Marketing Research Seminar Series

Doctoral Student Research Presentations

Welcome and Keynote Address

Generative AI models are increasingly used to produce marketing content. Since off-the-shelf models are misaligned with desired marketing outcomes, they are fine-tuned using content experiments that identify what content is correlated with higher engagement. Yet optimizing only for what works risks overfitting, reward hacking, and poor generalization, yielding content that succeeds in-sample but fail in new contexts or drift toward clickbait. We propose a principled knowledge-alignment framework that moves beyond merely what works to why it works. In our approach, an LLM iteratively generates hypotheses about mechanisms (e.g., emotional language, narrative framing) to explain observed performance differences on a small set of data (abduction), then validates them on held-out data (induction). The optimized set of validated hypotheses form an interpretable, domain-specific knowledge base that regularizes fine-tuning via Direct Preference Optimization (DPO), constraining the model toward generalizable principles. Our LLM-based approach extends the tradition of theory-guided machine learning to domains where relevant knowledge is tacit and therefore hard to explicitly encode in models. Using a dataset of over 23,000 A/B-tested news headlines across 4,500+ articles, we show that our knowledge-guided framework outperforms supervised fine-tuning, DPO and multi-dimensional DPO in improving engagement (click-through), while avoiding clickbait and maintaining lexical diversity.

Algorithmic personalization is pivotal to the digital economy worldwide. To generate personalized recommendations, companies frequently elicit consumer preferences by asking them to select categories of interest (e.g., video genres). This research examines whether preference-elicitation questions effectively capture consumer preferences and, if not, why this occurs and how it can be mitigated. Six pre-registered studies (along with five supplemental studies and a pilot study total N = 6,767) spanning eight product domains (e.g., videos, news, wine) reveal that consumers share less diverse preferences with algorithms than they actually possess. Instead, they focus on their core preferences while omitting tangential ones—a phenomenon termed preference filtering. It is driven by the belief that sharing diverse preferences with an algorithm makes it prone to misclassification. Redesigning the preference elicitation task to attenuate the perceived risk of misclassification can encourage consumers to share more diverse preferences with algorithmic recommenders, which deters the formation of ''filter bubbles''. Paradoxically, contrary to the lay belief in misclassification, two studies on a custom-built video-streaming website show that more diverse recommendations enhance consumer evaluation of the algorithmic recommendation service. These findings offer valuable insights for companies that rely on algorithms to engage consumers.

Consumers often solicit preferences from each other when deciding what to consume together. Prior work has shown that consumers are often hesitant to express a preference for joint consumption decisions, but expressing a preference for an appealing option can ease the decision-making process. We extend this work by evaluating the effectiveness of preference communication that rejects an unappealing option from a choice set. Despite only eliminating a single (and unappealing) option, such preference communication reduces decision difficulty for joint consumption because rejecting an unappealing option increases the perception of preference similarity with a consumption partner. As such, our effect is not observed when an unappealing option is rejected for reasons other than personal preference or when making decisions for individual consumption. Further, rejecting an unappealing option is a stronger signal of preference similarity in less established relationships. Together, this research contributes to the literature on decision making for joint consumption, interpersonal inference-making, and preference communication, and offers managerial insights for firms and individuals wishing to increase the effectiveness of decision-making for shared consumption.

Growing inequality continues to shape consumers' lives, widening the gap between the advantaged and the disadvantaged. This research examines how perceived disadvantage versus advantage influences consumer pickiness, defined as the latitude of acceptance around idiosyncratic ideal points. Across eight studies—including an analysis of consumer panel data, a field study at a local food pantry, and six preregistered experiments—we find that a sense of disadvantage leads consumers to be less picky, whereas a sense of advantage leads consumers to be more picky. These effects are driven by differences in psychological entitlement: disadvantage reduces entitlement, while advantage increases it, which in turn affects pickiness. Importantly, these differences emerge even in the absence of resource or external constraints, highlighting entitlement as a key psychological mechanism. We further find that the effects are moderated by social dominance orientation, such that the impact of disadvantage versus advantage on entitlement and pickiness is attenuated among individuals who do not endorse existing inequalities.

