Incrementality Testing: How to Prove Which Ad Spend Actually Works

Suppose Meta reports 1,800 retargeting purchases and Google reports

1,200 branded-search purchases for the same month. Those counts can

overlap, and attribution does not establish that an ad caused a

purchase. Compare them with backend outcomes before using them to

justify a budget.

This guide covers incrementality-test designs, Meta's and Google's

published eligibility rules, the calculations from conversions to iROAS,

and how to control campaign changes during a test.

What is incrementality testing?

**Incrementality testing is a controlled experiment that withholds ads

from a comparable control group and compares its outcomes with a test

(or treatment) group that could see the ads.** The difference between

the two groups is the incremental lift: the conversions your ads caused,

as opposed to conversions that would have happened anyway.

Attribution allocates credit across touchpoints. An incrementality test

estimates a counterfactual: the outcomes a comparable audience would

have produced without the tested ads. Compare that baseline with the

treatment outcome to estimate causal lift.

Say a Meta campaign holds 10% of its audience out of delivery for four

weeks. If the test group buys at 0.30% and the control at 0.24%, the

incremental lift is 25%, and 0.24% estimates what the test group would

have done without ads.

**Alt text: Incrementality testing diagram: an audience split into a

test group that can see ads and a control group that cannot, with the

gap between their conversion rates labeled incremental lift.**

Why platform-reported conversions overstate what your ads do

A platform-reported conversion means someone converted inside the

platform's attribution window after an ad interaction, not that the ad

changed anything. Windows can include view-through credit (Meta's lift

eligibility rules refer to a 1-day view setting), so someone who

scrolled past an ad and bought later that day counts even if they were

already headed to checkout. Each walled garden also sees only its own

touchpoints, so one purchase can be claimed by Meta, Google and LinkedIn

at once. [Attribution models such as last click and

data-driven](https://synterai.com/blog/marketing-attribution-models)

change how credit is split, not whether the sale was caused.

In [field experiments at

eBay](https://www.nber.org/papers/w20171), Blake, Nosko and

Tadelis found no measurable short-term benefit from brand-keyword ads.

Their non-experimental estimates also overstated overall paid-search

returns, and much of the non-brand spend reached frequent shoppers whose

purchases did not change. These were results for eBay's tested

campaigns, not a universal verdict on search advertising.

Retargeting can produce incremental purchases: [Johnson, Lewis and

Nubbemeyer's 2017 retailer

experiment](https://journals.sagepub.com/doi/10.1509/jmr.15.0297)

estimated a 10.5% purchase lift from display retargeting. A separate

[2019 study by Gordon and

colleagues](https://pubsonline.informs.org/doi/10.1287/mksc.2018.1135)

compared observational estimates with 15 randomized Facebook experiments

and found that the observational methods often did not recover the

experimental effects.

Incrementality testing vs. A/B testing, attribution and MMM

**A typical ad A/B test compares two versions of something, such as two

creatives or bid strategies, and both groups still see ads.** It tells

you which version wins, which is what [A/B testing ad

creative](https://synterai.com/blog/creative-testing-automation-guide)

is for, not whether advertising worked. An incrementality test changes

exposure itself: ads versus no ads, or more spend versus less.

Multi-touch attribution allocates credit across tracked touchpoints; MMM

estimates channel contributions from aggregate data. Experiment results

can inform those estimates. For example, Meridian can use incrementality

results to [calibrate channel ROI

priors](https://developers.google.com/meridian/docs/advanced-modeling/roi-priors-and-calibration).

Our [MTA-versus-MMM

comparison](https://synterai.com/blog/mta-vs-mmm) covers

the differences between the methods.

Method Question it What it needs Main blind spot

answers

• ----------------- ----------------- ----------------- -----------------

A/B test Which version Two variants, Says nothing

(creative or performs better? random split, about whether ads

setting) both exposed to drove the outcome

ads

Incrementality What did the ads A comparable One channel,

test cause? control group spend level, and

with no ads (or period at a time

less spend) and

enough volume

Attribution (MTA) Which touchpoints User-level Credits

were in the path? tracking conversions that

would have

happened anyway

MMM How much did each A long history of Correlational

channel aggregate spend unless calibrated

contribute over and outcomes with experiments

time?

