explōrātiō runs two of the market's leading MMM methodologies — Meta's Robyn and Google's Meridian — on your same data base. When both agree, the mix recommendation lands on the table with double weight; when they disagree, you know exactly where and why — without the classic risk of defending a number only one methodology could sustain. The result: a budget decision that survives the hard question in the meeting and the auditor who comes after.
Input → dual modeling → integrated readout
What goes into the study, what we run in parallel in the middle, and what you take away — with the project timeline below.
From contract to close-out
You send historical spend by channel¹ and the KPI you want to explain (revenue, conversion, leads), in whatever format you have today — platform exports, CSV, or spreadsheet. We return a readout of how much each channel is actually pulling the outcome, what changes if you move budget between them, and which decisions both methodologies agree on — and which still call for caution.
¹ assuming bases without classification errors; otherwise we'll assess the need to contract Data Clean
From the executive committee to the revenue team — when the question is mix, budget, and methodological defense in a meeting where someone will ask to rerun the numbers, not just a weekly performance report.
When the next committee meeting asks for a defense of the media plan, not a gut feeling. You need to arrive with a justified mix, budget scenarios ready, and methodological backing that survives the hard question — without having run the study in the dark or delegated to a tool nobody understands.
When the data team wants to open the hood and audit what was done. explōrātiō delivers an open specification of the dual study (Robyn + Meridian), CSV exports of estimates and decompositions by methodology, and technical documentation your team can read, rerun, and integrate — no black box, no closed platform.
When the end client asks for MMM, but the agency cannot sustain a study of this scale internally. explōrātiō steps in as the analytical arm — white label, co-signed, or behind the scenes — running the study end to end, including readout workshops and scope customization per the briefing.
One project design covers everything: base preparation, training both methodologies on the same base, cross-checking outputs with explicit readout of convergence and divergence, budget scenarios for the next planning round, and readout workshops so your team can apply the result — without redigesting the technical report for the next meeting.
Four readout fronts in the final report. Each one exists to answer a concrete question that will come up in the planning meeting, in committee, or in conversation with the data team — not to fill slides.
Robyn and Meridian trained on the same base. The report explores correlations and agreement between each channel's contributions — and the recommendation gains weight when both point the same way. When they don't, the report explains why, without forcing a conclusion.
Macroeconomic variables (inflation, interest rates, household debt), trade events, sales force actions, and sponsorships enter the model as context. Without this, the study confuses media effect with market effect — here, the separation is explicit.
Reallocation recommendations by channel and total budget scenarios for the next planning round. The deliverable guides the decision of “where to put the next dollar” — not an isolated chart you have to interpret alone.
Fit metrics for each model (NRMSE and Pareto frontier in Robyn; credibility intervals in Meridian), side by side in a comparable table. Where methodologies converge, the recommendation is firm; where they diverge, the report documents plausible causes (priors, functional form, fit quality) — no black box.
MMM Combined is not choosing Robyn or Meridian — it is training both on the same base and reading the contrast between them as additional evidence. The recommendation gains backing when both methodological traditions reach the same place; it gains honesty when they don't.
Frequentist (Ridge), adstock, Hill saturation, Pareto and NRMSE as fit anchors. In Combined, it runs always in pair with Meridian — to contrast structure and uncertainty, not to “pick” an automatic winner.
Bayesian (MCMC via JAX), delayed adstock, and posteriors. In Combined, it offers the Bayesian counterpoint to Robyn — the report cross-checks intervals and contributions to see where the uncertainty view reinforces or tensions the frequentist.
Four moves in the project: single base; two full training runs; joint readout (including correlations between contribution series or equivalent drivers); then budget scenarios and a single business narrative. One data onboarding — two aligned specifications for honest comparison.
We recalculate offline impacts and investments for better alignment with online data. Ad stock and contribution adjustments are also considered before modeling.
Data such as inflation, interest rates, household debt, and other macroeconomic measures are included in the model for indirect impacts. Investments in sales force, trade marketing, events, and sponsorships are also considered in the final model.
Report integrating both readouts, exports (CSV) of estimates and decompositions from each methodology, cross-check tables and documented assumptions — with README at handoff when it makes sense for the technical team.
The study does not replace reliable data or invent answers where the base cannot support them — but it delivers something isolated tools rarely cover: two reference methodologies on the same base, with explicit readout of the relationships between outputs.
Robyn and Meridian are free and auditable, but “just the library” still requires pipeline, Python, standardization, and time to consolidate readout. Running both in parallel on your own multiplies effort and risk of divergent schemas — Combined puts both on the same track and in the same correlation and convergence narrative.