Question Answering (QA) on customer-facing platforms (e-commerce, travel, education, and brand websites) often suffers from delayed, low-quality responses and limited user participation. While customer reviews are abundant, their unstructured nature limits direct use to answer specific questions, and Large Language Models (LLMs) alone lack product-specific details and subjective insights. To address these challenges, we propose and evaluate a novel QA framework that integrates LLMs with reviews through Retrieval-Augmented Generation (RAG). Our framework incorporates three components: (1) RAG for dynamically retrieving relevant reviews at inference time (2) a question–review type matching module that enhances topical alignment (3) an answerability classifier that determines whether a reliable answer can be generated. Using a data set of 500 Amazon questions, 2000+ human responses, and 14,000 review sentences, we systematically evaluate different model variants for both lexical similarity metrics (ROUGE-L) and human judgments. Our full model improves lexical similarity scores by 50% from baseline LLM answers, matches or exceeds 72% of human responses, and approaches the best human answers in clarity, relevance, and informativeness. In particular, human evaluations show that our full model performs particularly well on subjective questions and lexical similarity metrics fail to capture this performance gain. Overall, our findings show how LLMs and reviews can be combined to build scalable QA systems, while also revealing the limits of lexical similarity metrics and highlighting the importance of human-centered evaluation.

Salespeople frequently face the dynamic screening decision of whether to persist in a conversation or abandon it to pursue the next lead. Yet, little is known about how these decisions are made, whether they are efficient, or how to improve them. We study these decisions in the context of high-volume outbound sales where leads are ample, but time is scarce and failure is common. We formalize the dynamic screening decision as an optimal stopping problem and develop a generative language model-based sequential decision agent — a stopping agent — that learns whether and when to quit conversations by imitating a retrospectively-inferred optimal stopping policy. Our approach handles high-dimensional textual states, scales to large language models, and works with both open-source and proprietary language models. When applied to calls from a large European telecommunications firm, our stopping agent reduces the time spent on failed calls by 54% while preserving nearly all sales reallocating the time saved increases expected sales by up to 37%. Upon examining the linguistic cues that drive salespeople's quitting decisions, we find that they tend to overweight a few salient expressions of consumer disinterest and mispredict call failure risk, suggesting cognitive bounds on their ability to make real-time conversational decisions. Our findings highlight the potential of artificial intelligence algorithms to correct cognitively-bounded human decisions and improve salesforce efficiency.

Observational Learning (OL) is the process by which individuals learn by observing the actions of others. Online platforms increasingly provide statistics that serve as OL information, such as backer counts in crowdfunding, viewership on streaming platforms, and claimed rates in flash sales. While this firm-provided information can facilitate consumer social learning, it is vulnerable to manipulation. This paper identifies a previously undocumented deceptive tactic on Amazon’s Lightning Deals, where sellers artificially inflate the Deal-Claimed Rate (DCR), the real-time percentage of inventory claimed, to mislead consumers. We term this practice 'fake carting.' Analyzing 2.07 million Lightning Deals, we estimate that fake carting occurs in about 1% of all deals, but is heavily concentrated among high-DCR cases. When consumers observe a DCR of 80 percent or higher, there is a 36.5% chance it is manipulated. Holding product, price, and seller constant, we find that fake carting increases the sales effect of Lightning Deals by 23.9%, suggesting consumers are misled into overestimating product value and making suboptimal decisions. Survey evidence shows that consumers place high trust in DCR and remain largely unaware of manipulation. Our findings highlight a significant and underexplored deceptive practice in online markets: the manipulation of OL information.

What types of stories are most persuasive? In this paper, we introduce a new template for categorizing story types based on the specific emotional dynamics of text, or ''emotion sequences''—for example, whether a story begins fearful and ends with sadness, or vice versa. We present this as a new way to capture distinct narrative progressions that is tractable even in short-form media, and then apply this method to analyze the persuasiveness of different story types in online fundraising. Using transformer-based emotion classification tools, we measure the two-part emotion sequences of 14,000 medical fundraising pitches from GoFundMe.org and show that, among other findings, medical fundraising pitches that begin with a sad tone and end on a caring tone are significantly more likely to succeed. We then develop a simple new approach for testing the generalizability of these observational findings by using crowd-sourced, LLM-assisted rewrites to introduce particular emotion sequences to a sample of 40 randomly-selected fundraisers. We show that human-only rewrites generally fail due to skill deficits (and LLM-only rewrites can introduce salient informational changes from originals), but demonstrate that crowd-sourced, LLM-assisted rewriting offers an effective method for testing the out-of-sample application of research results by everyday online users. With this, we establish that pitches rewritten to feature our focal emotion sequences see a significant boost in perceived persuasiveness, even for some sequences associated with lower success in observational data, while placebo rewrites produce null effects. Furthermore, we show that increased identification with the protagonist of the fundraiser is the primary mechanism driving the observed effects.

