How to use Incrementality Testing
A practical guide to incrementality testing for paid channels: geo holdouts, ghost bids, and conversion lift studies — plus how to read incremental ROAS without fooling yourself.
Incrementality Testing
Experimental methods — geo holdouts, ghost bids, conversion lift — that measure the revenue a channel actually caused, not just claimed.
Incrementality testing is the experimental practice of measuring the causal contribution of a marketing channel by comparing exposed and unexposed groups. Instead of trusting platform-reported conversions or last-click attribution, you deliberately suppress ads for a randomised slice of your audience or geography and observe the gap in revenue.
The three workhorse methods are geo holdout tests (turn a channel off in matched regions), ghost bids (record auctions you would have won but don't serve), and platform-run conversion lift studies (Meta, Google, TikTok randomise users into a no-ads control). Together they answer one question: if you cut this spend tomorrow, how much revenue would actually disappear?
Last-click attribution gives every channel credit for conversions it touched last. The problem is that many of those buyers would have converted anyway — they searched your brand name because they saw a Meta ad three days ago, or because a friend recommended you. The platform booked the sale; the channel didn't cause it.
Incrementality testing fixes this by introducing a control group that doesn't see the ads. Whatever revenue the exposed group generates above the control is the true lift. Everything else is correlation dressed up as performance.
The three methods that actually work
Geo holdout tests are the most common starting point. You split your markets into matched pairs — comparable cities or regions with similar baseline revenue and seasonality — then pause the channel in one half for 2-4 weeks. The revenue gap between treatment and control geos, normalised against the pre-test baseline, is your incremental lift.
Ghost bids are the cleanest method when the platform supports them. The auction engine records every impression you would have won but withholds it from a random share of users. Because randomisation happens at the user level inside the same auction, confounds from seasonality, supply, and competitor activity disappear. Meta's Conversion Lift and Google's Ghost Ads work this way.
Conversion lift studies are the platform-managed version: you brief the platform on the conversion event, it randomises users into a control cell, and reports lift at the end. Lower operational lift than geo tests, but you're trusting the platform's measurement of its own performance — keep that bias in mind.
Synthetic controls beat naive matching
Don't pair geos by gut feel. Use a synthetic control method (or at minimum a 4-week pre-test baseline regression) to weight donor markets so the treatment and control groups have indistinguishable revenue trajectories before the test starts. A bad match is the single most common reason incrementality tests produce noise instead of insight.
Designing a geo holdout that produces a real answer
Pick at least 6-8 matched market pairs. Fewer than that and your confidence intervals will swallow any realistic lift — you'll end the test unable to distinguish a 15% incremental channel from a 40% incremental one. For smaller stores, run the test longer (4-6 weeks) rather than with fewer geos.
Run a pre-test period of equal length to validate the matching. If treatment and control diverge by more than 3-5% during the pre-period, your design is broken — re-match before you cut spend. Spending €40k on a holdout that turns out to be confounded is a worse outcome than not testing at all.
Reported vs incremental ROAS by channel (typical DTC apparel)
Platform-reported ROAS
Incremental ROAS (measured)
The gap between bars is where budget gets quietly wasted. Branded search and retargeting almost always show the largest delta — they intercept demand that already exists. Prospecting channels (cold Meta, TikTok, upper-funnel YouTube) typically have a smaller gap because they're genuinely creating new demand rather than harvesting it.
Reading incremental ROAS without fooling yourself
Incremental ROAS is the revenue lift divided by the spend that produced it. If pausing €50k of Meta retargeting in a holdout cost you €90k of measured revenue, your incremental ROAS is 1.8x — not the 7.6x the platform claims. The right comparison for budget decisions is incremental ROAS against your contribution-margin breakeven, not against the reported number.
Expect uncomfortable findings. Most stores discover branded search has an incremental ROAS between 1.5x and 3x, despite reporting 10x+. Retargeting often lands between 1.5x and 2.5x. The fix isn't always to cut — sometimes the channel is still profitable at its true rate — but you need the honest number to make the call. This is where incrementality directly improves your overall ROAS measurement: it tells you which reported numbers to trust.
