Incrementality testing: what marketers need to know in 2026

November 5, 2021

Tin Le

Updated on October 17, 2026: Adtek rewrote this entire article with the latest data, platform features and regulations from 2025 to 2026.

Incrementality testing measures the share of results that happen only because of your ads. Here is how it works: you keep one group of users or one group of regions from seeing the ads (the control group, or holdout), compare their results with the group that still sees the ads, and the difference is the lift. Ads Manager and Google Ads reports will not give you this number, because they also count orders from customers who would have bought anyway without any ad.

In 2026, incrementality testing is no longer something only large corporations can afford. Google says a test that could once cost more than USD 100,000 now has a minimum threshold of USD 5,000. Meta has published that, among the advertisers it studied, 31% of the incremental conversions Meta generated were wrongly credited to other channels by rule-based attribution models.

How is incrementality testing different from the numbers in your ads manager?

An attribution model answers the question “which ads did this conversion touch?” An experiment answers a different question: “without the ads, would this conversion still have happened?” The two numbers drift furthest apart in remarketing and branded keyword search, where the people seeing the ad were already planning to buy.

A hypothetical example: a cosmetics shop runs remarketing to 1,000,000 people and holds back 20% as the control group.

Metric	Group that sees ads	Control group
People	800,000	200,000
Orders	4,000	800
Conversion rate	0.50%	0.40%

Lift = (0.50% – 0.40%) / 0.40% = 25%. The number of incremental orders is 4,000 minus the orders that the 800,000-person group would have generated on its own at a 0.40% rate, which is 4,000 – 3,200 = 800 orders.

Now assume the spend was VND 160 million and the average order value is VND 500,000. If the platform claims all 4,000 orders, the reported ROAS is 12.5 and the CPA is VND 40,000. Calculated on the 800 incremental orders, the iROAS drops to just 2.5 and the incremental CPA rises to VND 200,000. A decision to double the budget based on 12.5 looks very different from one based on 2.5.

One caution: if the two groups are different sizes, like the 80/20 split above, you have to convert to a per-person rate before subtracting. Do not subtract total revenue directly.

If the lift is positive, ask the next question: is the incremental CPA lower than the LTV or the gross profit of one customer? A lift close to zero means the campaign is paying for orders that would have come in organically, so change the audience, change the message, or cut it. Negative lift is rare and usually comes from excessive frequency or a test that was set up incorrectly.

User-level holdouts and geo experiments

User-level holdouts

The platform randomly splits users into a group that sees the ads and a group that does not, then tracks conversions for both. This method is precise and runs right inside your ad account. The catch is that it only measures what the platform can see (the pixel, Conversions API, uploaded conversions), and each platform grades its own homework. Techniques for filling the control group’s ad slots, such as ghost ads or PSA ads, now sit inside the Conversion Lift tools themselves, so marketers rarely have to build them by hand.

Geo experiments

Instead of splitting people, you split regions. In one group of provinces you turn ads off (go-dark), keep them as they are, or raise the budget sharply (heavy-up), and then compare sales against a group of similar provinces. This approach measures real revenue from your sales systems, including physical stores, marketplaces and hotlines.

Google offers geography-based Conversion Lift, which splits by Google Marketing Areas to limit cases where someone sees an ad in one area but buys in another. The beta only supports holdback and go-dark, campaigns must target a single country, and you need to go through a Google representative. If you cannot use the platform tool, Google’s open-source Meridian GeoX library lets you design and analyze geo experiments yourself, including tests that compare several treatment groups against one shared control group.

Vietnam has one detail that needs recalculating from 2025 onward. Under Resolution 202/2025/QH15 of the National Assembly, the country now has 34 provincial-level administrative units, operating from July 1, 2025. Fewer units means fewer pairs of provinces to match, and each new province is larger: Ho Chi Minh City now includes the former Binh Duong and Ba Ria-Vung Tau provinces. If most of your orders come from Hanoi and Ho Chi Minh City, you cannot put either one in the control group and still get a fair comparison. In that case, run the test in the remaining provinces or use a user-level holdout.

Conversion Lift on Meta and Google Ads in 2026

In 2025, Google Ads moved Conversion Lift to Bayesian statistics. It now uses data from past campaigns as a “prior,” so tests need fewer conversions and less budget. In November 2025, Google announced the USD 5,000 threshold and said the share of studies that reach a conclusion had increased by up to 50%. Meta opened self-serve Conversion Lift in the Experiments section for eligible advertisers, and brought lift data into delivery through its incremental attribution setting.

