The short answer
Size a brand lift study from the smallest lift worth detecting. Pick one KPI, estimate its baseline, decide the smallest lift that would change a decision, then calculate the people needed per cell. At a 30% baseline, 95% confidence and 80% power, 300 people per cell detects a lift of about 10.5 points; detecting 5 points needs about 1,300 per cell.
Key takeaways
- Decide the smallest lift worth detecting before choosing a sample size.
- People per cell ≈ 7.84 × 2p(1 − p) ÷ lift², with p the baseline and lift as a proportion.
- 300 per cell at a 30% baseline detects about 10.5 points; 200 per cell detects about 13.
- Write the detectable lift into the plan so a flat result is read correctly.
The steps
- 1
Pick one primary KPI
Choose the metric the campaign is meant to move, such as consideration, and fix it before fieldwork. Sizing for one KPI keeps the study honest; other metrics are reported as secondary.
- 2
Estimate the baseline
Use last year’s tracker, a pre-wave or a category norm to estimate how many people already give the positive answer. Detectable lift is largest at baselines near 50% and smallest near the extremes.
- 3
Decide the smallest lift worth detecting
Ask what size of lift would change a decision, such as renewing a placement. That number, not the budget, should drive the sample.
- 4
Calculate people per cell
Use the formula below, or the table. Round up, and allow for exclusions: if you expect to remove 10% of completes for quality, recruit about 10% more.
- 5
Check it against the budget
If the sample is too expensive, either accept a larger detectable lift, share the survey wave with other brands, or narrow the question. Don’t quietly run a study too small to see the effect you expect.
- 6
Write the detectable lift into the plan
Record the detectable lift alongside the KPI before fieldwork. If the result isn’t significant, everyone can see whether the study was ever able to detect the lift in question.
The formula.
People per cell ≈ (1.96 + 0.84)² × 2p(1 − p) ÷ lift²
p is the baseline as a proportion and lift is the smallest lift worth detecting as a proportion, at 95% confidence and 80% power. For p = 0.30 and lift = 0.105: 7.84 × 0.42 ÷ 0.011 ≈ 300 people per cell.
Ready-reckoner at a 30% baseline.
| People per cell | Smallest detectable lift |
|---|---|
| 150 | about 15 points |
| 200 | about 13 points |
| 300 | about 10.5 points |
| 400 | about 9 points |
| about 515 | 8 points |
| about 1,320 | 5 points |
Detectable lift at 300 per cell, by baseline.
| Baseline | Smallest detectable lift at 300 per cell |
|---|---|
| 20% | about 9 points |
| 30% | about 10.5 points |
| 50% | about 11.5 points |
Platform tools work at a different scale because they survey inside their own inventory. Google’s guidance says detecting an absolute lift above 4% needs about 1,200 to 2,800 total responses, rising to 20,000 to 45,000 for a 1% lift.
Questions
Why not just survey as many people as possible?
Every complete costs money, and past a point extra people only narrow a range you no longer need. Sizing to the smallest lift worth detecting gives the cheapest study that can still answer the question.
Does the sample size apply to each sub-group?
No, it applies to each cell on the full sample. Sub-groups are smaller, so they can only detect larger lifts. That is why sub-groups should be fixed in advance and read with caution.