How to evaluate and choose a marketing mix model?

[KatatoniaSh]

Assuming that we have an MMM problem where the ground truth is unknown, I am wondering how we can choose the best model. Suppose that we have a dataset. To train the model, we can set different priors, different adstock and hill functions, and so on. Given those, how can we rank the models? Some metrics can be R2 on the train and test data, the number of converges, Bayesian Posterior Predictive P-value, R-hat. It is possible that we have a model with different settings which results in different ROIs, contributions.