	Google Ads Conversion Lift	Meta Conversion Lift
Published requirements	At least 1,000 observed conversions and a minimum study budget of USD 5,000	No fixed budget threshold found in public documentation; self-serve for eligible advertisers
Data sources	Conversion tags, enhanced conversions for web and leads, consent mode	Conversions API (recommended by Meta), Meta Pixel, App Events API
Duration	Minimum 7 days, more than 14 days recommended	Depends on the length of the purchase cycle
Reading the results	Certainty levels from 50% to 95%; below 90% the result is directional	Single-cell to compare publishers, multi-cell to compare strategies
Notes	Not available to every account; you need to contact a Google representative	Incremental attribution optimizes for incremental conversions directly in Ads Manager

According to Google’s documentation, the control group can be as large as 50%: the larger it is, the faster you get results, but the more orders you give up. Studies shorter than 14 days in industries with long conversion delays can understate absolute lift by up to 17%. You need at least 150 conversions in the test group and 65 in the control group. Google also advises measuring all campaigns of the same type together, because people in the control group still see ads from campaigns outside the study.

For incremental attribution, Meta reports an average performance improvement of 46%, based on 37 Conversion Lift studies from 30 advertisers that ran from July to October 2024. This is a figure Meta measured itself. When you turn the setting on, reported conversions usually fall, because Meta drops orders it believes would have happened without the ads. A CPA that looks worse does not necessarily mean the campaign got worse.

How incrementality testing fits with MMM and multi-touch attribution

The three tools run on three different rhythms. Multi-touch attribution gives you daily numbers for each campaign and is used for fast optimization. MMM (marketing mix modeling) uses weekly sales data over 2 to 3 years, accounting for seasonality, pricing and offline channels, and is used to split budget by quarter. Incrementality testing gives you causal evidence, but only for one channel over one period.

A Kantar study cited by Meta (May 2025, 1,935 decision makers) found that advertisers use an average of 3.8 measurement solutions, and 55% have run into conflicting results between those solutions. Meta and Google both propose the same fix: calibration, which means using experiment results as the benchmark for adjusting the other two tools.

Calibrating attribution: if reports record 1,000 orders and the experiment shows 600 incremental orders, multiply that channel’s reported numbers by 0.6 until the next test. It can go the other way too: Meta found that at the median, rule-based attribution models need a multiplier of at least 1.45x to match Meta’s true incremental value.
Calibrating MMM: feed experiment results in as a prior for the channel’s ROI coefficient. Meridian (Google) and Robyn (Meta), two open-source MMMs, both support this.

Experiments do not need to follow each person across multiple channels, so they are the least affected as cookies and mobile identifiers become harder to use, an issue Adtek covered in the article marketing analytics in a world without cookies.

Adtek’s view: designing a test for a mid-budget business

For a Vietnamese business spending a few hundred million dong a month on advertising, the goal of the first incrementality test is a conclusion reliable enough to make one budget decision. This is the process we recommend:

1. Write down the decision before you write the hypothesis

For example: “If Meta remarketing has an iROAS below 2, move 30% of the budget to non-branded search.” Without a decision attached, the test just produces one more report.

2. Pick the campaign most likely to “eat” organic orders

Remarketing, branded keywords and campaigns targeting existing customers usually show the biggest gap between reported numbers and real numbers, so they are where wasted money is easiest to find.

3. Choose a KPI close to revenue but with enough volume

Orders are best. If you only get a few dozen orders a week, add a mid-funnel KPI such as add-to-cart or form submissions as a secondary metric, as Google suggests.

4. Check tracking and the legal basis for your data

Conversions API and enhanced conversions send customers’ hashed emails and phone numbers to the platform. That is personal data under Vietnam’s Personal Data Protection Law No. 91/2025/QH15, in effect from January 1, 2026, so you need proof that customers have consented before you turn these features on.

5. Choose the method that fits your budget

If you have USD 5,000 and 1,000 conversions on Google, or Meta shows you as eligible in Experiments: use the platform’s Conversion Lift.
If not: run your own geo experiment. Match 6 to 10 provinces into pairs with similar sales over the last 8 weeks, and turn ads off in one province of each pair. Check whether the location lists in your ads manager have been updated to the new administrative boundaries.
Last resort: switch ads off in alternating weeks. This is easily distorted by seasonality and competitors, so treat it only as a directional signal.

6. Pick a clean time window

Avoid Tet (Lunar New Year), 11.11, 12.12, Black Friday and your own big promotions. If you are working on an 11.11 plan, make sure the test ends before the sale-season campaign starts, or push it to after Tet. Run it for at least 2 weeks, and longer for high-priced products.

7. Read the results by the rule you set

Do not edit the campaigns while the test is running. When the results come in, apply the rule from step 1 exactly, even if the number is not what you hoped for. Save the calibration factor for your weekly reports, and schedule the next test for 6 months later.

A quick checklist before you hit run:

Which budget decision will change based on this result?
Is the control group truly not seeing ads from other campaigns of the same type?
Do any promotions, product launches or sale seasons fall within the test period?
Does the customer data you send to the platform have valid consent?