Coauthors: Shuang Zheng (Renmin University of China), Xin Ye (Arizona State University), Liang Shen (University of Connecticut)

Abstract: Generative search is an emerging search paradigm that integrates Generative AI (GenAI) into traditional search engines. Increasingly adopted by platforms worldwide, this format presents consumers with a GenAI-generated response to their query prior to displaying conventional search results. Unlike keyword-based search, which relies on consumers’ ability to translate underlying needs into effective keyword queries, generative search allows consumers to directly express their underlying intentions in natural language and receive responses that guide subsequent exploration. This shift moves the consumer–search engine interaction upstream in the search process, to the stage of original problem formulation. By bringing this upper-funnel interaction, namely, the mapping from problem formulation to keyword articulation, onto the platform, generative search makes a previously unobservable stage of consumer search behavior empirically observable. However, it remains empirically unknown whether generative search can increase consumer purchases, and how consumer search behavior changes when search can begin from expressed intent rather than keywords. Using a large-scale field experiment conducted on Meituan, a leading Chinese technology platform, this paper provides the first empirical evidence on the effectiveness of generative search. We find that generative search significantly increases consumer purchases. Further analysis reveals that generative search facilitates more effective and diverse keyword queries, reduces exploratory browsing and clicking while concentrating evaluation within relevant categories and merchants, which helps explain the observed increase in purchases. These findings provide empirical insights for future consumer search models that allow search to originate at the intention-expressing stage.

Coauthors: Jieteng Chen, Jiwoong Shin 

Abstract: Generative AI has transformed content creation, enhancing efficiency and scalability across media platforms. However, it also introduces substantial risks, particularly the spread of misinformation that can undermine consumer trust and platform credibility. In response, platforms deploy detection algorithms to distinguish AI-generated from human-created content, but these systems face inherent trade-offs: aggressive detection lowers false negatives (failing to detect AI-generated content) but raises false positives (misclassifying human-created content), discouraging truthful creators. Conversely, conservative detection protects creators but weakens the informational value of labels, eroding consumer trust. We develop a model in which a platform sets the detection threshold, consumers infer credibility from labels when deciding whether to engage, and creators choose whether to adopt AI and how much effort to exert to create content. A central insight is that equilibrium structure shifts across regimes as the threshold changes. At low thresholds, consumers trust human labels and partially engage with AI-labeled content, disciplining AI misuse and boosting engagement. At high thresholds, this inference breaks down, AI adoption rises, and both trust and engagement collapse. Thus, the platform’s optimal detection strategy balances these forces, choosing a threshold that preserves label credibility while aligning creator incentives with consumer trust. Our analysis shows how detection policy shapes content creation, consumer inference, and overall welfare in two-sided content markets.

Coauthors:  Wenbo Wang (Hong Kong University of Science and Technology), Zijun (June) Shi (Hong Kong University of Science and Technology).

Abstract: Social media platforms often confine users to information silos, creating narrow loops of reinforced interests that limit exposure to diverse perspectives and exacerbate societal divisions. This study introduces a metric to systematically quantify this phenomenon through two dimensions: intrapersonal variety, reflecting the breadth of individual topic engagement, and interpersonal difference, capturing the divergence of user consumption from population norms. Focusing on TikTok, we combine demographic surveys with a field experiment wherein participants record their browsing behavior. We analyze video topics to compute this information silo index, identify silo formation patterns, and correlate them with user characteristics. Furthermore, we test experimental interventions by providing users with tailored feedback on their content consumption habits. Our results demonstrate the intensity of information silos, assess user awareness, and establish the efficacy of feedback mechanisms in disrupting these patterns to foster broader content engagement.

Abstract: A large-scale randomized field experiment on a two-sided platform shows that providing a recall option in consumer search significantly increases the probability of a successful match and, counterintuitively, leads to higher final transaction prices. These findings challenge the well-established result in the consumer search literature that recall has no impact on search outcomes in an infinite-horizon single-agent setting. We develop a theoretical model of consumer search in which consumers are uncertain about the price distribution and update their beliefs in a Bayesian manner. The model predicts that recall can increase both match rates and accepted prices as consumers re-evaluate past offers in light of updated beliefs. An incentive-aligned online laboratory experiment further corroborates these findings and the underlying learning mechanism in a controlled setting, showing that recall reduces reliance on outside options.

Abstract: Why do people report probabilities like 0.2, 0.5, or 0.8 rather than 0.317, for example? Survey evidence documents widespread heaping at round numbers in subjective probability elicitation, yet existing work treats this as measurement noise or satisficing rather than the outcome of rational optimization. We argue it is neither noise nor laziness, but the outcome of rational optimization under two kinds of cost: forming a precise internal belief is cognitively demanding, and so is expressing it with many decimal places. We model a sender who first refines her belief through hierarchical binary partitioning and then rounds it to a decimal grid, with convex costs on both dimensions. The optimal precision pair is finite and well-defined, and reported probabilities heap on round decimals as a natural consequence. A further result is that refinement has sharply diminishing returns once internal resolution becomes finer than what the decimal grid can express, thus, at that point, thinking harder simply cannot improve what one says. Introducing response time as the resource governing both dimensions ties deliberation directly to reporting precision. The proposed framework bridges rational inattention and bounded communication.

Abstract: The Russian advertising market underwent a dramatic structural transformation following the exit of many multinational brands in 2022. This talk examines how the composition of advertisers, categories, and media investments changed in response to this shock, drawing on large-scale industry data and market analysis. The departure of international brands led to a sharp decline in the number of foreign advertisers and a substantial reallocation of advertising budgets across domestic companies, localized brands, and new entrants from Asia. 

