Oded Netzer

As Taboola prepared to launch its GenAI Ad Maker, a tool that lets advertisers create ad images within the platform, it faced a consequential question about whether AI-generated ad images would perform well with real consumers, and under what conditions. Working with Taboola, we construct a quasi-experimental comparison of 4,633 ‘sibling ads,’ meaning AI-generated and human-made images launched by the same advertisers, in identical campaign settings, at the same time. These ads span more than 369 million impressions and 2.5 million clicks. We do not find a detectable average click-through-rate disadvantage for AI-generated images relative to their human-made siblings. However, this average masks an important boundary condition: ads with AI-generated images outperform human-made ads when the AI images do not ‘look like AI.’ We identify key visual features influencing consumers’ perceptions of whether an ad looks AI-generated. While AI generates sharper, clearer, and less blurry images with larger faces, consumers associate these features with human-made ads. On the other hand, in line with consumers’ expectations, aesthetics, lower warmth, and intense color saturation in ads signal AI generation. These findings support Taboola’s roll-out communication, advertiser guidance, and continued refinement of the GenAI Ad Maker.

Debate over synthetic data in marketing research has polarized into claims that LLMs make human respondents obsolete and calls to shun them entirely. We argue that both positions obscure the more useful question: not whether synthetic respondents work, but when. Building on Brand, Israeli, and Ngwe (2026), we make three contributions. First, we distinguish three types of synthetic data: ungrounded LLM responses, segment-level personas, and individual-level digital twins, and map each to the decisions it can support. Second, we develop a taxonomy of four families of accuracy measures and suggest that the wide discrepancies in reported twin accuracy (from near-perfect to near-chance) could largely reflect differences in what is being measured rather than in method quality. Aggregate measures often perform reasonably well even when little information is provided to the LLM and can mask the complete absence of respondent-level differentiation. Third, we introduce the forgotten question problem (a question omitted from a fielded study) as a promising setting for twin-based augmentation to existing studies. We propose an ex-ante answerability diagnostic requiring no ground-truth data: the R² of a random forest predicting twin outputs from twin training data. Across 108 attitude questions from a nationally representative survey (N = 3,063), an R² > 0.7 screen raises mean twin-human individual-level correlation by 15% and reduces the share of poorly answered questions from 25.9% to only 4.3%. Embedding similarity and experienced-researcher judgment provide correlated but weaker screens.

Scientists and practitioners are aggressively moving to deploy digital twins—large language model (LLM)-based models of individuals—across social science and policy research. We conducted 19 preregistered studies with 164 diverse outcomes (e.g., attitudes toward hiring algorithms and intention to share misinformation) and compared human responses with those of their digital twins (trained on each person’s previous answers to more than 500 questions). We establish an empirical benchmark for digital twin performance: Digital twins’ answers are only modestly more accurate than those from the (homogeneous) base LLM and correlate weakly with human responses (average correlation coefficient of 0.20). To guide future development, we document five ways in which digital twins distort human behavior: (i) insufficient individuation, (ii) stereotyping, (iii) representation bias, (iv) ideological biases, and (v) hyper-rationality. We make our full dataset and code public as a standardized testbed. Our results caution against premature deployment while laying the groundwork for the transparent, replicable, and iterative science necessary for responsible deployment of digital twins.

Echo chambers, social environments where like-minded individuals reinforce shared beliefs, have been extensively studied in political and public health contexts, yet their relevance for brands and consumer behavior remains unexplored. This paper explores echo chambers in a brand context and examines how consumers in brand echo chambers, i.e., surrounded by others with similar brand preferences, perceive and behave toward brands. It introduces two complementary measures of brand echo chambers using a large-scale dataset of over 870,000 X followers across 179 brands. The first measure reflects brand embeddedness among the focal social circle, while the second captures brand exclusivity within the same circle by identifying connections tied solely to the brand. We find that consumers in stronger brand echo chambers discuss brands more frequently and favorably, mention competitors less often, and exhibit stronger emotional engagement when discussing brands. These insights show that brands can leverage readily available social media data to derive echo-chamber metrics and track consumer networks and their relative exposure to the focal brand and competitors. These metrics can also serve as scalable segmentation and targeting signals by helping managers identify consumers whose network structures are associated with distinct patterns of brand and competitor related behavior.

