How incrementality testing works, and what Google's $5,000 Conversion Lift minimum changes

Platform dashboards report the conversions that followed an ad. They do not show how many of those conversions would have happened without it. Incrementality testing is the method built to estimate that difference.

The cost of entry has fallen. Google announced on November 11, 2025 that it had cut the minimum experiment budget for Conversion Lift from a level approaching $100,000 to $5,000, according to trade publication PPC Land. On October 7, 2026, Search Engine Land reported that Search and Performance Max campaigns now support the user-based version of the test. This article covers how the tests work, what Google requires, and how results feed marketing mix models.

How a holdout test separates caused sales from baseline sales

Search Engine Land describes Google's Conversion Lift as creating a holdout group that is not exposed to the measured campaigns. The behaviour of that group is compared with users who can see the ads. The difference estimates the incremental conversions.

The conversions in the holdout group are the baseline: people who converted without seeing the campaign. Conversions beyond that baseline, among exposed users, are the ones the test credits to the ads.

Holdout size is a trade-off. Search Engine Land notes that a larger holdout can improve the sample but withholds ads from more users. PPC Land adds that control groups carry an opportunity cost in lost conversions.

What geo-based tests do differently

A geo test divides the comparison by place rather than by user. Meta Open Source describes GeoLift as an open-source tool for calculating lift at a geographic level. The project's page also states that GeoLift is not a Meta product and is intended for research purposes only.

The pages read for this article do not set out GeoLift's method, data requirements or recommended test length. Those details sit in the project's documentation and repository, which were not reviewed here.

None of the pages read for this article covers a geography-based version of Google's Conversion Lift, so its terms are not described here.

What Google's $5,000 minimum and other requirements mean

PPC Land reports the minimum experiment budget fell from near $100,000 to $5,000. The same article gives conflicting dates for when the $5,000 figure first appeared: it cites Think Week in September 2025 and also May 2025, alongside the November 11, 2025 announcement. The timeline is therefore not settled in that source.

The sources describe $5,000 as a budget minimum. They do not say it is a monthly figure, so the phrase "$5,000 a month" should not be assumed. Search Engine Land states that participating campaigns need at least $5,000.

Search and Performance Max have a second hurdle. Search Engine Land lists at least 1,000 observed conversions, excluding conversions that use supplementary data, plus at least one conversion action compatible with Conversion Lift. Availability varies by account, and alpha and beta campaign types may still need Google's help.

In November 2025, PPC Land reported that Display, Search, Shopping and Performance Max required contacting an account representative. The October 2026 report moves Search and Performance Max to self-serve. The later report says account availability can vary and that Conversion Lift still isn't available to every advertiser.

How long a test should run and how feasible it is

PPC Land reports that studies can run as short as seven days, and that Google recommends at least 14 days, especially where conversions take longer to happen. For businesses with longer conversion lags, studies under 14 days saw Absolute Lift fall by up to 17%.

Google also shows feasibility ratings with budget recommendations. PPC Land reports a 60 to 90% chance of a conclusive result for a High rating and 0 to 30% for Low. These are Google's estimates, not guarantees.

The choice of conversion action matters. Per PPC Land, Google says upper- and mid-funnel actions such as pageviews, lead forms and add-to-cart show higher lift than purchases, and recommends measuring several conversion types. A campaign cannot join more than one study at a time.

PPC Land says Google's data shows most advertisers run about one to two studies per year. It also reports that Google says its improved statistical models produce up to 50% more conclusive results. That figure is Google's own claim.

The conversions in the holdout group are the baseline: people who converted without seeing the campaign.

Reading lift next to platform-reported ROAS

The two measures answer different questions. Platform-reported figures count conversions linked to ads. A lift test estimates the share of conversions that exposure caused, by comparing against users who did not see the campaign.

PPC Land reports that data-driven attribution and conversion-based or conversion-value bidding are associated with higher measured lift. It lists Target CPA, Maximize Conversions and Target ROAS as examples. The results appear at product or brand level in the Lift Measurement table.

The sources read for this article do not say how to convert a lift result into a corrected ROAS figure. Each result applies to the campaign, conversion action and dates tested.

How lift results calibrate marketing mix models

PPC Land describes Meridian as Google's open-source Bayesian marketing mix model, released in January 2025. It reports that Meridian can incorporate incrementality results from any channel or method to calibrate its models. The article frames the MMM as a broad view of all channels and external factors, and experiments as precise causal data for specific campaigns.

Google's Meridian documentation says priors can come from past incrementality experiments, such as lift tests or geo-experiments. They can also come from industry benchmarks, previous MMM results and internal expertise.

Meridian lets users set priors on business metrics such as ROI, mROI and contribution percentage. The standard deviation expresses confidence, where a smaller value means higher confidence. The documentation gives the example of a 95% chance that ROI falls between 2.0 and 6.0.

Informative priors act as a form of regularisation. Without past experiments or business intuition, Google advises starting with Meridian's default priors rather than flat ones, which can give unreliable estimates. It says choosing the right strength is usually iterative and involves checking out-of-sample fit. The page does not specify how strongly a prior should be weighted against the data.

PPC Land also notes that the Bayesian approach incorporates prior knowledge about expected outcomes and allows meaningful conclusions with smaller sample sizes. The documentation page reviewed does not detail how to convert an experiment result into a prior.

Related reading: How Google Ads' Smart Bidding decides how much to bid in every auction; How the learning phase decides when a Meta Advantage+ campaign is ready.