GEO Glossary

Priors in Marketing Mix Modeling

Priors in Bayesian MMM encode the analyst's beliefs about plausible parameter values before observing data. Why they matter, how to set them, and the AI variable specifics.

By Ramanath, CTO & Co-Founder at Presenc AI · Last updated: April 23, 2026

What Are Priors?

In Bayesian MMM, a prior is the probability distribution the modeler assigns to a parameter before observing data. The Bayesian inference updates the prior with the data to produce the posterior. The choice of prior matters most when data is sparse; with abundant data, the posterior is dominated by the likelihood and the prior's influence fades.

Why Priors Matter

Marketing data is almost always sparse relative to the parameter count, especially for newer channels with short history. Priors are what makes Bayesian MMM tractable in this setting; they encode the analyst's domain knowledge and constrain the inference to plausible parameter ranges.

For AI search specifically, the AI variable typically has 52 to 104 weeks of history, which is short relative to traditional channels. Informative priors on adstock half-life and saturation curves are the difference between an identifiable AI coefficient and an uninformative noisy estimate.

Types of Priors

Uninformative priors: Wide distributions that express minimal prior knowledge. Used when no domain knowledge is available. Risk: with sparse data, the model wanders to extreme parameter values that fit the noise.

Weakly informative priors: Distributions that bound parameter values to plausible ranges without strong commitment within the range. The standard recommendation for most MMM parameters.

Informative priors: Distributions that express strong prior knowledge, often derived from category benchmarks or previous model results. Used for parameters where domain knowledge is reliable, especially adstock half-life and saturation curve shape.

Calibration priors: Distributions derived from external experimental evidence (lift tests). Used to anchor MMM coefficients to causal ground truth.

In Practice

The right prior choice depends on data history and external knowledge. For new MMMs with short data history, informative priors derived from category benchmarks compensate for the data limitation. For mature MMMs with abundant history, weakly informative priors are usually sufficient because the data dominates.

The single biggest prior selection mistake for AI visibility variables is borrowing paid search priors, which produce too-short adstock and too-aggressive saturation for the AI channel's actual behavior.

How Presenc AI Helps

Presenc AI publishes category-specific prior guidance for the AI visibility variable derived from cross-customer analysis. The guidance gives modelers a defensible starting point rather than a vague default, which improves first-refit quality and reduces iteration time.

Frequently Asked Questions

Sensitivity analysis. Refit the model with slightly different priors and check whether key conclusions change. If they do not, the data is dominating and the priors are fine. If they do, the data is sparse and the priors are doing the work; tighten them based on stronger external knowledge or wait for more data.
Yes. Over-constraining a parameter forces the inference toward the prior regardless of what the data says. Symptoms include very tight posteriors, poor holdout fit, and low sensitivity of conclusions to changes in the data. The fix is to widen the prior toward the weakly informative range.
Three sources: category benchmarks (vendor-published or peer-shared), previous model refits (use last quarter's posterior as this quarter's prior), and external experimental evidence (lift test results). The combination of these sources usually produces well-anchored priors for established channels.
Adstock: geometric with half-life prior centered at two to four weeks for most categories. Saturation: Hill function with half-saturation centered in the middle of the observed AI visibility range. Coefficient: weakly informative normal centered at zero with moderate variance. These produce identifiable inference even with short data history.

Track Your AI Visibility

See how your brand appears across ChatGPT, Claude, Perplexity, and other AI platforms. Start monitoring today.