AI-assisted reviews reshape how reviews are composed, and these reviews increase downstream content choice, with associated platform-design risks.

In many markets, buyers select providers without observing their underlying skills or experience. We propose that customization requests can serve as a screening device: experienced providers are better placed to recognize when standardized solutions dominate unnecessary customization, so their willingness to decline a request becomes informative about expertise. We develop a model in which experienced providers opt for standardization when customization has low value, while inexperienced providers customize because they cannot distinguish when customization helps. We examine these predictions in a field setting on Fiverr, where freelancers create Facebook video ads and receive an optional customization request. Freelancers who accept the request produce ads with lower click-through rates, yet their decision to customize is uncorrelated with observable platform quality metrics. The same freelancers who customized also produce ads that perform worse in a separate unrelated task, indicating that the customization decision signals persistent quality differences. A provider's willingness to say no can be a credible signal of expertise.

When a brand expands its retail channels and allows third-party retailers to sell its products for the first time, does this result in positive spillovers or cannibalization from the brand's direct-to-consumer (DTC) business? We provide the first empirical study of the impact of a new wholesale channel on a brand's own-store sales. The findings are obtained from transaction-level data provided by a retailer that launched a new wholesale program, with staggered partner entry across U.S. markets over a five-year period. The new wholesale channel increased new customer acquisition, active customer retention, and reactivation of lapsed customers in the brand's own DTC channels. Spillovers were largest in non-urban markets with low competitive intensity. Different partner types generate distinct spillover patterns: major national retailers drive new customer acquisition, while independent specialty retailers drive active customer retention. Our findings suggest that wholesale channels complement rather than cannibalize a brand's direct business.

Digital media literacy interventions to counter misinformation sharing have focused primarily on teaching strategies to address reasoning-based techniques that manipulate users' evaluation of the information in social media posts. Yet emotion-based techniques that manipulate how users feel when reading a post may matter more for how misinformation spreads. In a preregistered field experiment with Facebook users in Kenya (N = 8,684), a setting underrepresented in this literature, we test separate text message courses countering each technique type in a pre-post survey design. We find that only the course countering emotion-based techniques improves users' sharing discernment, the difference between non-misinformation and misinformation sharing. Notably, the emotion-based course reduces sharing across all misinformation types, even posts using only reasoning-based techniques, showing that the emotion-based course generalizes beyond the specific technique type it targets. Treatment effects of the emotion-based course persist in a follow-up survey seven to eleven weeks later.

Artificial intelligence is transforming how firms identify, structure, and prioritize customer needs — the "voice of the customer." We review the types of user-generated data firms now use to understand customer needs, the challenges of analyzing each type, and the machine-learning and AI methods commonly applied. We match these methods to firms' goals and data structures and map them to applications, offering guidance on selecting appropriate methods for a given research question.

The "job talk" is a standard element of faculty recruiting, and how audiences treat candidates could have disparate impact on protected groups, including women. We annotated 156 job talks from five top-ranked engineering and science departments for 13 categories of questions and comments. Differences by gender in the number, nature, and total duration of audience questions and comments are neither material nor statistically significant — the median difference in the duration of questioning ranged from zero to under two minutes across departments. In some departments, candidates who were interrupted more often were more likely to be offered a position, challenging the premise that interruptions are necessarily prejudicial.