Types of incrementality tests for paid media

Platform conversion lift studies

**Meta incrementality testing runs through Conversion Lift, under

Experiments in Ads Manager**, at no additional cost. As of October 2026,

Meta's eligibility guide asks for a campaign started in the past year

with at least USD 5,000 in spend and 500 conversions, and self-serve

tests add a signal-quality bar: Conversions API with an event match

quality score above 5, or a supported source sending a lower-funnel or

app-install event ([Meta Business

Help](https://www.facebook.com/business/help/221353413010930)

Center). Eligibility is only a floor: for a well-powered test, Meta

recommends at least USD 10,000 in spend, a 10% holdout and more than 500

conversions during the test, and its setup page suggests at least 28

days.

**Incrementality testing in Google Ads also uses Conversion Lift, which

isn't available to every account; you request it through a Google

representative**. User-based studies need at least 1,000 observed

conversions and a USD 5,000 minimum campaign budget ([Google Ads

Help](https://support.google.com/google-ads/answer/12005564)).

Google's help pages also disagree on campaign types (the setup page

lists Search and Performance Max as supported; the comparison page

routes them through your rep), so confirm with yours.

These studies split people rather than regions. Use that design for an

eligible single-platform question; it does not measure the combined

contribution of other platforms.

Geo holdout and geo lift tests

A geo test switches spend off (a holdout) or up (a lift test) in some

regions and compares them with unchanged regions. Because the unit is a

region, it works across platforms and can measure sales only your

systems see, including offline. Google's geo-based Conversion Lift

handles offline conversions but needs campaigns targeting a single

country and usually a bigger budget than the user-based study.

A synthetic control builds the counterfactual from a weighted mix of

untreated regions using pre-test history, the method behind Meta's

open-source

GeoLift.

Its docs recommend 4 to 5 times the test length in stable pre-period

data, at least 25 pre-treatment periods across 20 or more geo units, and

a minimum of 15 days on daily data. Meta still recommends people-based

tests where possible because they have more statistical power.

**Alt text: Geo lift test line chart with illustrative data: daily

conversions in test regions track a synthetic control during the

pre-period, then rise above it once the test starts, with the gap shaded

as incremental conversions.**

The main watch-out is leakage: Google notes that people exposed in a

test region who convert in a control region reduce measured

incrementality.

Audience holdouts

For CRM lists, email and retargeting pools, you can build the control

yourself: randomly split the list before upload, send only the test half

to the platform, and compare both halves in your CRM. It's cheap, but

other campaigns can still reach the holdout, the same dilution Meta

warns about when competing ads run outside its own lift tests.

PSA and ghost ad designs

A PSA design buys media to show the control group a public service

announcement. Ghost ads instead record control users who would have

received the advertiser's ad. In Johnson, Lewis and Nubbemeyer's study,

this supported measurement under real-time delivery optimization and

achieved comparable precision with at least an order of magnitude less

spend than PSA and intent-to-treat designs.

What about pre/post comparisons?

Comparing four weeks before a change with four weeks after is weak

evidence, because seasonality, promotions, and the platform's own

learning land in the same window. Use it to pick what to test, not what

to fund.

Method How the control Best for Typical Watch-outs

group is built duration

• -------------- ----------------- ----------------- -------------- --------------

Platform Platform holds Single-platform Meta 28+ days; Spend and

conversion out part of the questions Google 14+ (7 conversion

lift audience allowed) minimums; one

platform only

Geo holdout or Matched regions Cross-platform 15+ days, one Leakage; needs

lift or synthetic and offline sales purchase cycle stable history

control

Audience You randomly Retargeting and One purchase Other

holdout split a list CRM lists cycle campaigns

reaching it

PSA or ghost Placebo ad, or Where a platform As the PSA costs

ads logged offers it platform study media

would-have-seen

users

Pre/post Your own history Hypotheses only n/a Confounded by

everything

else

How to calculate incrementality

The core formulas:

• Lift % = (test conversion rate - control conversion rate) ÷

control conversion rate

• Incremental conversions = test conversions - (control conversion

rate × test group size)

• iROAS = incremental revenue ÷ incremental ad spend

• **Cost per incremental conversion = incremental ad spend ÷ incremental

conversions**

In a holdout, incremental spend is all the test group's spend, since

the control gets none; in a spend-up geo test, it's only the extra

spend.

