Key takeaways
- Attribution assigns credit under a model; incrementality estimates what happened because exposure or spend changed.
- Choose the experimental unit, outcome, population, power, run time, and contamination controls before launch.
- Preserve assignment and analyze the intended groups even when delivery or compliance is imperfect.
- Use lift to calibrate decisions with uncertainty—not to manufacture a universal multiplier from one campaign.
01
Ask a causal question that matches a budget decision
A useful question is not “How many conversions did Meta report?” It is “How many qualified purchases occurred because this eligible audience could receive the campaign during this period?” Name the intervention, counterfactual, outcome, population, and decision the estimate will inform.
Platform conversion-lift studies use test and control assignment to estimate incremental conversions. Google also provides experiments intended to measure uplift in eligible campaign configurations. Availability, thresholds, and methods vary, so design around the actual product and account—not a slide from another advertiser.
02
Protect assignment and measurement
Define eligibility and exclusions before randomization. Keep conversion definitions, windows, consent, and data delivery stable. Track other campaigns, promotions, stock, pricing, outages, and sales changes that could affect the groups differently.
For geo experiments, assess market similarity, spillover, travel, national media, seasonality, and sample size. A neat map does not make cities independent.
Swipe to compare every column
| Design | Strength | Common threat |
|---|---|---|
| User-level randomized lift | Strong assignment within platform eligibility | Identity loss and cross-device contamination |
| Geo holdout | Can capture broader channel effects | Spillover and unmatched local shocks |
| Time-based pause | Operationally simple | Seasonality and unrelated change |
| Attribution comparison | Fast diagnostic | Not a causal experiment by itself |
03
Plan for an answer with wide uncertainty
Estimate power and minimum detectable effect before launch. If the business cannot support the required population, spend, or duration, narrow the outcome or accept that this test may not resolve the question. Do not repeatedly peek and stop when the line becomes favorable.
Report lift estimate, interval, test population, conversion definition, dates, spend, delivery differences, and important validity threats. “No conclusive lift” does not mean zero effect; it means the study did not separate the effect from noise well enough.
04
Calibrate decisions, then retest
Compare experimental lift with attributed conversions and marketing-mix or CRM views. Use the gap to improve planning and measurement, but do not treat one ratio as permanent across channels, offers, markets, or seasons.
Archive the design and result, repeat important questions, and prioritize experiments near real allocation choices. Incrementality is most valuable when the business is genuinely willing to change spend.
Primary sources and further reading
Use the source material to validate details against your own context and current platform configuration.
- Meta Blueprint: Measure Channel Impact With Conversion Lift
- Google Ads: Performance Max Experiments
- Google Ads: Experiments
This field note follows the XenGrowth editorial policy: primary sources where available, visible limitations, material review dates, and no invented first-hand experience.
Stay with the problem
