Google Ads Incrementality Test with Meridian GeoX: Setup Guide
The 9-step runbook for running a Google Ads geo incrementality test with Meridian GeoX. Setup, market selection, iROAS, and MMM calibration.
.png)
To run a Google Ads incrementality test with Meridian GeoX, pick one campaign, choose a design (holdback, go dark, or heavy up), split your geographies into treatment and control markets, align Google Ads spend with backend revenue by region, run the test long enough to detect real lift, calculate iROAS, then feed the result into Meridian MMM as a Bayesian prior. GeoX is not available to everyone yet. Run the framework now and port it into GeoX when it ships.
Executive Summary
Google previewed Meridian GeoX on May 5, 2026, an open-source, publisher-agnostic geo experimentation tool that feeds Bayesian priors into Meridian MMM. Testing begins later in 2026 and Google hasn't said when it will be available to everyone.
GeoX is not the same as Google's Conversion Lift. Conversion Lift measures Google Ads inside Google. GeoX measures any channel against backend revenue you control. The methodology isn't even new, Google's trimmed_match and matched_markets repos have been on GitHub for years. What's new is the packaging, the name, and the explicit tie to Meridian.
Open source moves the cost from license to labor. Running GeoX in production still requires data pipelines, an analyst, and contamination monitoring. If your measurement team is one performance marketer, a managed platform is cheaper than self-hosting.
Across 225 geo tests on Stella's platform, the median iROAS was 2.31x with 88.4% reaching significance. Brands running tests today will plug into GeoX immediately when it ships. Brands waiting will still be learning the basics in 2027.
The rest of this post is the runbook.
What is Meridian GeoX?
GeoX is Google's open-source tool for running geographic incrementality experiments. It supports three test designs: holdback, go dark, and heavy up. Results convert directly into Bayesian priors that calibrate Meridian MMM. GeoX is publisher-agnostic, so it can measure Meta, TikTok, podcasts, or any channel by geography, not just Google Ads. Testing begins later in 2026.
The mechanic is the same as any geo holdout. Split geographies into a treatment group and a control group, change media exposure in one, and measure the gap. GeoX adds structure around three specific designs:
- Holdback: keep ads active in most regions, pause them in a small set of holdout regions
- Go dark: pause media entirely in selected regions and measure the drop
- Heavy up: increase spend in selected regions and measure the gain
GeoX also runs multiple treatments against a shared control in the same study, which makes multi-cell tests cheaper. The output feeds into Meridian as priors, so the MMM learns from real causal evidence instead of inferring impact from correlation.
A useful frame: Conversion Lift is the platform handing you a lift number. GeoX is the methodology, open and inspectable, that lets anyone produce one.
Can you run GeoX today?
Not yet. GeoX entered testing in late 2026 and Google hasn't published a launch date. You can read the docs, look at the underlying trimmed_match and matched_markets repos, and run geo holdouts now using your current setup so you have baseline tests ready when GeoX ships.
The brands that get value from GeoX on day one are the brands already running disciplined geo tests today.
How to run a Google Ads incrementality test with Meridian GeoX
Nine steps. None are optional.
Step 1: Choose the campaign
Pick one campaign worth testing. The campaigns where platform ROAS and iROAS diverge most are usually:
- Branded search. Most conversions would have happened anyway. iROAS often lands 60-80% below platform ROAS.
- Performance Max. Mixes branded search, shopping, YouTube, and Display, so platform attribution is opaque. See the PMax-specific guide.
- YouTube and Demand Gen. Long conversion windows, weak last-click attribution. Most likely to be undercredited.
- Non-brand search. Cleanest test for whether the channel acquires new customers.
One test answers one question.
Step 2: Pick the test design
Holdback is the default for most DTC brands. Go dark is right when you suspect a channel is barely incremental. Heavy up is right when you've validated a channel and want to test scaling.
Step 3: Select treatment and control markets
This is the single biggest predictor of significance. Across Stella's 225-test benchmark, the 11.6% of tests that didn't reach significance almost always failed on pre-period matching, not sample size.
Good matching means:
- Similar revenue baseline in the 90 days before the test
- Similar seasonality over the past 12 months
- Similar media spend share across all channels
- Similar daily conversion volume
- Geographic separation (don't pair Manhattan and Brooklyn)
Most teams use DMAs as the geographic unit. Brands with retail footprints use trade areas. Pick what matches how your business runs.
Step 4: Prepare the data
Before launch, you need this dataset joined and clean:
- Daily revenue by geo (from Shopify or your CRM, not Google Ads)
- Daily spend by geo across every paid channel
- Daily orders by geo
- Promo calendar with start and end dates
- Inventory outages on revenue-driving SKUs
- Major site changes during the test window
- Campaign IDs and geographic targeting settings
- Treatment vs control assignment per geo
- Contribution margin
Critical: outcome data has to come from your backend, not Google Ads. Platform conversions are filtered by Google's attribution model, which is the thing you're testing against. Using them as your outcome variable defeats the experiment.
Step 5: Configure Google Ads location settings
This is where most geo holdouts get contaminated.
Google Ads has two location targeting modes:
- "Presence or interest" (the default): shows ads to people in, regularly in, OR showing interest in a location. That last category leaks ads into control regions whenever someone searches "best running shoes Seattle" from a control market.
- "Presence": shows ads only to people physically located in the targeted regions.
For a geo holdout, you almost always want Presence. See Google's docs on advanced location options for the mechanic.
Other contamination sources:
- Search partners and Display Network can override geographic restrictions. Disable both.
- PMax campaigns reallocate budget across geographies dynamically. See the PMax guide.
- Commuting zones can leak exposure across DMA boundaries. Avoid pairing adjacent markets.
Step 6: Run the test
Run length should be driven by statistical power, not calendar habit.
Pre-launch QA checklist:
- Treatment and control markets are statistically similar in the pre-period
- Control regions are fully excluded in campaign settings
- Location targeting is set to "Presence"
- Other channels are stable (no concurrent test on Meta or TikTok)
- Promo changes during the test window are documented
- Backend revenue tracking confirmed working by geo
Step 7: Calculate lift and iROAS
Three formulas:
Incremental revenue = Treatment revenue − Expected revenue (what treatment would have earned without the ads)
Expected revenue is where synthetic controls come in. They use a weighted combination of control markets to model what treatment would have done without the campaign. See Stella's guide to weighted synthetic controls.
iROAS = Incremental revenue / Incremental ad spend
Incremental profit = (Incremental revenue × Contribution margin) − Media cost
iROAS without contribution margin is a vanity number. A 2x iROAS is excellent at 60% margins and unprofitable at 30%. The decision is profit, not the ratio.
Step 8: Feed the result into Meridian MMM
This is what GeoX changes most. The output of a GeoX test becomes a Bayesian prior in Meridian MMM, which means the model learns from your real experiment instead of guessing from correlation.
In plain English, the MMM stops pattern-matching on historical data and starts anchoring its estimates to causal evidence. Stella's MMM runs 87% accurate on average in forward tests, and 95% when calibrated with iROAS from real experiments.
The eight-point gain matters because budget decisions live downstream of the model. An MMM at 87% misallocates a meaningful share of every quarter's spend. An MMM at 95% misallocates less.
Most teams don't have a single causal experiment to calibrate their MMM with. GeoX makes that gap embarrassing instead of invisible. For more, see Getting Started with MMM Using Google Meridian.
Step 9: Make the budget decision
The test should lead to an action.
The worst outcome is running the test, sharing the result, and changing nothing.
.png){kind=avatar}
.png)
.png)
.png)
.png)