Meridian GeoX: How to Plan a Better Paid Media Incrementality Test

Meridian GeoX is Google’s open-source framework for running geographic incrementality experiments across publishers. It uses matched test and control regions to estimate what would have happened without a media change, helping marketers move beyond platform attribution.

The software can support the analysis, but the experiment is only as credible as the market selection, business outcome, treatment and decision rule defined before launch.

What is Meridian GeoX?

Google describes Meridian GeoX as a global, transparent and publisher-agnostic incrementality solution. It supports single-cell and multi-cell studies and can evaluate holdback, go-dark and heavy-up designs.

GeoX can run independently or calibrate Google’s Meridian marketing mix model. Experiment results can become priors that help the MMM estimate channel ROI more accurately.

What question does a geo experiment answer?

A geo experiment estimates the additional outcome caused by a treatment. Examples include:

• What revenue did paid social create beyond existing demand?
• Does increasing YouTube spend raise total qualified leads?
• What happens when branded search is reduced in selected markets?
• Does a new channel generate incremental store sales?

Phrase the question around a business outcome and a specific change. “Does Meta work?” is too broad. “Does a 30% prospecting increase generate incremental first-time customer revenue?” is testable.

Choose the right outcome

The dependent variable should be consistently measurable by geography and time. Good outcomes include revenue, orders, qualified leads, activated accounts or store visits when tracking is sufficiently reliable.

Avoid using platform-reported conversions as the only outcome in a cross-publisher experiment. They inherit the attribution rules of the platform being evaluated.

Build comparable geographic groups

Use historical behavior

Match regions with similar baseline outcome trends, size and volatility. The objective is not to find markets that look identical on one average. It is to find controls that predict the test markets during the pre-period.

Check operational differences

Inventory, pricing, promotions, sales coverage, weather and local events can disrupt comparability. Document expected changes before assigning markets.

Limit contamination

Media can cross borders through commuting, connected television, shared DMAs and location inaccuracies. Choose regions and channels where treatment exposure can be controlled reasonably well.

Select the treatment design

• Holdback: maintain normal activity in test regions while withholding it from controls.
• Go-dark: stop an existing channel in selected regions to estimate what disappears.
• Heavy-up: increase investment in test regions while controls remain at baseline.
• Multi-cell: compare more than one treatment against a shared control.

Google notes that native multi-cell execution can reduce cost and time. It also adds interpretation complexity, so each cell needs a clear hypothesis and enough statistical power.

Run design analysis before launch

Use historical data to estimate the minimum detectable effect, test duration and budget required. If the experiment can only detect an unrealistic lift, it is unlikely to resolve the decision.

Define in advance:

• Primary outcome
• Pre-period and test dates
• Included markets
• Treatment budget
• Expected conversion lag
• Decision threshold
• Conditions that would invalidate the test

Protect the experiment while it runs

Do not change offers, tracking or sales operations in only one group unless that change is the treatment. Keep a log of outages, promotions, competitor activity and major regional events.

Monitor execution without repeatedly checking results and ending the test at a favorable moment. The analysis plan should specify when the final read occurs.

Interpret lift and uncertainty

Report the estimated incremental outcome, uncertainty interval, cost and incremental return. A positive point estimate with wide uncertainty may not justify scaling.

Compare the geo result with platform attribution and the broader stack in Attribution, Incrementality and MMM. Differences are expected and should be investigated, not averaged away.

Use GeoX results to improve MMM

When the experiment is credible and relevant, use it to calibrate channel priors in Meridian. Google’s documentation explains that GeoX can convert experiment results into priors for MMM.

Recency matters. A test from a different product, market or competitive environment may be less informative than a recent experiment that matches current conditions.

Frequently asked questions

Does Meridian GeoX work only with Google media?

No. Google describes it as publisher-agnostic, so it can evaluate media across platforms when geographic execution and outcome data are available.

Do small advertisers have enough data?

Not always. The answer depends on geographic scale, outcome volume, volatility and expected effect. A design analysis should determine feasibility before money is committed.

Can a geo test replace attribution?

No. Geo tests answer causal channel questions periodically. Attribution remains useful for operational reporting and diagnostics between experiments.

Sources

• Google Meridian GeoX overview, accessed September 8, 2026
• Google Meridian GeoX analysis methodology, updated August 28, 2026
• Google Meridian: Calibrating priors with experiments, updated September 2026