Can subtle nuances in the way aid is provided negatively impact recipient psychology and behavior? The present research explores the practice of sharing information about and/or a message from the giver to the aid recipient—what we term giver spotlighting—a marketing practice employed by sponsorship charities and community-based aid programs. Six preregistered experiments (N = 3,888), a Kenya-based field study and a series of five online experiments, explore the effect of giver spotlighting on recipients' willingness to recommend the aid organization to others in similar need states and return for more aid if it is needed, hence, examining potential knowledge spread of the aid opportunity and continued aid take-up. Our results suggest that giver spotlighting threatens recipients' self-efficacy and, subsequently, hinders recipients' willingness to recommend the aid program and to return if more help is needed. The present research provides novel insights into how aid organizations can create more effective campaigns that bring in more recipients while also improving existing recipients' experiences when utilizing their aid program.

Visual designs are often used in marketing (e.g., packaging, ads, media covers) to achieve a variety of business outcomes, like improved sales, click-through rates, and brand attitudes. Since designs are complex, unstructured data, it is difficult to determine what features drive their success in a way that is interpretable and managerially actionable. To address this challenge, I develop a novel methodological framework to automatically discover what interpretable features make visual designs in a given domain successful. I first leverage a deep generative text-to-image AI model (by fine-tuning Stable Diffusion 3.5 in my application) that adopts the role of designer and enables visual designs to be described by low-dimensional design representations. Then, I apply a novel adaptation of cutting-edge ''mechanistic interpretability'' methods—specifically ''sparse autoencoders'' typically applied to large language models—to scalably discover a taxonomy of interpretable and managerially relevant features predictive of success from these design representations. Finally, I generate image redesigns by manipulating features of interest to help managers scalably pilot data-driven design changes. I apply this framework to discover how book cover redesigns predict sales on Amazon.com using a unique dataset I collected of over 160,000 books. I discover a diverse set of interpretable features related to illustration, typography, composition, and layout. I then create realistic cover redesigns predicted to improve sales by manipulating those features (e.g., redesigns with lower contrast and less separation of text and graphical elements). In a holdout analysis with a rich set of control variables, including just 30 of these discovered features (out of 9,728) improves variation explained in sales by nearly as much as prices and by more than reviews. Back-of-the-envelope calculations suggest that a large publisher could leverage this subset of features to increase annual revenue by over $9.1 million, reflecting a change in sales equivalent to introducing an 8.5% price discount.

We study teachers' usage of a digital differentiation tool designed for reading comprehension assignments on an educational technology platform and examine the role of teacher usage in shaping the effectiveness of this tool. Using a structural model that incorporates a hidden Markov model and a two-stage framework for teachers' differentiation decisions, we investigate teachers' usage and implementation preference and evaluate the tool’s effectiveness in enhancing student performance and addressing achievement gaps. First, we find that overall usage of the tool is low, with even lower usage in high-poverty schools. Among those who do use the tool, implementation preferences differ by socioeconomic segment: teachers in high-poverty schools are more likely to prioritize higher-achieving students. Second, through counterfactual simulations, we find that while observed actual usage provides only modest benefits, full usage leads to significant improvements in student performance, particularly in low-poverty schools. To address the gap between potential and actual use, we analyze platform-led interventions. Reducing the cost of implementation increases usage but disproportionately benefits low-poverty schools. A combined strategy that includes both cost reduction and targeted training for teachers in high-poverty schools improves teacher usage and student outcomes across socioeconomic segments. These findings highlight the importance of addressing usage barriers in realizing the full benefit of educational technologies.

Peer-to-peer payment platforms such as Venmo have profoundly transformed social and financial interactions, generating rich behavioral, relational, and network data. Yet, understanding the dynamics of user engagement and growth dynamics in non-contractual settings remain challenging. This study proposes a unified framework, the Multi-Stream Temporal Graph Neural Network (MuST-GNN), that integrates transactional patterns, linguistic features, and network structure. Designed as a joint-task model, it predicts two behavioral outcomes critical to platform growth: engagement among existing users and the acquisition of new users through first-time interactions. Using comprehensive data from Venmo's early growth period and evaluated in a live-update setting, the findings are threefold. First, MuST-GNN substantially outperforms both CRM-based and graph-only models, with predictive improvements of over 30% for engagement and nearly 20% for acquisition relative to a baseline that excludes multimodal signals. Second, linguistic signals provide the strongest predictive lift, highlighting the relational and contextual value embedded in transaction notes and emojis. Third, attention analyses show that the model dynamically shifts its reliance on different modalities over time. Network position is more predictive when little behavioral history exists, transaction patterns become more important as user activity grows, and language signals dominate as social ties deepen. This research provides a robust framework for understanding how social connections, financial behaviors, and communication jointly drive consumer engagement and growth on networked platforms.

