Program

I will present the experimental design and technical infrastructure underlying the Internet Behavior Experiment (IBE). The IBE is a platform for large-scale randomized field experiments designed to examine online behaviors and associated consumer outcomes while exogenously manipulating participants’ exposure to advertising, tracking, and targeting. Its central contribution is the development of a novel technological infrastructure that securely and privately collects rich, longitudinal behavioral data across devices and enables randomized interventions across those devices. Launched in May 2025, the IBE is enrolling participants on a rolling basis.

Lead calling is a vital aspect of B2B prospecting, yet research on salesperson speech that enhances its effectiveness remains limited. Drawing on Symbolic Interactionism theory and grounded analysis, we develop a conceptual framework to examine how salespeople use communicative acts to convert leads. We identify linguistic acts (legitimizing, piquing, nudging) and paralinguistic acts (intensity, speech rate, pitch) as key elements of salesperson speech and conceptualize their impact on speaking turn level lead receptivity as the conversation unfolds. Additionally, we assess the influence of lead receptivity on conversion at the conversation level. Using data from 783 lead calling conversations containing 11,566 speaking turns between salespeople and B2B leads, we empirically test our hypotheses. We find that a 1 SD increase in lead receptivity raises conversion likelihood by 10.22%. Moreover, the impact of linguistic and paralinguistic acts on lead receptivity depends on the timing of their delivery within the conversation. For instance, speaking with higher, compared to lower, pitch earlier (later) in the conversation heightens receptivity by 13.88% (6.22%), whereas high (compared to low) levels of legitimizing raise receptivity by 2.99% earlier in the conversation but reduce it by 24.50% later in the conversation. Implications for theory and practice are discussed.

Digital platforms have fundamentally reshaped how consumers access entertainment. Audiences are increasingly moving away from traditional formats like physical media (e.g., CDs) towards on-demand digital platforms. This migration has profound implications for content creators and distributors, particularly because digital platforms not only distribute content, but they also influence how and when it is consumed through their release strategies and promotional tools.
Streaming platforms such as Spotify offer users extensive libraries and tools to curate their own listening experiences or follow curated playlists. Among these, platform-managed playlists, some with millions of followers, serve as promotional tools. Placement in these playlists can significantly boost a song’s exposure. Existing research has examined how musical attributes contribute to a song’s success. Over time, commercially successful songs have become increasingly simpler, with more repetition and less complexity in terms of pitch, harmony, and rhythm. Prior studies also suggest that the most successful songs strike a balance between familiarity and novelty, offering just enough uniqueness to stand out without alienating listeners. What remains underexplored, however, is the role of playlist inclusion as a driver of success.
This paper examines how such inclusion affects streaming counts and popularity metrics, and what types of songs are most likely to be featured. A particular emphasis is placed on musical and/or lyrical complexity, leveraging existing insights about how simplicity has become a commercial advantage in modern music. In sum, this research offers a novel, data-driven approach to understand the interactions between content and platforms in the digital entertainment economy.

This paper estimates how gun demand responds to gun violence and the information exposure it generates. Combining 4 million firearm transactions from a major online marketplace with incident-level data from the Gun Violence Archive (2010–2018) and measures of local news coverage, we study how exposure to information about gun violence shapes purchasing. We treat news coverage as a measure of information exposure — the salience and reach of gun-violence information across local markets — and consider incidents ranging from high-salience mass shootings to more localized events. The paper characterizes the demand response to gun violence and its information environment, examining how that response unfolds over time and across places.

Consumers form impressions of marketplace actors not from single messages, but from sequences of signals encountered over time. Prior work has treated identity as a stable reference against which each new message is judged. We argue identity is itself dynamic: a moving reference that audiences update with every new signal. The question, then, is not whether a single message fits, but what degree of deviation from an evolving identity most effectively engages audiences. Combining text embeddings on 20,000 sponsored Instagram posts from 940 influencers with preregistered experiments, we find an inverted U-shaped relationship between variation and engagement. Moderate breadth outperforms both narrow and highly diffuse identities, and new content performs best when it deviates moderately from prior posts: shifts into related categories (Beauty→Fashion) outperform both repetition and moves across distant ones (Beauty→Food). Moderate variation works because it sustains interest while preserving authenticity. Further, the effects depend on influencer lifecycle and post goals. Together, these findings reframe identity as a dynamic, cumulative inference and offer a scalable way to measure how it unfolds.

