advertising and marketing

Industrial digital advertising lacks a systematic, engineer-oriented methodology for budget pacing—i.e., uniform spending over time. Method: This paper introduces the first structured, production-ready pacing algorithm taxonomy and tuning paradigm, unifying deterministic, stochastic, and learning-based strategies. It integrates control-theoretic techniques (PID/MPC), online optimization, probabilistic modeling, and real-time feedback to jointly optimize spend rate, ROI, and smoothness. Contribution/Results: The framework significantly improves budget attainment and delivery stability: on major DSP platforms, it reduces average spend deviation by 37% and over-spend rate by 52%. It fills a critical gap in the real-time bidding pipeline by establishing a reusable, scalable infrastructure for budget control—bridging theory and large-scale ad systems engineering.

Online behavioral advertising (OBA) suffers from conceptual ambiguity and fragmented empirical evidence, hindering theoretical advancement and practical guidance. Method: Through a systematic literature review and interdisciplinary theoretical integration, this study develops the first precise definition of OBA and a unified theoretical framework that jointly models advertiser-controllable variables (e.g., data transparency, personalization intensity) and consumer-level characteristics (e.g., privacy concern, digital literacy). Contribution/Results: The study maps the distribution of existing empirical findings and identifies three critical research gaps: (1) mechanistic black boxes, (2) contextual boundary conditions, and (3) cross-cultural variations. It proposes empirically testable future research directions and policy design principles that balance advertising efficacy with privacy protection. This work provides an integrative scholarly foundation for advancing OBA theory, optimizing precision advertising practices, and informing data governance policy.

Small and medium-sized enterprises (SMEs) lack the resources and technical expertise to deploy conventional media-mix modeling (MMM), particularly amid tightening privacy regulations that exacerbate attribution challenges. Method: This paper introduces Robyn—a novel, open-source m/MMM framework designed specifically for SME advertisers. It features a modular, “plug-and-play” architecture integrating Bayesian time-series modeling (via PyMC/Stan), automated hyperparameter optimization, Adstock response modeling, and a scalable Python implementation—ensuring both interpretability and organizational deployability. Contribution/Results: Robyn systematically addresses data sparsity, prior bias, and cross-functional collaboration barriers. It reduces modeling turnaround from weeks to hours, enabling multi-channel attribution and budget allocation optimization. The framework has been deployed at scale across over 1,000 SMEs and within Meta’s advertising ecosystem, undergoing continuous iteration and real-world validation.

Digital advertising platforms (e.g., Meta Ads) suffer from algorithmic opacity, hindering advertisers’ understanding of audience targeting, pricing mechanisms, and ad relevance—thereby impeding data-driven decision-making. To address this, we propose SODA: the first explainable advertising analytics framework integrating multimodal text-image models with large language models (LLMs). Our method introduces a natural-language–based interactive explanation interface tailored for non-technical marketing professionals, enabling automated competitive ad summarization, attribution analysis, and click-through rate (CTR) prediction. By synergistically combining eXplainable AI (XAI) techniques with natural language generation and understanding, SODA enhances predictive accuracy while delivering actionable, trustworthy AI-assisted insights. Evaluated in real-world deployment scenarios, SODA significantly improves interpretability without compromising performance, empowering marketers to make informed, auditable decisions grounded in transparent model reasoning.

This study investigates whether privacy-enhancing technologies (PETs) meaningfully reduce consumers’ perceived privacy violation (PPV) from online advertising. Method: Drawing on the dual-privacy theoretical framework, we conducted an online experiment with U.S. users to compare PPV across distinct data practices—behavioral targeting, on-device processing, cohort-based targeting, and context-based targeting—while varying tracking and personalization mechanisms. Contribution/Results: We provide the first empirical evidence of substantial consumer misperceptions regarding technically defined privacy protections: on-device processing yields only marginal PPV reduction; cohort-based targeting confers no advantage over individual-level behavioral targeting; and tracker-free contextual targeting significantly lowers PPV—approaching levels observed with no ads at all. Notably, users exhibit no statistically significant preference difference between tracker-free non-targeted ads and ad-free browsing. These findings challenge prevailing assumptions about PET efficacy and offer critical empirical grounding for privacy-by-design principles and regulatory policy.

This study addresses the critical role of the first three seconds—the “hook phase”—of video advertisements in capturing user attention and driving conversion, a challenge compounded by the multimodal nature of this content that limits traditional analytical approaches. To overcome this, we propose an integrated framework combining multimodal large language models (MLLMs) with topic modeling. Our method employs a frame sampling strategy that balances representativeness and diversity to extract visual and acoustic features, leverages BERTopic to abstract themes from MLLM-generated descriptions, and incorporates audio attributes alongside ad targeting metadata. Evaluated on large-scale real-world data from a social media platform, the approach significantly enhances prediction accuracy for key performance indicators such as return on ad spend, marking the first effective integration of MLLMs and topic modeling for hook-phase analysis.

This work addresses a critical limitation in existing automated bidding algorithms, which often conflate ad impression or click revenue with true advertiser value and neglect the counterfactual benefit of organic search traffic when auctions are lost, leading to inefficient budget allocation. The paper introduces a novel approach that models ad value as the marginal causal effect between winning and losing an auction. By leveraging payment information revealed through the second-price auction mechanism, the authors design an online bidding strategy that achieves theoretically optimal regret bounds across multiple feedback models. This method substantially outperforms counterparts in first-price auction settings and significantly enhances both advertising efficiency and return on investment.

Existing ad generation methods rely on multiple models and only capture average click preferences, lacking cross-modal awareness and personalization capabilities. This work proposes Uni-AdGen, a unified autoregressive model that achieves end-to-end joint generation of image-text advertisements for the first time. By integrating foreground-aware modeling, instruction fine-tuning, and a coarse-to-fine multimodal user preference understanding mechanism, Uni-AdGen accurately captures individual interests from users’ click histories to guide ad generation. The contributions include constructing PAd1M—the first large-scale personalized ad dataset—introducing the Product Background Similarity (PBS) evaluation metric, and demonstrating significant performance gains over existing approaches in both general and personalized ad generation tasks, yielding results with higher realism and user relevance.

This work addresses the lack of efficient bidding algorithms in brand advertising auctions that can fully exploit the stability of user engagement and the rapid feedback inherent in such campaigns. We propose a lightweight model predictive control (MPC) framework that, for the first time, integrates online isotonic regression with MPC to construct monotonic bid–spend and bid–conversion relationships in real time from streaming data. This approach enables accurate, low-overhead real-time bidding without relying on complex models. Extensive simulations demonstrate that our method significantly outperforms existing baselines, achieving superior performance in both cost control and spending efficiency. Moreover, the framework exhibits high scalability and strong practical viability for real-world deployment.