From Attribution to Action: Causal Incrementality and Bandit-Based Optimization for Omnichannel Customer Acquisition in Retail Media Networks

Abstract

Retail Media Networks (RMNs) have become a central marketplace for performance marketing, yet the measurement and optimization loop remains fragile when decisions are driven by correlational signals such as last-click attribution or conventional Return on Ad Spend (ROAS). Two problems appear repeatedly in practice: (i) attribution systems that mis-allocate conversion credit across multi-touch journeys, and (ii) bidding and budgeting policies that over-invest in placements that look profitable only because organic demand is misattributed to ads. This research article proposes an integrated framework that unifies machine learning attribution, causal incrementality measurement (incremental ROAS, iROAS), and sequential decision-making via multi-armed bandits (MABs). The contribution is a closed-loop architecture in which a causal estimator produces uncertainty-aware incremental value signals at campaign and placement levels, and a contextual bandit policy allocates budget to maximize cumulative incremental profit under realistic constraints (delayed conversions, inventory coupling, advertiser objectives, and privacy). The article specifies a deployable pipeline, catalogs experimental and quasi-experimental designs suitable for RMNs and omnichannel acquisition, and provides a simulation study illustrating how iROAS-aligned bandits can outperform ROAS-driven heuristics when organic baseline and confounding are present. The discussion emphasizes validity threats (interference, spillover, novelty, and measurement leakage), governance requirements (auditability, explainability, and privacy-by-design), and a research agenda for real-time causal estimation and multi-objective bandit learning in high-dimensional retail settings.