The Marketing Context Graph: A Governed Substrate for Causal Marketing Measurement and Auditable Decision Provenance - Causal inference as a governed layer on a bitemporal context graph

Abstract

Marketing decisions increasingly depend on two capabilities that have evolved separately: causal marketing measurement (distinguishing real advertising effects from the correlational patterns in marketing data, the domain of Causal AI) and governed, graph-based context for the AI agents now trusted to act on those effects, the domain of Agentic AI. Both are limited by the same architectural gap. Problem 1: measurement is correlational by construction (not only in media-mix models but in touch-based attribution and observational methods generally), so a channel’s direct, indirect, and total effects are hard to separate and to calibrate against experiments. Problem 2: the agents acting on those effects lack persistent, governed, historically accurate context, and to answer questions such as which campaigns to scale or how to optimize the mix they must reason over incremental effects, uncertainty, evidence strength, applicable populations, and decision lineage, not language. We introduce the Marketing Context Graph (MCG), a bi-temporal, governed property graph, formally a tuple G = (E, R, γE, γR), in which campaign entities, spend vectors, conversion outcomes, and external macro factors are connected by typed, directional, causally identified edges carrying bi-temporal attributes and reified decision traces. We give it a precise set-theoretic semantics, define a point-in-time snapshot operator Σ(tv, tr), and prove it is free of look-ahead contamination. Crucially, the MCG is a governed substrate, not a new estimator: the measurement methods applied on it are classical: mediation / path analysis, inverse-probability weighting, and transport. We (i) inject an experiment-identified halo edge as an informative structural prior into a Gaussian-conjugate Bayesian MMM, showing it reduces posterior variance along the collinear directions (Remark 1) while separating that variance claim from identification (Proposition 2); (ii) compose each channel’s total effect by traversing the graph, so budget decisions improve as cross-channel mediation grows and gain nothing when no structure exists; (iii) reconstruct complete, auditable decision provenance; and (iv) transport an identified edge across populations exactly when the graph’s selection diagram licenses it, and refuse when it does not. Our evaluation is decision-centric rather than fit-centric. Across 8,000 simulated budget decisions the graph cuts budget-decision regret by up to 43% as cross-channel mediation grows (a paired within-replication reduction); look-ahead-free backtesting exposes 70% over-investment that a flat store conceals; the graph answers the provenance queries that require multi-hop, policy, temporal, and audit structure a flat fact/vector store cannot represent; and governed transport stays low-error as populations diverge, where naive transfer’s error grows linearly with divergence. Evidence is synthetic by design, isolating each mechanism against a known oracle; all results are reproducible from released code.