Closed solutions speed up procurement, but often hide specification, weigh on subscription, and limit export to BI. You depend on the vendor's roadmap to reproduce the method.
Integrated service: one data design, training of both methodologies by the same team, joint analysis of outputs (correlations, convergence, explained divergences) and mix recommendation that honors both methodological traditions — with explicit fit metrics, exportable outputs, and methodologies the technical community already discusses in the open.
From base diagnosis to documented close-out — with a typical window for each phase. What is in the signed scope is what gets done; that is the yardstick for measuring “done” at the end.
Mapping channels, calendar, KPI, and context variables; taxonomy and completeness check before the first model run.
Dual training (Robyn and Meridian) on the approved base; aligned calibrations so outputs are comparable in the validation and cross-check phase.
Interpretation workshops: contributions and elasticities (or equivalents), agreement between methodologies, correlations between drivers, and where it makes sense to dig into divergence — with business sanity checks and limits for decision.
Documented package (report, exports, assumptions). Continuity for new channels, seasonality, or base updates per agreed cadence.
Three study sizes, one-time delivery. Scope is sized by number of brands, channels, exogenous variables, and geography — we align in the initial 45-minute conversation with a sample of your case. Unlike SaaS platforms, you don't pay a recurring subscription; you pay for the study and receive the deliverable.
R$ 17.947
R$ 45.947
R$ 75.457
For agencies that want to deliver the study under their brand, and for data teams that want the study in the technical standard they integrate, specific formats by conversation. Annual recurrence models (1 semiannual retrain + 2 readouts) and pilot packages with discount for the first study — just ask.
MMM is about scope — brands, channels, geography, variables. And sometimes scope changes: an extra brand enters, an additional exogenous variable appears, or you need to look at a region not in the plan. Instead of hiding the rule, we keep the table in view: what triggers an add-on, what becomes a new study, and what each case costs on top of the contracted package.
+35%
Change up to 25% of signed scopeEx.: Essential (R$ 17.947) → add-on of R$ 6.281.
+50%
Change above 25% of scopeEx.: Essential (R$ 17.947) → add-on of R$ 8.974.
25%
Limit by channels changedEx.: Essential — 6 channels in scope. Request adds 1 channel (17%) → targeted (+35%). From the 2nd extra channel (≥33%) → broad (+50%).
25%
Limit by exogenous variables addedEx.: Essential — 3 variables in scope (macro + seasonality + 1 extra). Request adds 1 variable (33%) → broad. More sensitive criterion because each variable reopens Diagnosis.
30d
One round after formal readoutRequests consolidated within the window. Written acceptance. After that = new scope.
New
Outside add-on — separate quoteAnother brand outside scope, geography change (National ↔ regional), engine swap/addition (Robyn/Meridian → other), retrain outside agreed cadence, historical series extension.
Criteria A and B measure % of scope changed · higher index prevails · targeted add-on +35% or broad +50% on package · all within signed scope
MMM only delivers reliable readout when history is stable and channels have real variation over time. Use the cards below to validate whether it makes sense before the conversation — and if something doesn't fit, we'll say frankly which explōrātiō product to tackle first.
Minimum 18 months of series, preferably weekly. Monthly only as exception — reduces within-month seasonality readout.
Three or more channels with varying spend in history. Offline and other media: weekly submission with impacts already calculated and aligned to the model.
CSV or spreadsheet, media exports and KPI on the same calendar; column-by-column mapping documented. The base should arrive treated and ready for the models.
Macro, trade, sales force, and events — the more context, the better the model separates media from external factors.
Typically 4 to 6 weeks from approved base to validated model, depending on data quality and training rounds.
Secure links and encrypted transfer; data stays in your environment. After study delivery, full deletion — LGPD compliance and no reuse in other projects.
MMM measures average effect over time — it does not replace all marketing analysis. In some situations another explōrātiō product delivers a faster answer, or prepares the ground for MMM to make sense later.
MMM works well at weekly or monthly cadence; to track creatives and adjust ad by ad, use your media operational reporting — outside this product's scope.
If campaign names, exchange rates, and sales are in conflict, no model fixes that on the fly. First Data Clean, then MMM.
Anticipating category demand with search and trends is Trends Analyzer or another forecasting design — not to be confused with historical media decomposition.
The model needs contrast over time; with fewer than three channels with real variation, the readout becomes opinion with high R².
Both, always in this product: scope is the integrated study, with joint training on the same base and dedicated readout to cross-check results — correlations, convergences, and explained divergences. We do not position Combined as a standalone sale of a single methodology.
