A diagram of variables and their conditional probability relationships, used to model how causes relate to effects in a system. Agencies encounter Bayesian networks most often in attribution modeling, where the question is: given what we observed, what actually caused this outcome?
Also known as belief network, Bayes net, probabilistic graphical model
A Bayesian network represents variables as nodes and their conditional dependencies as directed edges. The structure encodes assumptions about causality: an arrow from A to B means A has a direct probabilistic influence on B. Together, the graph and its probability tables define a complete model of how outcomes in the system are generated.
Once a network is defined and calibrated with data, it can answer queries in any direction. Given that a conversion happened, what is the probability each channel contributed? Given that a channel received a higher budget, what change in conversion probability would we expect? The same model answers attribution questions and scenario planning questions using the same structure.
Unlike purely correlational models, Bayesian networks require the modeler to specify a causal structure. That constraint is also a feature: the assumptions are explicit, auditable, and correctable when domain knowledge changes.
Attribution is the central measurement problem in agency work, and Bayesian networks are one of the more principled tools for solving it. They force explicit reasoning about how channels interact rather than treating touchpoints as independent variables in a regression.
They separate correlation from causation. Standard last-click or data-driven attribution models measure which touchpoints correlate with conversion. Bayesian networks can be structured to model which touchpoints cause conversion by incorporating causal assumptions from domain knowledge. That distinction changes how budget decisions should be made.
Scenario planning becomes tractable. Because the model represents the full probability structure of the system, you can query it prospectively: if we shift 20% of display budget to paid social, what happens to estimated conversion probability? That is more useful to a planning conversation than a historical correlation table.
They require expertise to build correctly. A Bayesian network is only as good as the causal assumptions encoded in its structure. Agencies evaluating vendors using this approach should ask how the network structure was defined, whether domain experts were involved, and how the structure is validated and updated over time.
A media analytics team builds a Bayesian network to model the client’s conversion funnel, encoding prior knowledge that TV drives initial awareness, search captures in-market intent, and display retargeting closes the loop. When the model is calibrated on twelve months of campaign data, it produces channel credit estimates that differ materially from the platform’s default attribution model, which had credited search with a disproportionate share of conversions. The Bayesian model reveals that much of the search volume was already-converted intent driven by TV exposure, not independent search-driven demand. The revised attribution changes the budget recommendation significantly.
The generative AI foundations module of the workshop covers how today’s models work, what they can and can’t do, and how to choose between them.