Marketing Mix Modeling Is Back: How to Measure Growth Without User-Level Tracking

On January 12, 2026, Meta stopped returning 7-day view and 28-day view conversions through its Ads Insights API. Any report that pulled those windows lost that credit overnight. Changes like this keep coming, and they’re the reason I now put marketing mix modeling into every measurement plan I write.

I’ve run demand gen budgets through each round of signal loss. Apple’s App Tracking Transparency prompt arrived with iOS 14.5 in April 2021. In my experience, platform-reported conversions and finance’s revenue numbers drifted apart soon after. Attribution tools can only credit what they can see. Marketing mix modeling, or MMM, skips the individual user. Instead, it works from weekly totals: spend by channel, sales, prices, promotions and seasonality. So it keeps working when cookies, device IDs and view-through windows disappear.

Why Marketing Mix Modeling Is Back in 2026

The method itself is old. Packaged goods companies ran regression models on TV and print spend long before anyone tracked a click. Digital marketers dropped it because pixel tracking was faster and cheaper. As a result, most performance marketing programs assumed you could trace every conversion to an ad.

That assumption keeps breaking. In October 2025, Google said it would retire most Privacy Sandbox technologies, including the Attribution Reporting API. Its stated reason was low adoption. Chrome kept third-party cookies, but the planned privacy-safe replacement for cross-site conversion measurement is gone. If you have tried to keep retargeting working without third-party data, you have watched the same erosion from the audience side.

Meanwhile, the tools for marketing mix modeling got cheaper. A first model used to mean hiring a specialist consultancy. Now Google and Meta both publish free, open-source MMM code, with documentation and sample notebooks.

Google Meridian and Meta Robyn Compared

Google Meridian

Meridian became available to everyone on January 29, 2025, after Google tested it with hundreds of brands. It is a Bayesian model written in Python. So you can give it what you already believe about a channel, such as a lift test result. The model then blends that prior with your data. Meridian also measures reach and frequency instead of raw impressions. In addition, it can add Google search query volume as a control for organic demand.

Google’s documentation recommends geo-level data when you have it, because pooling regions gives the model more observations than national totals. If you lack in-house data science, Google also trains and certifies outside partners. It had more than 20 at launch.

Meta Robyn

Robyn comes from Meta’s Marketing Science team. It runs in R, with a beta Python API. It fits ridge regression models and uses the Nevergrad optimizer to search combinations of adstock and saturation. Adstock is how long an ad’s effect lingers. Saturation is the point where extra spend stops paying back. Robyn then returns several candidate models that balance fit against plausibility, and an analyst picks one. Meta built it for granular datasets with many independent variables, which suits digital and direct response advertisers.

In practice, the choice comes down to your team. Meridian fits a Python shop with regional data, while Robyn fits an R shop with many digital channels. Neither one removes the labor, though. Someone still has to collect the data, check the output against common sense and explain it to finance.

When Marketing Mix Modeling Fits Your Business

Marketing mix modeling learns from variation over time, so it needs history. The Robyn analyst’s guide sets a minimum of two years of weekly data. It also recommends about 10 observations for every independent variable. With 104 weeks, that allows roughly 10 variables in total, counting channels, promotions and seasonal controls.

MMM fits well when:

• You spend across several channels, including some that clicks can’t measure well, such as TV, podcasts, out-of-home or YouTube.
• Your spend moved over time through flighting, launches or regional tests. A flat budget gives the model nothing to learn from.
• Marketing shows up in a weekly outcome. For B2B with long sales cycles, model pipeline created rather than closed revenue.
• You need a budget answer. MMM tells you how to split next quarter’s money across channels. It won’t tell you which ad creative won.

However, it fits poorly for a young company with one main channel and a year of data. In that case, a clean holdout test on that channel will tell you more, and sooner.

Pairing MMM With Incrementality Tests

An MMM finds patterns in historical data. When two channels always rose and fell together, the model cannot separate them. Experiments break that tie. Robyn’s guide strongly recommends calibrating the model with experimental results treated as ground truth. It also suggests running incrementality studies on a regular schedule. Meridian takes the same input in a different form: lift results become priors.

The loop I use for marketing mix modeling has four steps:

1. Build a first model and note which channels have the widest uncertainty ranges.
2. Run a geo holdout or platform lift test on the biggest uncertain channel.
3. Feed the result back into the model as a prior or calibration point.
4. Refit each quarter, then test the next channel.

Testing also got cheaper. At Google Marketing Live in May 2025, Google said advertisers could run incrementality tests with experiment budgets as low as $5,000. Still, a platform’s own lift study grades its own homework. Where you can, run geo tests that you design and control.

The Data You Need Before You Start

Most of the work in a first model is data collection. Robyn’s guide sorts inputs into paid media, organic activity and context variables. That split works as a checklist even if you choose Meridian.

• An outcome measured weekly, such as revenue, orders, sign-ups or pipeline, by region if possible. Use the number finance reports, not platform conversions.
• Paid media spend plus impressions, clicks or GRPs per channel per week. Google offers advertisers an MMM Data Platform for its own media data. Meta’s Ads Insights API also includes a breakdown built for MMM exports.
• Organic activity, such as email sends, PR coverage, social posts and webinars.
• Context, such as price changes, promotions, holidays, stockouts and competitor launches.
• Every lift test you trust, with its dates and confidence range.

Two years is the floor. If you only have monthly data, Robyn’s guide recommends four to five years. Also clean up channel names before you start, because a channel split differently each year will confuse the model. Finally, agree on which marketing KPIs to track before the first output lands, so nobody argues about the outcome variable later.

Next Steps for Your First Marketing Mix Modeling Project

• Pull at least 104 weeks of outcome data from finance and get sign-off on the definition.
• Export spend and delivery per channel per week, with the same channel names across every year.
• List every promotion, price change, outage and launch with its dates.
• Pick a tool. Choose Meridian if your team works in Python and has regional data, or Robyn if it works in R. Otherwise, hire one of Google’s certified Meridian partners.
• Choose one uncertain channel and schedule a geo holdout for next quarter.
• Keep your multi-touch attribution model for daily optimization, and use marketing mix modeling for quarterly budget splits.