Your Marketing Model Is Misleading You: How to Audit It

Marketing mix model dashboard showing misleading attribution data

Author:

Ara Ohanian

Published:

October 30, 2025

Updated:

November 20, 2026

On September 10, 2026, Google said it was building agentic features into Meridian, its open-source marketing mix model, “to help you audit data quality, resolve errors, and guide model building in real time.” In May it announced it was bringing Meridian into Analytics 360. The company selling much of your media also supplies the model, the planner and the assistant that checks it.

None of that makes Meridian wrong, but the questions must come from you. In July, the Coalition for Innovative Media Measurement proposed a standard factsheet for mix models: data sources, transformations, priors, validation methods, uncertainty ranges, refresh cadence and limitations. Ask for it before reading any ROI figure.

Four questions that expose a mix model

Did the data choose the curves?

Every mix model assumes how long an ad keeps working (adstock) and how fast returns fall as spend rises (saturation). Robyn pairs a Hill curve with geometric or Weibull decay. In a 2017 Google paper, Yuxue Jin and colleagues found that with small samples, priors “have a big impact on the posteriors” and can bias estimates. Meridian charts how much the data, versus the prior, set each channel’s carryover. Ask to see it.

Which channels sit on their priors?

Meridian’s default priors are “moderately informative.” Its health checks flag channels whose posterior ROI barely moved from the prior, which “suggests the data signal for these channels was not strong enough to update the model’s beliefs.” A channel whose output matches its input is the modeler’s opinion with a chart around it.

Which channels were calibrated?

Meridian’s calibration guide warns that “experiments and MMMs often have different estimands,” and that results from another period “might not be directly applicable.” Robyn accepts only a point estimate as calibration input, dropping the test’s uncertainty. Ask which tests fed the model, how old they are, and which channels had none.

How wide are the ranges?

Ask for the credible interval on every ROI, not the midpoint. Meridian’s docs urge caution with response curves “beyond the range of the available historical media data,” and its optimizer by default keeps each channel within 70 to 130% of its spend on a fixed budget. A plan to triple a channel is usually extrapolation. Robyn returns several candidate models, and “a final model should be selected by the user.” If the runners-up rank channels differently, the ranking was a choice.

Where marketing mix modeling for DTC brands goes wrong

For a brand spending mostly on search and social, the first risk is search claiming demand it did not create. Advertisers often spend more when demand is stronger, so a model that sees only spend and sales over-credits the ads, as a 2018 Google paper showed. In one case, a naive model put search ROAS almost 15 times higher than an experiment did; controlling for search query volume brought it much closer. Ask whether yours controls for branded search. Red flags, per Google’s docs: a negative baseline or one channel dominating all others, and a negative R-squared. Inputs pass through too: our breakdown of dashboard numbers shows which are modeled, and tracking gaps from signal loss go in with them.

Let a lift test settle the disagreement

A sensible order: take the channel where the model wants the biggest change, test it with a geo or platform lift study (our guide to incrementality testing compares them), and feed the result back as a prior. Google’s GeoX library is built for that. A rule of thumb: if one test shifts several unrelated channels, the model is less stable than its deck suggests. Read the test by its range, as our piece on reading an incrementality result explains. The model proposes, the test decides, and budget follows the test.

That loop is the daily work of performance marketing, and our work with a client usually starts with an audit of its current marketing.

Frequently Asked Questions

How can you check an MMM forecast before moving budget?

Ask for its out-of-sample fit on weeks it was not trained on. Meridian’s documentation calls this worth checking, though not the main test of fit. Check that the recommended spend sits within the range the model has seen. Then test the forecast: shift budget as the model suggests in some regions only, hold the rest steady, and compare. If those regions move as forecast, scale the change.

How should you present MMM results to a CFO who distrusts attribution?

A sensible approach is to answer finance’s first question: how could this be wrong? Show ranges, not point estimates, and say which channels an experiment has checked. Express returns as contribution margin, not revenue. Present the budget change as a bounded test, with a set shift, a fixed period and the result that would reverse it. Approving a test asks for less trust than approving a forecast.

How much historical data does a mix model need?

Meta’s Robyn guide sets a floor of two years of weekly data, with about ten observations per variable. Google’s Meridian documentation is stricter: in its example, two years of weekly national data gives about four data points per parameter, too few to estimate a model reliably, and three years with fewer channels may yield only directional results. Regional data adds observations. With less history, start with lift tests and model later.

Is an open-source model good enough for a DTC brand?

The code is rarely the weak point. Meridian is free and well documented; the hard work is the data, the priors and the calibration tests. Robyn’s future is less certain: AdExchanger reported in July 2026, citing unnamed sources, that Meta is winding it down. Open code is not automatically neutral either: Meridian’s defaults come from Google, which also sells you search and YouTube ads.