Our faces are said to be windows into the soul. But can they also reflect who we are as consumers? Can facial images predict brand preferences? To answer these questions, we analyze a unique dataset of over 100,000 single-face Twitter profile pictures linked with brand followership data for 444 brands across categories and brand personality metrics. Using advanced machine learning for automated face analysis, we demonstrate that consumers' social media profile faces can reveal their preferences between rival brands (study 1). We further show that consumers who follow brands with similar brand personality traits (e.g., glamorous, rugged) tend to look alike (study 2). Finally, we identify facial appearance characteristics associated with brand personality traits; for example, followers of "smart" brands like Barnes & Noble are likely to wear reading glasses, while "edgy" brands like Diesel attract bearded men (study 3). Together, our findings show that brands can extract actionable insights from consumers' facial images, both to address the "cold start" problem by predicting preferences before any consumption actions are observed and to better understand the brand's identity through the aggregated traits of its followers. This work is among the first to demonstrate large-scale associations between faces and brand preferences.

An idea is more than a simple collection of words or ingredients that make up the idea. What makes an idea original or appealing is how these elements are combined in the context in which they appear. This research leverages representation learning methods, specifically word embeddings, to measure internal coherence among the components of a creative idea and capture the relationship between its coherence and idea evaluation or success. Using a large-scale online recipe dataset with over 57K recipes, as well as an ideation dataset of 1,057 ideas, the authors investigate how the fit among idea components relates to downstream measures such as popularity and evaluation. The results consistently show that internal coherence of an idea promotes its popularity, trial, and post-trial evaluation. Counter to prior research on creativity, which suggests that creativity is mostly associated with positive outcomes, the study finds that ideas with unique ingredients have lower popularity or trial, but higher ratings given trial in the context of food recipes. Based on these findings, the authors develop a generative recipe tool that suggests recipe improvements by adding, removing, or substituting ingredients ( http://recipecreativity.com/ ).

From advertisements and social media to sales pitches and customer service calls, marketplace interactions often involve a mix of linguistic (i.e., words), auditory (e.g., sounds), and visual (e.g., images) content. But while these different modalities, or streams of information, frequently interact, most existing research (and practice) analyzes only one modality at a time. Further, while some work has started to consider multiple modalities, it often treats them as independent, failing to consider the interrelationships between them. To fill this gap, the current work introduces the concept of intermodality, proposes an integrative theoretical framework regarding different ways modalities relate to, and interact with, each other, and outlines a unifying approach for modeling multimodal communications. This includes (i) systematically aligning levels of representation across modalities, (ii) considering how modalities can be combined (i.e., through separate or joint extraction), and (iii) functional forms and considerations for modeling different types of intermodal relationships. Overall, this work provides insight into how linguistic, audio, and visual content interact, sheds light on how to study intermodality, and offers researchers and practitioners concrete guidance on how to better identify key drivers of relevant outcomes.

Major life transitions, such as pregnancy and childbirth, reshape lifestyles and purchasing priorities, yet causal evidence on how consumers reallocate spending across product categories remains limited. We quantify the effects of first-time parenthood by linking a large-scale transactional panel to verified birth records. To identify causal effects, we implement a difference-indifferences design augmented with causal forests, enabling flexible comparisons between households entering parenthood and carefully matched controls. We uncover a pronounced and dynamic behavioral trajectory. Spending and trip frequency decline early in pregnancy, rebound late in gestation, and increase sharply after childbirth. Post-childbirth, supermarket spending and ticket size rise by approximately 20-25%, accompanied by a 10% increase in promotional reliance and a systematic reallocation away from discretionary categories (e.g., dining, travel, beauty, pets) toward health-and family-oriented goods (e.g., fresh foods, yogurt, pharmacy items, baby products). Reduced leisure and personal-care spending is offset by higher expenditures on essential household categories, indicating lifestyle adaptation rather than uniform budget tightening. Heterogeneity is substantial: effects are larger among younger parents and households with higher baseline spending. These results identify childbirth as a major inflection point in consumer behavior and offer actionable guidance for retailers.