What Is Robyn?
Robyn is the open-source marketing mix modeling framework developed and maintained by Meta's marketing science team. It is implemented in R, uses a hybrid Bayesian-genetic-algorithm approach to model fitting, and is designed for production marketing measurement at brand scale. Released publicly in 2021, Robyn has become one of the two dominant open-source MMM frameworks (alongside Google's LightweightMMM).
Why Robyn Matters
Robyn made production-grade MMM accessible to teams without dedicated econometrics staff. The framework handles adstock and saturation transforms automatically, runs hyperparameter optimization via genetic algorithm, and produces Pareto-optimal model candidates ranked by fit and stability. The opinionated workflow reduces the analyst's scope of choices to those that matter most.
For brands measuring AI search, Robyn is one of the standard frameworks for adding the AI variable. The hybrid Bayesian approach handles informative priors on adstock and saturation, which is critical for short-history channels like AI visibility.
How Robyn Works
The user provides a CSV of weekly data: spend by channel, exposure or impression metrics, the outcome variable (typically revenue or conversions), and controls (seasonality, macroeconomic variables, holidays). Robyn runs a genetic algorithm to search over adstock and saturation hyperparameters, fits a regularized regression at each candidate, and returns a Pareto frontier of models trading off fit (NRMSE) against decomposition stability (DECOMP.RSSD).
The analyst selects a model from the Pareto frontier based on judgment about the trade-off and inspects budget allocation recommendations derived from the response curves. Calibration against incrementality tests can be enforced through the calibration_input argument.
In Practice
Robyn workflows in production: weekly data refresh into the input CSV, monthly full refit with manual Pareto selection, quarterly stakeholder review, annual major revisit of model spec and channel list. The framework handles most of the operational lift; the analyst spends time on prior selection, validation, and stakeholder communication.
For the AI search variable, the conventional pattern is to enter LLM share of voice as a paid_media_var with a constant placeholder in the corresponding spend column. Adstock and saturation priors should be informative (geometric half-life two to four weeks, Hill saturation with half-saturation in the middle of observed range) rather than uninformative defaults.
How Presenc AI Helps
Presenc AI exports weekly AI visibility data in the exact format Robyn expects. The CSV column names, week format, and adstock-ready structure plug into Robyn's input pipeline without further transformation. Default category-level priors for the AI variable are published alongside the export so the variable enters Robyn with informative starting points.