Google's GeoX is out of beta. Go dark or heavy up? Pick the geo test by the question it answers

KEY TAKEAWAYS

1. Go-dark and holdback tests measure average lift against zero spend. That is the result Meridian's calibration tool is designed to take.
2. Heavy-up and go-dim tests measure the return on a slice of spend. They inform budget changes and check the model. They are not recommended as priors.
3. Budget does not lower the minimum detectable effect. It sets how inefficient a channel can be and still show up as a result.
4. Google's 31% cost saving claim has no published method. The documented checks, such as the R-squared and A/A thresholds, are what to rely on.

Google's geo experiment library is now a finished product. Most of the coverage explains what a geo test is. The harder choice comes first: which kind of test to run. Go dark, go dim, heavy up and hold back sound like settings on one tool. They measure different things, and only some of them give a number your marketing mix model is built to accept.

What changed in September?

Two releases, one week apart.

• 2 September 2026. Google released Meridian 2.0, its open-source marketing mix model, and said Meridian had passed one million downloads. The release moves the default engine to JAX, which Google says is twice as fast and four times as memory efficient.
• 9 September 2026. Meridian GeoX, the geo experiment library first shown on 5 May, left beta and became generally available worldwide.

The pieces connect. Meridian now scores each channel on how much it needs an experiment. GeoX designs and reads the experiment. A new calibration module turns the result into a prior for the next model run. GeoX can also run on its own, and it can test media outside Google.

One claim needs a caveat. A GeoX product manager said the library cuts budgets by more than 31% for large advertisers and design time by more than 94%. PPC Land and Animo both point out that no sample, test period or comparison set was given. Animo sells a competing measurement product, but on this point it is right. Treat both figures as unproven.

Which question does each test type answer?

Each test type compares the regions you change with regions you leave alone. What you change decides what you learn.

Google's own documentation draws the go-dim line plainly. A partial go-dark measures the marginal efficiency of the slice you cut, not the baseline value of the channel. Rockerbox's glossary makes the same point about heavy-ups: they cannot establish whether baseline spend is incremental.

The calibration column comes from Meridian's guide to setting priors from experiments. Its calibration builder is designed for experiments that measure average lift against zero spend. Results from tests with a non-zero comparison point are accepted, but Google advises against them. In its words, they "introduce an estimand mismatch" with the model, which estimates every channel against a world with no spend.

This matters because diminishing returns are the normal case. The average return of a channel and the return on its last 20% are different numbers. Feed the second into a model that expects the first and you will understate the channel.

Does a bigger test budget buy a more sensitive test?

No, and this is the most common misreading.

GeoX's FAQ says the minimum detectable effect, the smallest percentage lift the test can tell apart from noise, depends only on how volatile your regions are and how well they match. The troubleshooting guide adds that changing the budget does not alter that limit.

What budget changes is the efficiency the test can detect. Google calls this the design-implied cost per incremental conversion. If your KPI is revenue, it is one over incremental return on ad spend (iROAS). A summary of the GeoX paper on Zenn gives the relation: required budget equals the detectable lift, times the baseline, times the cost per incremental unit. Rearranged, budget divided by the detectable incremental revenue gives the most expensive result the test can still see.

Google's rule of thumb follows. The design-implied cost should not be lower than the cost you expect the channel to achieve. If it is, the test is underpowered for that channel.

A hypothetical sizing example

The numbers below are invented to show the arithmetic. Assume one revenue KPI and spend spread in line with revenue.

• Treatment regions hold 30% of revenue, the GeoX default cap. Over a 28-day test they would normally bring in $1,500,000.
• The design's minimum detectable effect is 4%. The test needs to see at least $60,000 of incremental revenue.
• The channel spends $500,000 per 28 days nationally, so $150,000 in the treatment regions.

Read the last two rows together with diminishing returns. A 20% change only reads clearly if that slice of spend returns at least $2 per $1. The slice you would add or cut is usually your least efficient money. So a small heavy-up on a mature channel will often end with no significant result. That does not prove zero effect. It means the test could not see an effect that size.

The go-dark row has a cost the table does not show. If the channel's true iROAS were 1.5, the treatment regions would lose about $225,000 of revenue while saving $150,000 of spend, before margin. That is the price of a clean average.

Should a geo test calibrate your MMM at all?

Measurement people disagree, and the disagreement is useful.

Aryma Labs argues against calibration. Experiments run for weeks, models span years. Tests isolate one channel, models capture interactions. Experiments report lift, models work in sales. Its advice is to use experiments to validate a model, not to adjust it. The post gives no figures of its own.

Recast takes the other side. Its view is that a well-run experiment is generally more trustworthy than a model alone. It says to apply each result only to the period it covered and let older tests count for less. Sellforte adds a qualifier. A calibrated model should not match the test exactly, because the test is the authority only for its own campaigns, dates and spend level.

Meridian's calibration module answers two of Aryma's objections mechanically. It widens the uncertainty of tests that ended long ago, using a 52-week half-life. It scales short tests up for effects that land after the test, and widens their uncertainty too. It also widens tests run at a small fraction of the channel's normal daily spend. Aryma's third concern, that a test measures something different from the model, is the one Google itself concedes for marginal tests.

Our reading: calibrate with go-dark and holdback results, run near normal spend, on the same channel definition your model uses. Use heavy-up and go-dim results to check the model's response curve and to make the budget call. Haus makes the matching point from the vendor side. A test of Meta as a whole will not map cleanly onto a model that splits Meta into prospecting and retargeting.

How long should the cooldown be?