Consumers are often drawn to offers that promise unlimited access to a product or service (e.g., unlimited monthly mobile plans). Because actual consumption opportunities are typically finite, most explicitly unlimited offers (e.g., ''unlimited minutes per month'') could be reframed as superficially limited (e.g., '' 44,640 minutes per month''). Although explicitly unlimited offers are seen as more subjectively valuable (i.e., attractive), superficially limited offers win out on monetary valuation (i.e., willingness to pay, estimated price). Two processes explain why superficially limited frames—despite imposing superficial constraints—elevate valuation. First, their high discrete usage limits serve as anchors that increase anticipated usage. Second, these limits permit comparisons with other (necessarily smaller) finite offers that are simpler to price. Consumers spontaneously recruit and scale up from these reference prices when assessing a superficially limited offer's monetary value. The extent to which a consumer's interest in or preference for an offer is predicted by subjective versus monetary valuation—and thus which offer frame dominates—depends on how preferences are elicited and what information consumers have access to (e.g., prices). This work moves research on unlimited offers in a qualitatively new direction and illustrates the theoretical and practical importance of distinguishing between subjective and monetary valuation.

Consumers increasingly encounter products and services positioned to support mental wellness, yet little is known about how they perceive these offerings. This research demonstrates that consumers perceive mental wellness offerings to be more luxurious relative to physical wellness offerings and documents a novel lay belief that underlies this perception. Consumers believe the pursuit of mental wellness is a luxury—desirable but non-essential, and a symbol of exclusivity for those with surplus resources. This lay belief drives greater luxury perceptions of mental wellness offerings even when the same product is advertised to support mental rather than physical wellness. These enhanced luxury perceptions of mental wellness offerings, in turn, increase purchase intentions, particularly when consumers are motivated by a desire to consume luxury. Further supporting the role of the lay belief, luxury perceptions of mental wellness offerings arise in health maintenance contexts but not in treatment contexts, where such products are seen as more of a necessity, rendering the lay belief less applicable. The practical implications of these insights for brand and product positioning, product development, and advertising are discussed.

When analyzing a sequence of customer interactions, it is important for firms to understand how these interactions align with key objectives, such as generating qualified customer leads, driving conversion events, or reducing churn. We introduce a transformer-based framework that models customer interactions in a sequence similar to how a sentence is modeled as a sequence of words by Large Language Models. We propose a heterogeneous mixture multi-head self-attention mechanism that captures individual heterogeneity in touchpoint effects. The model identifies self-attention patterns that reflect both population-level trends and the unique relationships between touch points within each customer journey. By assigning varying weights to each attention head, the model accounts for the distinctive aspects of the journey of each user. This results in more accurate predictions, enabling precise targeting and outperforming existing approaches such as hidden Markov models, point process models, and LSTMs. Our empirical application in a multichannel marketing context demonstrates how managers can leverage the model's features to identify high-potential customers for targeting. Extensive simulations further establish the model's superiority over competing approaches. Beyond multichannel marketing, our transformer-based model also has broad applicability in customer journeys across other domains.

Serialized content—such as episodic fiction, TV series, or educational courses—presents consumers with a dynamic choice: pay for immediate access, wait for free access, or quit. These decisions are shaped not only by platform policies but also by the content itself. Understanding how these factors jointly influence consumption is critical for platforms aiming to optimize engagement and monetization. Content, however, is high-dimensional (capturing multiple layers of meaning and narrative structure) posing challenges for traditional estimation methods. We address this challenge by extending adversarial inverse reinforcement learning (AIRL)—a GAN-based estimator that recovers reward and policy functions from observed behavior. Methodologically, we (i) prove identifiability of action-dependent utilities, (ii) incorporate unobserved consumer heterogeneity, and (iii) develop a content representation method that map unstructured text into features predictive of engagement and payment. We apply the model to 15 months of reading data from a serialized fiction platform (24,000 users, 6,000 chapters, 151 books). Estimates reveal two dominant segments—''Pay & Read'' and ''Wait & Read''—that together account for 87% of consumption and 99% of purchases. Validation against field experiments shows that the estimated rewards are stable and disentangled from environment dynamics. Counterfactuals yield three key insights. First, providing content for free can counterintuitively increase purchases within the consumer, by allowing them to move past likely churn points and creating additional purchase opportunities. Second, customized policies based on heterogeneous patience can increase consumption and purchases by 4% and 23%, respectively. Third, leveraging unstructured content to align pricing with narrative peaks can significantly boost monetization. More broadly, we introduce adversarial estimation as a scalable tool for dynamic models with high-dimensional states.