Brand Lift Study in India: How to Design, Size and Read One

Pick one to three. Every extra metric adds questions, and Google's own set-up lets you "choose up to 3" lift metrics from association, awareness, consideration, favorability, purchase intent and video ad recall.

Ask the same questions in the same order to both groups. Put awareness before consideration and consideration before intent, because each question primes the next.

The whole study depends on the control group looking like the exposed group. There are three ways to get one.

Whichever you use, match the groups on what you can: city, age band, gender and NCCS. Contamination breaks a control. Google notes that if a study targets people who already viewed the video, it "currently does not build a control group" (Google Ads Help).

In a survey with a panel, the steps are: screen for category buyers, randomise into exposed and control, show the ad to the exposed group only, then ask the same questions of both. Put a filler item or neutral ad in the control, so both groups spend the same time on the screen.

Size the study before you field it. For two groups and a yes-or-no metric, the number per group is:

n = (z(α/2) + z(β))² × [p₁(1−p₁) + p₂(1−p₂)] ÷ (p₂ − p₁)²

Here p₁ is the control rate you expect, p₂ is the exposed rate you hope to detect, z(α/2) is 1.96 for 95% confidence and z(β) is 0.84 for 80% power. Power is the chance of detecting a real lift of that size.

Worked case. Control awareness 20%, target lift 5 points (p₂ = 25%). (1.96 + 0.84)² = 7.85. Then p₁(1−p₁) = 0.16 and p₂(1−p₂) = 0.1875, which sum to 0.3475. Divide by 0.05² = 0.0025 to get 139. Multiply: 7.85 × 139 ≈ 1,091 per group, so about 2,200 respondents in all.

Read down the lift columns: halving the detectable lift roughly quadruples the sample. Quota rules and weighting add variance, so allow a margin on top. As a cross-check, Google's table lists a detectable absolute lift of more than 4% at 1,200 to 2,800 responses (Google Display & Video 360 Help). For the test itself, see statistical significance testing.

Google and Meta run lift studies on their own ads. An independent survey measures the campaign from outside. They answer different questions.

Know what the platform reports. Google's absolute lift is "calculated by subtracting the positive response rate of the baseline group from the exposed group", and "no lift detected" means "there was no statistically significant difference" (Google Ads Help). Google also says the confidence interval is usually 90%, which is looser than the 95% used above. Meta reports brand lift as accounts reached multiplied by the lift in per cent. Its own example is 1,830,000 accounts reached and an 11.6-point increase, giving about 212,000 additional accounts (Meta).

Platform studies also need budget. Google says the minimum depends on the channel, country and number of questions (Google Display & Video 360 Help). A small campaign may not qualify, which is one reason to run an independent survey.

The figures are illustrative, not from a real study. A brand ran a six-week campaign. We survey 1,000 category buyers who saw the ad and 1,000 matched buyers who did not, sized for a 5-point lift from a 20% base.

Awareness. Pooled share = 520 ÷ 2,000 = 0.26. Standard error = √(0.26 × 0.74 × (1/1,000 + 1/1,000)) = 0.0196. z = 0.08 ÷ 0.0196 = 4.08. The 95% interval for the gap is 4.2 to 11.8 points. Relative lift is 8 ÷ 22 = 36%.

Reading it. Awareness moved, and it would pass even if you divided the 5% level by three for three tests (0.0167). Consideration and intent moved in the right direction, but at 1,000 per group the study could not separate 3 points from noise. That is not proof of no effect. It means this sample was too small for that metric. To confirm a 3-point consideration lift from a 35% base, you would need about 4,000 per group by the formula above.

Lift is not sales. Awareness and intent are steps toward buying. A lift study does not show revenue.

Forced exposure flatters the ad. People who are told to watch pay more attention than people scrolling past. Use the result to compare ads and groups, not to forecast market awareness.

Survey answers are stated, not observed. Intent overstates behaviour, and respondents may guess what you want to hear.

Short window. A study taken during the flight misses decay and long-term effects.

Small lifts need big samples. Report the interval, not only the point estimate, and say which metrics were under-powered.

Match on NCCS, city and age. A campaign shown more in Tier 1 cities will reach a different audience from the control if you do not. Set the same quotas for both groups. Our Tier 2 and Tier 3 guide covers sampling beyond the metros.

Run the survey in the campaign's language. If the ad ran in Tamil, show it and ask about it in Tamil. A lift read in English underweights the people the ad was made for.

Geo holdouts are possible but crude. If a campaign ran in some cities only, unexposed cities can act as a control. Cities differ in awareness before the campaign, so measure a baseline in both.

Track the screen. If most of the campaign was mobile video, show the test ad on a phone.

If your job is an independent lift read for an Indian campaign, our platform supports the forced-exposure version, run as a survey with matched cells: the ad shown in one cell and a neutral filler, or no ad, in the control. You can target by city, age band, gender and NCCS, set interlocking quotas across gender, age and city, and use video or image stimuli in a question. We do not tag who was served your live ads, so for a true platform holdout use Google or Meta. The two-group example above needs 2,000 responses, bought as two Pro bundles of 1,000.

Pro is sold in bundles of 400 to 1,000 responses and costs 20% less per response than Plus (pricing). Enterprise plans come with a dedicated research team and are priced on request.

Our panel has a 25M+ consumer footprint, and the survey-eligible panel is a subset that depends on your targeting. Members complete Aadhaar-based KYC once, and each study receives a zero-knowledge proof that the respondent is verified. Surveys run in 8+ Indian languages, and speeders, straight-lining and inconsistent answers are removed before analysis. See panel methodology.

The limit: the panel is app-based, so respondents are smartphone users only, and offline households need face-to-face fieldwork.