← All playbooks·Digital attribution

What you need first

• Campaign flight dates and the platform's delivery/exposure log, from the DSP or ad server
• Minimum detectable effect and required sample size per cell, from the platform's or vendor's brand lift power calculator (e.g. Kantar, Dynata, or the walled garden's own tool)
• Control group method, ghost ads/PSA or geo holdout, from the brand lift product spec of the platform running the study
• Baseline awareness, recall, or intent benchmark for the category, from the prior tracking wave or a syndicated tracker
• Media budget and reach forecast, from the media plan
• Significance threshold and confidence level agreed with the client, from the measurement brief

The procedure

1. Define the exposed and control cells using ghost ads/PSA or a geo holdout, sized against the platform's minimum detectable effect calculator, producing two audience pools comparable on reach and demo mix
2. Set survey questions for aided awareness, ad recall, and purchase intent that match the client's existing brand tracker wording, producing a comparable question set
3. Launch the flight with the brand lift tag firing across both cells, producing raw exposure counts and survey response counts per cell
4. Hold reporting until the platform's minimum sample threshold is hit per cell, producing a topline result stable enough to read
5. Calculate absolute lift (exposed% minus control%) and relative lift (absolute lift divided by control%), producing the headline lift figure per metric
6. Run a two-proportion significance test on exposed versus control, producing a p-value and confidence interval per metric
7. Segment lift by frequency band, creative, or placement where cell sizes allow, producing a diagnostic breakdown for mid-flight optimization
8. Write the result up against the pre-agreed success threshold, producing the go/no-go recommendation for the next flight

Worked through with numbers

Netherlands campaign, €150k budget, 2.4M unique reach forecast. Exposed cell returns 1,850 survey responses, control cell 1,720, both above the platform's minimum sample threshold. Aided awareness: 777/1,850 = 42% exposed, 602/1,720 = 35% control. Absolute lift = 42% - 35% = 7 points. Relative lift = 7/35 = 20%. Pooled proportion p = (777+602)/(1850+1720) = 1379/3570 = 0.386. Standard error = sqrt(0.386 x 0.614 x (1/1850 + 1/1720)) = sqrt(0.237 x 0.00112) = sqrt(0.000266) = 0.0163. Z = 0.07/0.0163 = 4.29, p < 0.0001. Read it as a real, statistically solid 20% relative gain in aided awareness, strong enough to justify scaling the flight, but check the ad recall and intent metrics move in the same direction before calling it a clean win.

Where it goes wrong

• Compare exposed against a true control that could have seen the ad but didn't, using ghost ads or a geo holdout, rather than against the general population, since a population comparison inflates lift with selection bias
• Wait for the platform's minimum sample threshold before reading toplines, since early reads on partial samples can swing the lift number by double digits in either direction
• Rely on the platform's own deduplication for cell assignment, since cross-device and cookie churn will leak exposed users into the control cell if you try to match IDs yourself
• Report the confidence interval alongside relative lift, since a 20% relative lift on a 7-point absolute gap reads very differently from a 20% lift on a 2-point gap and the client needs both numbers to judge it

How to know it is right

Terms used

• Brand Lift
• Incrementality

← All playbooks Calculated in ValueMetrics →