Home / Blog / How to Run an Incrementality Test: A 10-Step Geo-Holdout Che...
Industry Insights2026-11-1811 MIN READ

How to Run an Incrementality Test: A 10-Step Geo-Holdout Checklist (2026)

By Ryan Roberts, founder of North American Media Experts

To run an incrementality test, hold out a matched set of regions or users from your ads, keep everything else equal for at least 14 days, then compare outcomes against the exposed group. The gap is the lift your ads actually caused. Google now runs these experiments from a $5,000 minimum, so most advertisers can afford one.

Platform dashboards report what your ads touched. An incrementality test reports what your ads caused. Those are different numbers, and the gap between them is usually where the wasted budget lives. Retargeting, branded search and view-through CTV conversions in particular tend to claim credit for sales that were going to happen anyway.

Marketers know it. In the Haus 2026 Marketing Decision Confidence Index (n=500 US decision-makers, fielded January 2026), 60% said they trust independent incrementality testing most, ahead of media mix modeling at 40% and in-platform reporting at 37%. Yet the IAB and BWG Global State of Data 2026 report found that three in four marketers say their attribution, incrementality and MMM programs are not delivering the speed, accuracy or trust they need.

This checklist is the order we use to set up a test that produces a number a CFO will accept. It assumes no data science team. If you can export daily sales by region and pause spend in a few markets, you can run it.

Before you start: three things you need

If you are still deciding which measurement method fits your business at all, start with our comparison of last-click, data-driven attribution and MMM, then come back here.

Step 1: Write down the one decision the test will make

A test without a decision attached is a science project. Before touching any settings, write one sentence: "If the lift is above X, we will Y. If below, we will Z." For example: "If CTV drives an incremental ROAS above 2.5, we move $15,000 a month from retargeting into CTV. If not, we cut CTV to a brand-only flight."

That sentence forces you to choose a single channel, a single KPI and a threshold that matters to the business. It also stops the most common failure we see: a test that comes back ambiguous and gets read however the channel owner prefers.

Step 2: Pick the right test design

There are three practical designs. Choose based on your channel and your footprint, not on what sounds most rigorous.

DesignHow it worksBest forWatch out for
Platform conversion liftThe platform randomly withholds ads from a share of users and compares conversionsMeta, YouTube, Demand Gen and other single-platform questionsMeasured by the platform being tested; only covers that one channel
Geo-holdout (matched markets)You pause spend in selected regions and compare them with a synthetic control built from untouched regionsCTV, DOOH, programmatic, search, and any cross-channel questionNeeds enough regions and stable regional sales history
Time-based on/offYou switch a channel off for a period and compare against the forecastSmall, single-market businessesSeasonality and promotions easily contaminate the read

For most advertisers we work with, the geo-holdout is the default. It uses your own sales data, it works for channels that cannot track individual users, such as connected TV and digital out-of-home, and it lets you test more than one platform at once. Our DOOH foot-traffic attribution benchmarks show why this matters: the vendor lift numbers for screen media vary widely, and a geo test is how you find out which end of the range you sit at.

Step 3: Choose the KPI and confirm it moves

Use the metric closest to money that you can measure by region every day. For ecommerce that is orders or revenue. For lead generation it is qualified leads, not form fills. For retail it is store sales or visits.

Then look at its daily volatility. If a region's sales swing 30% week to week for no reason, a 5% lift will be invisible. Noisy KPIs need longer tests, more regions or a bigger holdout.

Step 4: Select test and control markets

The test regions (where you change spend) and the control regions (where you leave it alone) must have moved together historically. Pick them with data, not by instinct.

Two to five test markets is a common starting point. GeoLift's own walkthrough evaluates test groups of that size.

Step 5: Run a power analysis before you spend a dollar

A power analysis tells you whether the test can detect the lift you care about. Skip it and you risk spending four weeks of budget on a result that says "inconclusive."

Platform tools do this for you now. Google's lift experiments give a feasibility rating before launch; according to PPC Land's November 2025 coverage, a "High" rating carries a 60% to 90% chance of a conclusive result and a "Low" rating 0% to 30%. Do not launch on a Low rating. For a geo test, GeoLift's power functions simulate the smallest detectable lift for each combination of markets, duration and budget.

If the minimum detectable lift is larger than any lift you could plausibly see, change the design: add markets, extend the duration or increase the spend difference.

Step 6: Set the duration to cover a full purchase cycle

Run for at least 14 days. Google allows seven-day lift tests but recommends 14 as the minimum, and its analysis found that shorter tests under-reported absolute lift by up to 17% for advertisers with long conversion lags. EMARKETER's guidance is three to four weeks with properly sized holdouts, and geo tests run by Haus on TikTok averaged 21 days.

The GeoLift rule of thumb is simple: the test should contain at least one full purchase cycle. A B2B advertiser with a 45-day sales cycle should not read a 14-day test on closed revenue. Measure a leading indicator, such as qualified pipeline, or run longer.

Step 7: Lock everything else down

During the test, the only difference between test and control markets should be the spend you changed. That means no new promotions in one group and not the other, no creative swaps mid-test, and no budget reshuffling on other channels that targets the same regions.

Write down the start and end dates, the markets, the spend in each, and anything outside your control that happened, like weather or a competitor sale. You will need that log when someone challenges the result.

Step 8: Read the result as a range, then calculate iROAS

A good test reports lift with a confidence interval, not a single number. Report both. Then convert lift into incremental ROAS (iROAS): incremental revenue divided by the spend that produced it.