I discuss how firms adapted their marketing strategies under these conditions, including brand localization, rebranding, indirect media placement through digital channels, and shifts in category-level competition. The talk also examines the evolving structure of media demand, the implications for television and digital advertising markets, and scenarios for the potential return of multinational brands. 

These developments provide a rare real-world setting for understanding how large exogenous shocks reshape advertising markets, brand competition, and media allocation decisions.

Coauthors:  Argyro Katsifou and Ralf W. Seifert (École Polytechnique Fédérale de Lausanne)

Abstract: Assortment planning deserves much attention from practitioners and academics due to its direct impact on retailers’ commercial success. In this paper we focus on the increasingly popular retail practice to use combined product assortments with both “standard” and more fashionable and short-lived “variable” products for building up store traffic of “loyal” and “non-loyal” heterogeneous customers and enlarging the sales due to the potential cross-selling effect. Addressing the assortment planning as a bilevel optimization problem, we focus on decision-dependent uncertainties: the retailer’s binary decision about product inclusion influences the distribution of the product’s demand. Furthermore, our model accounts for customers’ optimal purchase quantities, which depend on budget constraints limiting the basket that a customer is able to purchase.

In this work, we propose iterative heuristics using optimal quantization of demand and customer’s budget distributions to define the total assortment and the inventory level per product. These heuristics provide lower bounds on the optimal value. We conduct a comparison to other existing lower bounds and we formulate upper bounds via linear (LP) and semidefinite (SDP) relaxations for the performance evaluation of the heuristics and for an efficient numerical solution in high-dimensional cases. For managerial insights, we compare the proposed approach with three assortment planning strategies: (1) the retailer does not carry variable products; (2) the retailer ignores the cross-selling effect; and (3) the maximum space allocated to each product is fixed. Our results suggest that variable assortment boosts the retailer’s profits if the cross-selling effect is not neglected in the decision about products’ quantities.

Abstract: Influencer marketing has experienced rapid growth in recent years. In the video game industry alone, approximately 40,000 YouTube influencers are actively promoting various games. Given that multiple influencers are often engaged simultaneously in typical promotional campaigns, and the collective impact of these promotions (e.g., sales outcomes) is only observed at an aggregate level, understanding and optimizing the selection of influencers is crucial. However, this task is challenging due to the need to accurately predict how campaign outcomes respond to the promotional efforts of participating influencers. Two primary challenges arise in this context. First, this response model is high-dimensional and nonlinear, characterized by numerous main effects and interactions due to the large number of influencers involved. Second, the number of observed campaign outcomes is often limited, particularly for short-lived products like video games, complicating reliable estimation. Existing methods face a trade-off: some effectively manage high dimensionality but struggle to capture nonlinear relationships (e.g., LASSO), while others provide flexible nonlinear modeling but perform poorly with a small number of observed outcomes (e.g., many machine-learning approaches). In this paper, we propose a Bayesian Additive Regression Tree with structural priors (BARTwSP) approach to simultaneously address these challenges. BARTwSP integrates tree-based machine learning techniques, of which the additive aggregation of trees is found to well approximate any complex nonlinear relationships, with the Bayesian framework which imposes a strong prior on the tree depth to ensure a shallow tree structure, thereby reducing overfitting and enhancing inference with small samples. By incorporating structural prior knowledge, we can further regulate variable splitting probabilities, addressing the ``small N large P" problem. Using a Monte Carlo simulation that replicates the real-world challenges, we demonstrate that the BARTwSP model predicts outcomes more accurately with at least 18% reduction in MSE than the benchmark models, including LASSO and other machine-learning approaches in high-dimensional (large P) settings with limited observations (small N). The results further indicate that this advantage is linked to the BARTwSP model's superior ability to select relevant variables used in the data generation process characterized by prominent between-group and within-group sparsity. We illustrate the benefits of our proposed model using data from a leading global video game developer. We generate structural prior knowledge using large language models (LLMs). Our findings show that: first, the BARTwSP model predicts daily active users (DAU) more accurately with the best MSE in this real-world setting. Second, leveraging the estimated response function, we optimize the selection of influencers subject to influencer-specific costs or an overall campaign budget. The proposed model can increase the DAU by 21% under the current budget constraints, and it selects only 48% of the original influencers but requests some of them to post more frequently. Lastly, using an endogeneity-corrected version of the model, we quantify the effects of individual influencers to better understand the factors driving the optimization results.

Coauthors: Jincheng Fang and Shijie Han (Hong Kong University of Science and Technology)

Abstract: Video content such as short-form social media clips and livestream commerce has become central to contemporary digital marketing, serving not just as visual media but as strategic vehicles for narratives, persuasion, and product demonstration. Standard Large Multimodal Models (LMMs) are limited in capturing these marketing-relevant elements because prompt-based approaches often miss persuasive logic and produce factually weak outputs. To address this gap, we develop a Reflexive Agentic Measurement Framework tailored to marketing video analytics. The framework involves nine AI agents in two phases. The first phase constructs agent behavior, involving iteratively discovering latent narratives and persuasive cues and then autonomously generating context-aware coding prompts. The second phase generates and verifies content analysis: decomposing marketing videos into discrete propositions tied to key marketing elements and validating each element against the raw footage to ensure factual reliability. Applying this framework to 5,000 livestream commerce videos, we show that it delivers richer and more accurate descriptions of marketing strategies—such as dynamic screen layouts, product effect demonstrations, and narrative framing—than standard prompt-based methods and human-crafted baselines. Our approach also achieves this at a low marginal cost per video minute ($0.005 per video minute), offering a scalable methodology for automated marketing video coding. By enabling systematic measurement of previously elusive strategic elements in marketing videos, this research advances measurement tools for digital marketing and supports data-driven investigations into the drivers of video marketing effectiveness.