| name | incrementality-testing |
| description | Design and run incrementality tests that measure true lift from ads — geo holdouts, platform lift studies, significance math, and iCPA decisions. Use when a user wants to know whether ads actually cause conversions, design a holdout test, or decide budget based on real lift. |
Incrementality Testing
Attribution measures correlation; incrementality measures causation. The question is always the same: would this conversion have happened without the ad?
Start with measure_incrementality where the platform and data support it; design a manual test when it does not.
1. Define the question
Pick one testable claim: channel incrementality ("does Display drive anything?"), budget incrementality ("does doubling spend double conversions?"), or tactic incrementality ("is retargeting real?"). One variable per test. Pull the baseline first: pull_<platform>_ads_performance for spend and attributed conversions, ga4_run_report for a platform-neutral conversion count.
2. Choose the method
- Geo holdout: run ads in test regions, go dark in matched control regions. The gold standard, but needs regional volume. Pause delivery in control geos via
pause_campaign or geo exclusions.
- Platform lift study: Meta and Google both offer native conversion-lift experiments with randomized holdout groups. Least effort, but you trust the platform to grade its own homework.
- Ghost ads / PSA: serve an unrelated ad to the control group so both groups pass through the same auction. Works on any platform, costs the PSA impressions.
3. Design it so the answer is trustworthy
- Match regions on population, historical CVR, revenue per conversion, and seasonality. Validate with 8-12 weeks of history: if historical CVRs differ by more than ~5%, pick different regions.
- Size the control at 20-30% of the addressable market.
- Duration: minimum 4 weeks for most channels; 2 weeks only for high-volume search or Meta; add 50-100% for long sales cycles. Avoid holidays or extend through them.
- Sample size: small lifts need big audiences. Detecting a 20% lift on a 1% baseline CVR takes roughly 400K users; a 10% lift takes 1.5M. If you cannot reach the numbers, test a bigger intervention instead of a subtle one.
4. Run it clean
Freeze everything else: no creative swaps, no budget moves, no new campaigns in test or control markets mid-flight. Use list_campaigns to confirm nothing else is live in the control geos, and set_spend_alert to catch drift. Measure conversions from a neutral source (GA4), not the tested platform.
5. Analyze
Lift = (CVR_test - CVR_control) / CVR_control
Incremental conversions = Conversions_test - (Exposed_test x CVR_control)
iCPA = Test spend / Incremental conversions
iROAS = Incremental revenue / Test spend
Run a two-proportion z-test on test vs control CVR and report the lift with its 95% confidence interval, not just the point estimate. A lift whose interval crosses zero is not a result.
6. Decide
Compare iCPA to attributed CPA:
- iCPA within ~1.5x of CPA: highly incremental. Scale it (see kill-scale-rules).
- 1.5-3x: moderately incremental. Hold and optimize.
- 3-5x: weak. Shift budget with
optimize_budget and re-test the reallocation.
- 5x+: the channel is mostly claiming credit for organic conversions. Cut it or keep it only for awareness.
Brand search typically tests at 15-35% incrementality, non-brand search 40-65%, display prospecting 5-20%, retargeting 20-45%. If your result lands far outside these ranges, re-check the test before celebrating or panicking.
Rules
- One variable per test. A test that changes budget, creative, and audience proves nothing.
- Neutral measurement only. The platform being tested does not get to score the test.
- Pre-register the success threshold before launch; do not move the goalposts after seeing data.
- Feed results back into planning: incrementality-adjusted returns belong in mmm-budget-planner and attribution conversations, not in a slide graveyard.