Because frequentist and Bayesian approaches tackle the same problem with different hypotheses — and see uncertainty differently. When both indicate the same pattern (a channel pushing the KPI, for example), the recommendation gains real weight to defend in committee. When they disagree, the report explains why (priors, functional form, fit quality) — a more honest view than choosing one methodology, ignoring the other, and discovering the error only in the next round.
Not required. We work with spreadsheets and treated exports within scope; the essential part is coherent granularity and calendar. Data lake flows can speed up recurring work, but are not a prerequisite for the study.
In practice, 4 to 6 weeks after the base is locked, including training of both methodologies, cross-checking outputs, and business validation. If data requires heavy cleaning before that, we add the Data Clean timeline.
MMM answers a different question: how much of KPI variation over time associates with each channel and context variables. It does not replace MTA at cookie level, but avoids some biases when CRM does not cover the full journey.
Processing and storage follow what is agreed in contract and your company policies — generally, without depending on our cloud for raw data. At close-out, we document where artifacts sit, retention recommendations, and LGPD alignment, coordinating with your legal or DPO when needed.
Methodological report integrating both lines of evidence, documented assumptions, tabular exports by methodology, and cross-check material (where it makes sense: agreement, correlations between equivalent drivers), plus interpretation workshops. The core of the decision is the joint assessment; exports support audit and your BI.
Technical terms that appear in the report, workshops, and CSV exports. It is here because whoever operates data (BI, analytics, revenue) will want to open and audit — whoever only receives the recommendation does not need to read it.
Statistical model that estimates how channels, media, and context factors associate with KPI variation over time — basis for budget allocation, not a substitute for cookie-to-cookie attribution.
In this product: Robyn and Meridian trained on the same base, with explicit readout of convergence, correlation, and divergence between outputs.
Share of the outcome (or modeled KPI) attributed to a media driver or exogenous variable in the model decomposition.
Splitting the KPI into parts explained by each channel, baseline, and controls — allows comparing relative weight of investments.
Carry-over effect: the impact of an investment does not disappear on the day the ad ran; it decays over time per calibrated parameters.
Curve where, beyond a certain investment level, spending more on the channel yields lower marginal return — common functional form in Robyn.
How much the KPI responds proportionally to a percentage increase in spend on a channel — useful for simulating budget scenarios.
Regularization that reduces overfitting when many correlated channels exist — anchors estimates without zeroing coefficients like Lasso.
Set of solutions where you cannot improve fit and parsimony at the same time; Robyn uses it to choose models balanced between complexity and error.
Normalized root mean squared error — model fit metric against historical data; lower (within scope acceptability), better comparable-scale fit.
How investment enters the equation after adstock and saturation — defines the response curve per channel before regression.
Estimates parameters as point values (with intervals derived from sample fit), without explicit priors as in Bayesian — central approach of Robyn.
Combines observed data with priors (calibrated initial beliefs) and produces posterior distributions — uncertainty is part of the answer, not just a single number.
Initial assumption about parameters (e.g., a channel's elasticity) before seeing data; influences the result when the series is short or noisy.
Parameter distribution after incorporating data — in the report appears as central estimate and credibility interval.
Markov Chain Monte Carlo — method to sample the posterior when no closed analytical solution exists; Meridian runs via JAX for performance.
Range where the parameter or contribution falls with high probability (e.g., 90%) — counterpoint to the frequentist single value.
Google numerical library used by Meridian to accelerate Bayesian training on GPU/TPU when available.
Parameterized carry-over variant in the journey — effect peak may occur days after exposure, not only on day zero.
When both methodologies point the same way (e.g., same channel as relevant driver) — reinforces the mix recommendation.
When contributions or elasticities disagree — the report documents plausible causes (fit, priors, functional form).
Measure of alignment between equivalent outputs from both methodologies; not to be confused with raw correlation between channel spend in the spreadsheet.
How well the model reproduces historical KPI — compared between methodologies with their own metrics (NRMSE in Robyn, Bayesian diagnostics in Meridian).
MMM Combined delivered within signed scope. Ready for the planning committee to defend mix, for the revenue team to run budget scenarios, and for the data team to audit the methodology — no “almost ready,” no endless review round, no half-deliverable.
The initial conversation is technical — we look at a sample of your data, validate whether history and channels fit a well-built MMM, and say frankly: there is a study here that solves it, or you need Data Clean first, or MMM is not yet the path for your question. If yes, we design scope, channels, exogenous variables, and timeline on the spot. If not, you leave with the diagnosis anyway.