MMM, MTA or incrementality: a guide by data volume

Pick the method whose published minimum your data meets. Robyn's documentation asks for "a minimum of two years of historical weekly data" for a media mix model, Google's Meridian GeoX documentation says to "include a minimum of 10 geos" for a geo experiment, and Google Ads recommends "at least 200 conversions and 2,000 ad interactions" in 30 days for data-driven attribution, which assigns credit from observed paths, not from a test you run. Count your weeks, your regions and your conversions against those bars first.

What does each method need, according to its own documentation?

Four statements, each from the tool's own pages:

1. Media mix modeling (MMM), Meta's Robyn. The analyst's guide says an MMM "will need a minimum of two years of historical weekly data" and recommends "1 independent variable: 10 observations".
2. MMM, Google's Meridian. Its data-needs page calls its guidance "rough and directional". In its example, two years of weekly data against 26 parameters gives "four data points per parameter", "too low to estimate the model reliably"; three years against 10 parameters gives "roughly 15", where "you might be able to glean some directional information".
3. Geo experiments. The Meridian GeoX FAQ says to "include a minimum of 10 geos for a single-cell design" and takes daily data only. Meta's GeoLift best practices recommend, at minimum, "25 pre-treatment periods of 20 or more geo-units". The two tools don't set the same bar.
4. Multi-touch attribution (MTA), Google. The Google Ads data-driven attribution page says every conversion action is eligible "regardless of conversion or interaction volume" and recommends "at least 200 conversions and 2,000 ad interactions in supported networks within a 30-day period", adding that "the performance of data-driven attribution improves with more data". GA4's attribution help page says the model learns from "both converting and non-converting paths" and that its counterfactual estimates for Google ad exposures come from "data from randomized controlled trials".

MTA needs no spend data and no test. A path report has no group you deliberately held back, which is what an experiment adds, and the GA4 page gives no sample sizes or results for the trials it mentions.

Why do small stores struggle to meet those bars?

One published result, then two sums. None of the pages cited here gives a minimum order volume for a small Shopify store, so the evidence comes from the documentation above and from studies of large advertisers.

1. Experiments need volume. Lewis and Rao (The Quarterly Journal of Economics, 2015, peer-reviewed) report twenty-five large field experiments with major U.S. retailers and brokerages, most reaching millions of customers. The median confidence interval on return on investment was "over 100 percentage points wide", and informative experiments "can easily require more than 10 million person-weeks".
2. Mix models run out of variables. For illustration: 52 weeks of data at Robyn's ratio of one variable to 10 observations supports about 5 variables, and three channels, a promotion flag and a holiday flag already use 5.
3. Experiments run out of people. For illustration: a 4-week test that needs 10 million person-weeks needs 2.5 million people, treatment and control together. Treat it as an order of magnitude: it reflects very volatile sales per customer, not a threshold for your store.

Which method fits which situation?

Rule of thumb, using only the thresholds above:

1. Under two years of weekly history, or flat budgets: below Robyn's stated minimum, so hold off on a mix model. Move one channel's budget in steps and log the dates; that builds the variation a model needs.
2. Fewer than 10 regions you can target separately: below GeoX's minimum for a single-cell design. GeoX says that with few geos "statistical power is likely to be low, leading to high MDEs or requiring very large budgets", where an MDE is the smallest change in conversions a test can detect.
3. Neither bar met: read GA4's paths as a description of how buyers arrive, not of what caused orders. To test a cut, pause the channel in one region for at least 15 days of daily data, GeoLift's stated minimum, and compare orders with the other regions. A flat result means the effect was too small to see, not that it is zero. The holdout test planner counts days.
4. Both bars met: run a geo experiment on the channel where a wrong call costs most. Meridian's pages say geo experiments can "calibrate your MMM", so the two can check each other.

How do you check your own data in an hour?

1. Count weeks. List weekly orders and weekly spend per channel for as far back as you have them. Pass: 104 weeks or more with no gaps. Fail: fewer.
2. Check the variation. Did each channel's weekly spend ever move? Robyn warns that when "TV activity has remained constant for the whole time period, the model can have difficulty determining how TV has impacted sales", and Meridian says "insufficient variation in the media spend adversely impacts national models". Flat spend is a fail for that channel.
3. Divide. Weeks divided by channels plus controls plus time parameters. Robyn recommends 10 or more per variable. Meridian calls 4 too low and says about 15 might give "some directional information".
4. Count regions. In your Shopify orders export, count the shipping provinces or states with orders on most days; GeoX points to "higher volume and fewer zero-count days" to steady a volatile series. Pass: 10 or more for GeoX, 20 or more for GeoLift. Fail: fewer.
5. Count conversions. In Google Ads, add up the last 30 days of purchases and ad interactions; skip this step without it. Pass: 200 conversions and 2,000 ad interactions or more. Fail: fewer, where Google says data-driven attribution still works with less precise credit.
6. Decide. Apply the rule of thumb above and write down which bar you missed.

Sources, 30 September 2026: An Analyst's Guide to MMM (Meta, 2024); Amount of data needed (Google, last updated 3 June 2026); Meridian GeoX FAQs (Google); GeoLift Best Practices (Meta, 2025); The Unfavorable Economics of Measuring the Returns to Advertising (Lewis and Rao, The Quarterly Journal of Economics, 2015); Get started with attribution (Google Analytics Help); Data-driven attribution (Google Ads Help).

Related answers

• MMM vs MTA: which question each one answers
• Marketing mix modeling
• Incrementality testing
• Do you need attribution software? A sizing test
• Incrementality test duration and sample size