Longer than the minimum Google suggests in one place. The GeoX FAQ says one to two weeks. Its implementation page says two weeks or more. Vendor data favours the longer end.

Across 190 YouTube geo tests, Haus reports that iROAS rose by 79% on average during a post-test window that averaged 13 days. In 640 Meta tests, the average gain after the test was 26%. These are one vendor's averages across its own customers, and Haus says results for similar brands can differ three to four times. The direction still holds: a short cooldown understates video and upper-funnel media. For more on how Google counts delayed YouTube effects, see Demand Gen view-through conversions.

A seven-step method

1. Write the decision first. "Is this channel worth keeping?" needs a go-dark. "Should we add 30%?" needs a heavy-up. "Should we launch?" needs a holdback.
2. Let the model pick the channel. Meridian's model health report scores each channel from 0 to 100. Channels under 67.5 are recommended for testing. An implausibly high return is flagged when spend share times ROI exceeds 20.
3. Check feasibility. You need at least 10 regions, daily data and roughly three times the test length in history. Seasonal businesses should have a year. The design should reach an R-squared of 0.8 and pass the A/A check.
4. Size it. Compare the design-implied cost with the cost you expect. If the test cannot see a plausible result, run longer or treat more regions. Do not just add budget.
5. Set up cleanly. Use Physical Presence location targeting, lock the region lists, start every platform on the same day and keep the campaign types aligned with your model's channels.
6. Run three to ten weeks, then a cooldown of at least two weeks.
7. File the result by type. Average-lift results go into calibration. Marginal results go into the budget decision and the response curve check. For reporting in Meta alongside this, see Meta incremental attribution or 7-day click.

When is GeoX the wrong tool?

When you have one market, weekly data or a single Google campaign to judge. For that last case, Google Ads Conversion Lift is moving to a Bayesian method with a $5,000 minimum spend, though Google says most accounts still use the older method. It needs at least 150 conversions in the treatment group and 65 in control to report. That covers one platform, not your whole mix. If your question is about one Search campaign's settings, a marginal-cost check like the one in AI Max auto-upgrade is cheaper than any experiment.

The line that matters

A geo test result is only as useful as the question it was designed to answer. Average and marginal lift are both real numbers. They belong in different places.

Paid media measurement and AI visibility are converging on the same discipline: test, then believe. If you want to see where your brand stands in AI answers, the waitlist is open.

Questions

What is the difference between a go-dark test and a heavy-up test?

A go-dark test turns a live channel off in the treatment regions while the control regions carry on as normal. It shows what the channel contributes at current spend. A heavy-up test adds spend in the treatment regions. It shows what the extra money returns. The first measures an average, the second a margin.

Does a bigger budget make a GeoX test more sensitive?

No. Google's GeoX documentation says the minimum detectable effect depends on how volatile and how well matched your regions are. Budget changes the efficiency the test can detect. A small budget can only confirm a very efficient campaign. To lower the detectable effect, run longer, treat more regions or use a higher-volume metric.

Can I use a heavy-up test result to calibrate Meridian?

Meridian accepts it, but its documentation advises against it. The calibration builder is designed for experiments that measure average lift against zero spend. A heavy-up or partial go-dark result measures marginal lift, which answers a different question. Use it to check the model's response curve and to make the budget call.

How many regions and how much data does a GeoX test need?

Google's FAQ sets a minimum of 10 regions for a single-cell test and recommends 50 to 100 or more for good matching. Data must be daily. Its sample design runs 28 days. Seasonal businesses should prepare at least a year of history. Google suggests one to two weeks of cooldown in one place and two or more in another. Plan for at least two.

Sources

1. Google, Reaching New Heights in Marketing Measurement: The Next Era of Meridian business.google.com
2. GitHub google/meridian, Meridian 2.0: Introducing Meridian GeoX and Automated Calibration github.com
3. Google for Developers, Meridian GeoX FAQs developers.google.com
4. Google for Developers, Meridian GeoX types of experiments developers.google.com
5. Google for Developers, Meridian GeoX troubleshoot design output issues developers.google.com
6. Google for Developers, Meridian GeoX implementation of geo testing developers.google.com
7. Google for Developers, Meridian GeoX API reference developers.google.com
8. Google for Developers, Set custom ROI priors using past experiments developers.google.com
9. Google for Developers, Channel calibration recommendation and score developers.google.com
10. Google Ads Help, About Bayesian methodology in Conversion Lift support.google.com
11. PPC Land, Google's Meridian GeoX exits beta claiming 31% cheaper geo experiments ppc.land
12. Search Engine Journal, Google Launches Meridian GeoX Globally searchenginejournal.com
13. Zenn, Understanding How Meridian GeoX Works from the Paper zenn.dev
14. Haus, Do YouTube Ads Perform? Lessons From 190 Incrementality Tests haus.io
15. Haus, The Meta Report: Lessons from 640 Haus Incrementality Experiments haus.io
16. Aryma Labs, Why Feeding Experiment Results Back into MMM is the Biggest Scam in Marketing Measurement arymalabs.substack.com
17. Recast, How to Use Lift Tests to Calibrate an MMM getrecast.com
18. Sellforte, Why can a calibrated MMM show a different incremental ROAS than a conversion-lift study? sellforte.com
19. Animo, Google's Meridian GeoX Just Went Live: What It Actually Changes animo-tech.com
20. Rockerbox, Incrementality testing glossary help.rockerbox.com
21. Haus, Why incrementality testing belongs with your MMM and where to start haus.io

NEXT STEP

See where your brand stands.

Your Digital Authority Score, who AI recommends for your target questions, and the placements that close each gap. Your first report is free.

Get your free report →