In today’s omnichannel world, many retailers radically transform their physical stores to meet changing customer preferences, typically emphasizing categories aligned with the new positioning, de-emphasizing categories that do not, and leaving a significant portion of categories unchanged in assortment size, composition, and presentation. This study examines how such a transformation impacts category performance. A quasi-experimental analysis of a German retailer that radically shifted from a low priced, transaction oriented store to a high priced, experience oriented format centered on fresh and ready to eat products, shows positive effects on sales of emphasized categories (+18 %) and negative effects on de-emphasized categories (-4 %). Interestingly, unchanged categories also experience negative sales effects (-6 %). The study then investigates which unchanged categories are more or less adversely impacted depending on eight category characteristics (8Ps): purchase frequency, penetration (capturing customers’ ‘familiarity’ with the category); perishability, price level, processed level (reflecting the category’s ‘fit’ with the new store concept); promotion intensity, portfolio and private label share (capturing a category’s prominence in the transformed store). Results indicate that the negative effects on unchanged category sales are amplified by familiarity (purchase frequency) and reduced by prominence (promotion intensity, private label share). Fit also matters: processed and higher-priced categories suffer more, while those high in perishability experience less negative effects.

Firm agents benefit from being perceived as helpful in collaborative conversations. Might how agents talk (i.e., how often they pause) influence how helpful they seem? Though speakers who spend more time in silence while talking are often perceived negatively, we suggest that brief pauses while speaking can be beneficial. Specifically, we argue that pausing encourages verbal assents from conversation partners (e.g., “yeah” or "uh-huh"), which leads them to perceive the speaker more positively. Analysis of customer service calls and four experiments support this account. Our findings indicate that, in collaborative conversations, brief pauses while speaking can make a person seem more helpful because they encourage conversation partners to assent.

Decision makers frequently need to predict crowd opinion, namely, the collective judgment and behavior of large groups of people. Examples include forecasting whether a new product will succeed and anticipating how consumers will perceive the benefits of a market offering. Although much prior research has examined the extent to which large language models (LLMs) can replicate human responses, little is known about how LLMs compare with humans in predicting aggregate crowd outcomes. In particular, can LLMs outperform humans when the task is to anticipate the behavior and perceptions of other people at scale? We propose that LLMs may hold an advantage in crowd prediction tasks because they are less susceptible to humans’ cognitive biases and are trained on extensive corpora that encode broad population-level regularities. We examine this question in the context of crowdfunding, where project descriptions are standardized and project outcomes are observable. To evaluate benefit inference, we first collect hedonic and utilitarian evaluations for a representative sample of 100 crowdfunding projects from 150 respondents to establish benchmark perceptions. We then ask an additional 150 respondents and eight LLMs to predict these aggregate perceptions under comparable conditions. We further compare the ability of humans and LLMs to predict project success outcomes. The results reveal a performance divergence across tasks. Humans outperform LLMs in predicting aggregate benefit perceptions, whereas LLMs outperform humans in forecasting project success. In addition, LLMs perform relatively better in predicting hedonic than utilitarian benefits, suggesting that they may be less affected by cognitive biases associated with hedonic inference. For project success prediction, we find evidence that LLMs incorporate a broader range of available information, whereas humans rely more heavily on salient dimensions. Together, these findings contribute to emerging research on AI-based social inference and collective outcome prediction.

As part of their social commerce strategies, brands not only promote their products using their own accounts (brand-owned accounts) but also rely on authorized third-party sellers (authorized seller accounts). However, it is unclear whether posts from authorized sellers reinforce the brand’s messages (synergistic effect) or create clutter that weakens brand communications (antagonistic effect). To investigate this, we analyze data from 1,291 brand-owned and 1,037 authorized seller accounts associated with 202 brands selling on Douyin between January and December 2023. We find that their interaction is negative, indicating channel antagonism. Building on this finding, we evaluate temporal spacing as a non-price coordination mechanism.
Using what-if simulations under typical posting levels, we show that asynchronous posting leads to a 5.2% increase in brand sales. However, while brand-owned accounts benefit from temporal spacing, authorized sellers do not, creating an asymmetric incentive structure that complicates voluntary coordination. Our findings contribute to the channel management literature by (1) documenting a new form of intra-channel conflict, (2) introducing temporal spacing as a coordination mechanism distinct from the contractual instruments studied in prior work, and (3) revealing that channel partners face asymmetric payoffs from coordination, motivating the design of incentive contracts to implement coordinated content strategies.

Research on clickstream considers a customer’s online browsing behavior as a path. However, considering customer browsing behavior as a path may prevent us from gaining enormous information. Relying on two related studies, this paper attempts to look at customer browsing behavior from a network perspective. In Study 1, through qualitative interviews, and using template and text analysis, the paper establishes that indeed customer online browsing may look like a network and that considering the browsing as interconnected nodes of websites may provide insights regarding customer purchases. Building on the tenets of network theory, the paper proposes that actual edges, betweenness centralization and network density of a customer browsing network influence customer purchases (amount) and the relationships are moderated by variety of websites visited and volume of information gained. In Study 2, using data for 1262 customers’ browsing behavior for online food delivery product category and using a robust empirical approach, the paper finds support for the effects of network characteristics and contingencies on consumer purchase amount. In Study 3, we provide a classification algorithm that helps customer segment using observed network characteristics. Using various robustness analyses, the paper contributes to the network theory, customer browsing literature with clickstream data, and respond to calls for considering customer browsing from a network perspective.