Marketing mix modeling (MMM) estimates how much each marketing channel contributed to sales, using weekly totals of spend, exposure and revenue rather than tracking individual customers. For an ecommerce brand, it answers the questions attribution cannot: what TV, podcasts or Meta actually added, and where the next dollar should go. Google and Meta now give away the modeling software, so the cost and the risk sit in the data underneath: two to three years of clean, complete weekly history from every channel. This guide sets out the data an MMM needs, where ecommerce brands usually fall short, and what that data work costs at published prices. It is from Brand Media Works, an AI consulting and marketing data engineering firm for ecommerce brands.

What data does a marketing mix model need?

Google's Meridian documentation and Meta's Robyn guide ask for the same core data, collected every week:

  • The outcome: revenue, orders or another KPI. Meridian requires a KPI that can be added up across regions and weeks, so rates such as conversion rate do not work.
  • Media exposure and spend for each paid channel. Meridian asks for one exposure metric per channel, such as impressions, and accepts spend as a fallback. Meta's guide prefers impressions as the model input, with spend collected to calculate return. Meta's guide says to avoid clicks for digital channels, because clicks miss people who saw an ad and bought later.
  • Non-media marketing: promotions, discounts and prices, usually as weekly price data or a flag for each promotion week.
  • Controls: seasonality, holidays and other factors that move sales on their own.
  • Organic activity such as email and social posts, which Meridian accepts as organic media with no spend.

Two rules decide whether the data is usable. Meridian needs a complete dataset with no missing values. And Meta's guide asks for actual delivery, not the media plan: the impressions and spend that actually ran, not what was planned.

What does the data work cost, and when is it worth building in-house?

The modeling software costs nothing: Meridian is released under the Apache-2.0 license and Robyn under MIT. Meridian runs on Python and Google recommends at least one GPU. The cost is in collecting, cleaning and keeping the weekly data, which Meta's guide calls one of the most important but time-consuming steps in an MMM study.

  • Storage and pipelines. Google's BigQuery Data Transfer Service loads Google Ads data with no transfer orchestration charge, Fivetran's Free plan covers 500,000 monthly active rows, and BigQuery's first 1 TiB of queries and 10 GiB of storage each month are free (read 18 October 2026). Three years of weekly channel totals is a small table, so the running cost of the warehouse is usually small.
  • The build. One data consultancy publishes about $4,000 for custom pipelines from one or two sources with automated refresh, and about $10,000 across around three sources (read 25 October 2026). An MMM usually needs every paid channel, the store and the email platform.
  • The vendor route. Meta's guide says a third-party MMM vendor is usually the most expensive option, though it needs the least of the brand's time. Funnel's pricing page estimates $2,250 a month, billed annually, for its digital measurement product at $500,000 to $1M of monthly ad spend (read 29 October 2026), or $27,000 a year (our arithmetic).

Building the data foundation in-house pays off when the brand wants to own the history, run the model more than once a year and reuse the same tables for reporting and testing. A vendor is the better choice when no one in-house can maintain the pipelines or read the model. Marketing data engineering describes how Brand Media Works builds the warehouse an MMM reads from, and how to build a marketing data warehouse covers the steps.

How much history does an MMM need?

  • Meridian: at least two years of weekly data for a model by region, and three years for a national model. With only monthly data, at least three years.
  • Robyn: at least two years of weekly data, and more than two years, such as four to five, if only monthly data exists.

Model size matters as much as length. Meta's guide recommends at least 10 observations for every input variable, so two years of weekly data, 104 weeks, supports about 10 variables (our arithmetic). Google's Meridian documentation gives a national example with 26 parameters on two years of weekly data, about 4 data points per parameter, and calls that too small to estimate reliably. Cutting to 10 parameters on three years of weekly data gives about 15 per parameter.

That is why regional data helps. Meridian prefers data by region, because the number of data points is the number of regions times the number of weeks: Google's example of 105 regions over three years gives 16,380 data points.

Why do ecommerce brands run short of MMM data?

Three common gaps:

  1. History expires. Meta's Insights data can only be pulled for the last 37 months, and Google Ads limits granular reporting data to a 37-month retention window (as documented by Airbyte, read 18 October 2026). A national Meridian model needs 36 months. A brand that only reads its numbers inside each ad platform is always close to losing the oldest year.
  2. Exposure is missing. Many brands keep spend and platform-reported conversions, but not weekly impressions by channel, and not delivery for channels like podcasts or influencers.
  3. Promotions are not recorded. Discounts, sales and price changes often live in emails and spreadsheets. Without a weekly record, the model credits a sale week to whichever channel spent most.

The fix is the same in each case: load every source into one warehouse on a schedule, keep the history, and record promotions as data. Meta also offers an MMM Feed with Meta impressions and spend split by day, country, region, placement and campaign.

How does incrementality testing fit with an MMM?

Meta's Robyn guide strongly recommends calibrating an MMM with experiments, such as Meta Conversion Lift or geography-based tests, and running them regularly. The test gives a measured result for one channel, and the model uses it to keep its estimates for that channel in check. Incrementality testing for ecommerce brands sets out how to run those tests and what they need from the brand's data.

Book a discovery call

Frequently asked questions

What is marketing mix modeling?

A statistical method that estimates how much each marketing channel and outside factor contributed to sales, using aggregated weekly data rather than individual customer tracking. Google describes it as privacy-safe for that reason.

Is Google Meridian free?

Yes. Meridian is open source under the Apache-2.0 license. Meta's Robyn is open source under the MIT license. The costs are the data work, the computing time and the people who run and read the model.

Can a smaller ecommerce brand run an MMM?

Only with enough history and enough variation. A brand with two years of weekly data can support about 10 inputs (our arithmetic from Meta's 1 to 10 guide), so it may need to combine small channels, which Google also advises for channels with low spend. A brand with less history is usually better served by incrementality tests first.