Meta's lift percentage divides incremental conversions by the

conversions it estimates the exposed audience would have had without

ads, and Google reports the last two as iROAS and iCPA. Some tools call

incremental ÷ test conversions "incrementality" instead; that's the

share of test-group conversions the ads caused, a different number.

Worked example (illustrative). A 28-day user-level holdout with a

10% control group, measured on backend purchases at a fixed \$150 order

value, in USD:

1. Test group: 900,000 people, 2,700 purchases (0.30%). Control:

100,000 people, 240 purchases (0.24%).

1. Lift = (0.30% - 0.24%) ÷ 0.24% = 25%.

1. Without ads, the test group would have made about 0.24% × 900,000 =

2,160 purchases, so 540 were incremental.

1. Incremental revenue = 540 × \$150 = \$81,000. On \$27,000 of ad

spend, iROAS = 3.0.

1. Cost per incremental purchase = \$27,000 ÷ 540 = \$50.

These are point estimates. The illustrative 95% interval puts

incremental purchases between about 249 and 831, corresponding to iROAS

of about 1.4--4.6. Although the interval is above zero, it crosses the

profitability threshold: at a 40% contribution margin (revenue less

variable costs such as goods, shipping and payment fees, divided by

revenue), break-even iROAS is 1 ÷ 0.40 = 2.5. Positive lift alone

therefore does not establish profitability.

iROAS vs. platform ROAS

Suppose the platform attributed 1,800 purchases to the same campaign.

Platform ROAS is 1,800 × \$150 ÷ \$27,000 = 10.0, more than three times

the iROAS. Platform ROAS measures revenue that followed the spend; iROAS

measures revenue the spend created. A good ROAS target is a separate

question that depends on your margins.

How to run an incrementality test, step by step

1. Name the budget decision the test will settle. Write it as a

rule: "If Meta prospecting iROAS is below 2.5, move 30% of it to

search."

1. Pick one outcome measured outside the ad platform: backend orders,

CRM pipeline, closed-won revenue or GA4, ideally captured with

[server-side conversion

tracking](https://synterai.com/blog/server-side-conversion-tracking).

For [long B2B sales

cycles](https://synterai.com/blog/b2b-marketing-attribution),

choose an outcome observable within the test window, such as

qualified pipeline, rather than assuming closed-won revenue will

mature in time.

1. Choose the design. Use the platform's lift study for one

eligible platform, geo for cross-platform or offline questions, and

an audience holdout for owned lists.

1. Size it before you spend. The minimum detectable effect is the

smallest lift the test can reliably detect. The 900,000/100,000

example detects a 20% lift with about 80% power but has only a 28%

chance of catching a true 10% lift. An even split would need about

685,000 people per group to detect that 10% at 80% power and 5%

significance, using the standard [two-proportion

calculation](https://www.stat.ubc.ca/~rollin/stats/ssize/b2.html).

Google's setup flow and GeoLift both estimate power for you.

1. Freeze other changes in the test and control cells. Busy

accounts break tests: smart bidding reallocates, rules fire, someone

refreshes creative, an agent shifts budget toward yesterday's

winner. Meta asks for a consistent budget and no edits that restart

the learning phase. Google treats bid and budget changes as business

as usual in its user-level studies but warns that creative and

audience changes can affect results. GeoLift asks for local and

national media to stay constant across markets. Write a

change-freeze list, switch off rules that touch the test, and keep

the bid strategy fixed. If agents run your accounts through Synter,

put the budget, bid, and other write tools that touch the test cells

on the workspace deny list for the test window. Under the default

Suggest level, those changes run immediately and are audited, while

a [deny-listed tool is blocked

outright](https://docs.synterai.com/mcp/approval-model).

1. Run it and monitor the plumbing, not the result. Confirm the

holdout got no ads and tracking stayed healthy. Google shares geo

results mid-study but recommends waiting until the end.

