Marketing mix modelling, causal inference, and the measurement problems underneath both — written from client work, not from theory alone.
Almost every MMM tutorial teaches adstock as a recursive carryover formula. Google Meridian does not implement that formula. The difference changes what your decay rate means, how half-life should be computed, and whether your Hill saturation parameters are interpretable at all.
Language models will answer any survey question you put to them. The question is whether those answers correlate with what a real panel would have said, and where the correlation breaks.
Print carried well in the model, a 1.9 week half-life and a respectable adstock multiplier. It still did not pay for itself in eleven of fourteen regions.