Typical incrementality factor by channel (incremental ÷ reported revenue)
| Channel | Incrementality factor | Confidence | Common test method |
|---|---|---|---|
| Branded search | 0.10 – 0.25 | High | Geo holdout |
| Non-brand Google search | 0.55 – 0.75 | Medium | Geo holdout / ghost bids |
| Google Shopping | 0.60 – 0.80 | Medium | Geo holdout |
| Meta retargeting | 0.20 – 0.35 | High | Conversion lift |
| Meta prospecting | 0.70 – 0.90 | Medium | Conversion lift / geo |
| TikTok | 0.65 – 0.85 | Lower | Conversion lift |
| Affiliate / coupon | 0.15 – 0.40 | High | Code-suppression test |
These ranges are starting hypotheses, not destinations. Your incrementality factors depend on brand awareness, repeat-purchase rate, and audience overlap across channels. A new beauty SKU with low brand recognition will see higher Meta prospecting incrementality than an established apparel store whose customers already know the brand.
Operationalising incrementality without breaking the team
Run one geo holdout per quarter on your two largest channels. Rotate which channels you test — annual coverage of every major channel is more valuable than re-testing branded search three times. Build the schedule into your media plan so finance and the agency know when revenue dips are by design.
Feed the results into your bidding and budget rules. If incremental ROAS on Meta retargeting comes back at 1.8x and your contribution-margin breakeven is 2.2x, you have a clear action: cap retargeting frequency, narrow the audience, or cut the budget by 30% and rerun. The test only pays back if you act on it.
What incrementality testing can't tell you
A single holdout measures average lift across an audience, not lift per impression or per creative. It won't tell you which ad worked, only that the channel as a whole did or didn't drive revenue. Pair quarterly incrementality with ongoing creative A/B tests and MMM for the full picture.
Incrementality testing FAQ
Attribution divides credit for conversions that already happened across the touchpoints that touched them. Incrementality testing measures whether those conversions would have happened at all without the channel. Attribution is bookkeeping; incrementality is causation.
Two to four weeks for high-volume channels with at least 8 matched market pairs. For smaller stores or lower-volume channels, extend to 4-6 weeks rather than cutting the number of geos. Run an equal-length pre-test period to validate that treatment and control geos track within 3-5% of each other before you start.
Geo holdouts get statistically shaky at that volume — you won't have enough geos with meaningful revenue. Stick to platform-run conversion lift studies (Meta, Google) which randomise at the user level and need less absolute volume, and run code-suppression tests on affiliate or coupon channels.
Compare against your contribution-margin breakeven, not your reported ROAS. If your contribution margin is 45% and you want a payback in the first order, your incremental ROAS breakeven is around 2.2x. Above that is profitable on the first purchase; below requires LTV to make the math work.
Most people searching your brand name already intend to buy from you — they saw an ad, got a recommendation, or are returning customers. The branded search ad intercepts a conversion that would have happened on the organic result one line below. Typical incrementality is 10-25% of reported revenue.
They're directionally useful but biased toward showing the platform's value. Meta's Conversion Lift will tend to credit Meta; Google's lift studies will tend to credit Google. Use them for relative comparisons within a platform (audience A vs B) and cross-check absolute numbers with an independent geo holdout.
Marketing mix models estimate channel contribution from historical data using regression. Incrementality tests produce ground-truth lift numbers that calibrate the MMM — without experimental priors, MMMs can drift badly. Best practice is to use incrementality results as Bayesian priors in the MMM.
A ghost bid is an auction your platform wins but deliberately doesn't serve to a random share of users, recording the impression that would have happened. It produces a clean user-level control group inside the same auction, making it the gold-standard method when the platform supports it (Google Ghost Ads, Meta Conversion Lift).
Yes, by design — that's how you measure lift. Budget the expected revenue dip into your quarterly forecast (typically 1-3% of total revenue for a 3-week single-channel test) and pick a period without major launches or peak season. If you can't afford the dip, run a smaller-geo test or a platform conversion-lift study instead.
Once a year for established channels, or whenever you make a structural change — new creative strategy, audience overhaul, major bidding-model switch, or a 30%+ budget change. Incrementality factors drift as audience saturation, brand awareness, and competitor pressure shift.
Track CAC, channels, and funnel conversion in one place
Metricuno connects ad spend, funnel events, and revenue so you can see CAC by channel, cohort, and campaign — without stitching together five tools.