1. Read the result and act on the rule from step 1.

Reading the results and acting on them

Incrementality test results land in one of four places:

• Positive, and the whole interval clears break-even. Keep or scale,

within the spend range you tested.

• Positive, but the interval straddles break-even. Hold the budget

and rerun with more power rather than cutting a channel that works.

• No significant lift. Google reports this when the groups don't

differ significantly. Check whether the test could detect the lift you

care about; a flat result from an underpowered test is missing

evidence, not proof of zero.

• Negative. Check for leakage, tracking breaks, or a promotion that

hit one group. If it holds, cut.

The incrementality factor is incremental conversions ÷

platform-reported conversions: 540 ÷ 1,800 = 0.30 in the example.

Multiply that campaign's platform-reported conversions by 0.30 until

the next test, and budget on the discounted number. An incrementality

factor holds for that campaign, spend level, and period, not for double

the budget. In Synter, the agent that pulled the readout can make the

budget change in the same conversation, subject to your autonomy level

and spend limits.

Use incrementality results to calibrate MMM priors and check your

attribution model's channel ranking. Our [measurement-tools

roundup](https://synterai.com/blog/best-ai-tools-paid-media-attribution-measurement-2026)

covers how attribution, incrementality and MMM fit together. Google's

study guidance describes one or two studies a year as common and

recommends testing before major budget decisions. Retest when spend,

creative strategy or market conditions change enough to make the earlier

estimate less relevant.

Running incrementality tests across ad platforms with Synter

If your question is about one platform and you meet its minimums, run

that platform's lift study first. Synter is for what comes next: geo

tests, attribution against closed-won revenue rather than each

platform's own count, and acting on the result across platforms.

Geo lift or synthetic control from the same chat. Our [read-only

measure_incrementality

tool](https://docs.synterai.com/mcp/tools) accepts the

platform, method (geo_lift or synthetic_control), and test/control

region codes. Incrementality testing is included on Scale and Custom.

This is a measurement tool; its listed inputs do not establish that it

changes regional targeting or joins CRM revenue into the lift

calculation.

Closed-won revenue in the same workspace. Our attribution treats CRM

closed-won as the truth set, pulling deals, pipeline stages, and revenue

from HubSpot, Salesforce, and other CRMs, so [closed-won

revenue](https://synterai.com/features/attribution) sits

next to the ad data you're testing.

A record of what changed mid-test. Every write through Synter's

safety gate lands in the workspace audit log, and our [change

journal](https://synterai.com/security-governance) shows

who changed which field, with old and new values. Edits made directly in

a platform's own interface bypass Synter; for Google Ads,

get_account_changelog returns the platform's change history, capped at

30 days.

Act where you read the result. Request budget changes across the 16

connected ad platforms, subject to each API's capabilities and your

permissions. Review the measurement and proposed change before executing

it.

To try it, [connect Synter's MCP

server](https://synterai.com/mcp) to Claude or Cursor (code

MCP100 currently adds a \$100 ad credit voucher) and ask the agent to

scope a geo lift test on your largest platform: regions, duration, and

detectable lift.

Contact our team to

discuss incrementality testing across your ad platforms.

FAQ

What is the difference between an A/B test and an incrementality test?

A typical ad A/B test compares two versions of an ad or setting, and

both groups see ads. An incrementality test holds a control group out of

ads, so it measures what advertising caused.

How do you prove incrementality?

Run a controlled experiment, measure the outcome outside the ad

platform, freeze other changes, and report lift with a confidence

interval.

How do I calculate incrementality?

Lift % = (test rate - control rate) ÷ control rate. Incremental

conversions = test conversions - (control rate × test group size). iROAS

= incremental revenue ÷ incremental spend (all test spend in a holdout).

What is Meta incrementality testing and how does it work?

Meta's Conversion Lift randomly assigns your eligible audience to a

test group that can see your ads or a control group that can't, then

compares conversions. You set it up under Experiments in Ads Manager.

How long should an incrementality test run?

At least one purchase cycle. Meta suggests 28+ days, Google typically

more than 14, and GeoLift a 15-day minimum on daily data.

How much budget do you need for an incrementality test?

Meta and Google minimums start around USD 5,000; Meta recommends USD

10,000 for a well-powered test. Smaller lifts need more spend to detect.