Here is a worked example using illustrative numbers. Test markets received $12,000 of spend and delivered 1,240 orders. The synthetic control says they would have delivered 1,000 orders without the ads. That is 240 incremental orders, a 24% lift. At a $120 average order value, incremental revenue is $28,800, so iROAS is $28,800 divided by $12,000, or 2.4. If the platform reported 600 conversions for the same spend, it was claiming 2.5 times more credit than the ads earned.

Compare iROAS with your break-even, which is 1 divided by gross margin. At a 40% margin, break-even is 2.5, so the channel in this example is close to paying for itself but not quite. Platform ROAS would have said it was a clear winner. For context on what platforms report, see our average ROAS benchmarks by industry and channel.

If your CTV or programmatic numbers look too good to be true, they usually are. Get a free paid media audit and we will show you which of your channels are worth testing first and how to size the holdout.

Step 9: Make the decision you wrote down in Step 1

Go back to the sentence. If lift cleared the threshold, move the budget. If it did not, cut or restructure. If the interval straddles the threshold, the honest answer is "not proven," and the next step is a longer or larger test, not a judgment call.

Then use the result to correct your everyday reporting. If platform-reported conversions overstate incremental conversions by 2.5 times, divide platform numbers by 2.5 when you plan budgets until the next test. Teams building a media mix model should use test results to calibrate it; our guide to MMM without a data science team covers how.

Step 10: Put the next test on the calendar

Incrementality is not a one-time audit. Channel effects change as creative fatigues, competitors shift budget and platforms change their algorithms. The IAB's 2026 data shows the industry moving from testing a few times a year to always-on measurement. A practical cadence for a mid-sized advertiser is one major channel test per quarter, rotating through your biggest line items, plus a retest after any large budget change.

Start with the channel where platform reporting and business results disagree most. For most of the accounts we audit, that is retargeting or connected TV. Our 2026 CTV benchmarks for CPM, VCR and ROAS are a useful sanity check on what platform-reported CTV performance should look like before you test it.

The 10-point incrementality test scorecard

Score one point for each item you can answer yes to before launch. Eight or more and the test is ready. Below six, fix the gaps first.

#CheckPass if
1Decision writtenOne sentence states the threshold and the budget move it triggers
2Design chosenConversion lift, geo-holdout or on/off, chosen to fit the channel
3KPI chosenBusiness-system metric, measurable by region daily
4History availableAt least 90 days of daily regional data
5Markets selectedTest and control chosen by historical fit, anomalies excluded
6Power checkedMinimum detectable lift is below the lift you expect
7Duration setAt least 14 days and one full purchase cycle
8Confounders lockedNo promotions, creative changes or budget moves that differ by group
9Readout plannedLift reported with a confidence interval and converted to iROAS
10Next test bookedFollowing test on the calendar before this one ends

Common mistakes to avoid

Methodology and sources

Figures in this guide come from the Haus 2026 Marketing Decision Confidence Index (fielded January 2026, n=500, reported by EMARKETER), the IAB and BWG Global State of Data 2026 (reported by MarTech, February 2026), EMARKETER's 2026 incrementality FAQ, PPC Land's November 2025 coverage of Google's lift-experiment changes, and Meta's open-source GeoLift documentation. The iROAS example uses illustrative numbers to show the calculation and is not client data.

Frequently asked questions

How long should an incrementality test run?

At least 14 days, and longer if your purchase cycle is long. Google recommends a 14-day minimum for its lift experiments and found that shorter tests under-counted lift by up to 17% when conversions lag. EMARKETER advises three to four weeks, and Haus TikTok geo tests averaged 21 days.

How much budget do I need to run an incrementality test?

Less than you might think. In November 2025 Google cut its minimum spend for incrementality experiments to $5,000, down from close to $100,000. For a geo-holdout test you run yourself, the real constraint is statistical power: you need enough conversions in the test regions for the expected lift to show above normal week-to-week noise.

What is the difference between incrementality testing and attribution?

Attribution assigns credit for conversions that already happened to the touchpoints that preceded them. Incrementality testing measures how many conversions would not have happened without the ads, by comparing an exposed group to a held-out control. Attribution tells you where conversions were touched; incrementality tells you what the ads actually caused.

What is a good iROAS?

It depends on your margin. Incremental ROAS is incremental revenue divided by test spend. Calculate your break-even iROAS as 1 divided by your gross margin: a business with a 40% gross margin breaks even at 2.5. Anything above break-even is a channel paying for itself on a causal basis, which a platform-reported ROAS cannot tell you.

Can a small business run a geo-holdout test?

Yes, if it sells across enough regions with stable sales history. Meta's open-source GeoLift tool works from a simple file of location, date and sales, and its walkthrough uses 90 days of pre-test data across 40 markets. A business selling in only one or two cities is better served by a platform conversion lift study or an on/off time-based test.

Next steps

Pick your biggest channel, write the decision sentence and pull 90 days of regional sales this week. That alone will tell you whether a geo test is feasible. If you would rather have it designed for you, our programmatic team sets up and reads incrementality tests across CTV, DOOH and display, and our paid search and paid social teams do the same for Google and Meta, with no minimum spend and no retainer fees.

Want this applied to your own budget? Get a free media plan built around your goal and budget, delivered within one business day. No pitch. Or book a call with Ryan.

Keep reading

All articles

Know where your next dollar should go.

Tell us your goal and budget. We send a channel-by-channel media plan within one business day. Free, no pitch.