Brand Lift Study: Setup, Math and How to Read It

A Brand Lift study is the one number in an awareness report that comes from an experiment instead of a delivery log. The platform holds the ad back from a random part of your audience, surveys both groups, and reports the difference. The answer is narrow: did people who could see the ad answer one survey question differently from people who could not.

Narrow is fine. The trouble starts when the result gets read. "+13% brand lift" and "+4 points" can be the same study, "no lift detected" usually means "too few answers" and not "the ad did nothing", and a lift found in one age bracket out of eighteen cuts is more likely noise than insight.

$5,000YouTube minimum for one question in Kazakhstan, spent in the first 10 days

2,452responses in total to detect a 4-point lift at a 30% baseline

9,642responses in total to detect 2 points: half the lift, four times the sample

Every platform requirement below was checked on 5 October 2026 against the vendor's own help pages. They change; re-check before you plan.

What the study actually is

Three steps, and only the first one is unusual: a holdout, one survey for both groups, a subtraction.

Because the split is random, the two groups differ only in whether the ad was allowed to reach them. That is what separates a Brand Lift study from a before/after survey, where the season, a competitor's launch and your own PR all move the number along with the ad.

Google Ads and YouTube

The control group is people who were eligible for the ad and were not shown it.²

Five numbers from one experiment

Every platform reports some subset of the same five figures. They all come from two proportions: the positive response rate in the exposed group, written p_e below, and the same rate in the control group, p_c. Google calls the second one the baseline.

L_{ab s} = p_{e} - p_{c} L_{r e l} = \frac{p _{e} - p _{c}}{p _{c}} L_{h e a d} = \frac{p _{e} - p _{c}}{1 - p _{c}}

lifted users = L_{ab s} \times reach cost per lifted user = \frac{spend}{lifted users}

Absolute lift is a difference in percentage points. Relative lift divides it by the baseline. Headroom lift divides it by the people who were still there to be won: everyone who did not already answer positively. These are Google's definitions.³ Yandex uses the same absolute and relative formulas and calls lifted users "uplift reach".¹⁰ Meta reports the point difference as "brand lift percent" and the head count as "brand lift".⁶

A worked example. The numbers are illustrative, chosen to be easy to check:

Campaign reach: 1,200,000 people. Spend: ₸6,000,000.
Question: "Which of these brands have you heard of?"
Control group: 2,000 responses, 600 positive. Baseline rate 30.0%.
Exposed group: 2,000 responses, 680 positive. Exposed rate 34.0%.

One experiment, two proportions. Every other figure in the report is derived from their difference.

Metric	Arithmetic	Result
Absolute lift	34.0% - 30.0%	4.0 points
Relative lift	4.0 / 30.0	13.3%
Headroom lift	4.0 / 70.0	5.7%
Lifted users	4.0% × 1,200,000	48,000
Cost per lifted user	₸6,000,000 / 48,000	₸125

Three "lifts" - 4.0, 13.3 and 5.7 - describe one result. Whoever writes the report picks the one that looks best, so decide which one you read before you open it.

How sure is it

The 4.0 points came from 4,000 people, not from 1,200,000. Ask another 4,000 and you get a slightly different number. The standard error of the difference says how different:

S E = \frac{p _{e} ( 1 - p _{e} )}{n _{e}} + \frac{p _{c} ( 1 - p _{c} )}{n _{c}}

With the numbers above, SE = 1.47 points. Google's reports use an 80% two-sided interval, which it explains as a 90% chance that the true lift is above the lower bound.³ That interval is the lift plus or minus 1.2816 standard errors:

Lift: 4.0 points, interval from 2.1 to 5.9.
Lifted users: 48,000, interval from 25,335 to 70,665.
Cost per lifted user: ₸125, anywhere from ₸85 to ₸237.

Google also shows a "certainty", defined as one minus the p-value.³ Here the z-score is 2.71 and the certainty, a one-sided figure that agrees with the lower bound of that interval, is 99.7%. Meta reports a confidence percentage and treats a result above 90% as reliable.⁶

Now keep the same rates and cut the sample to 400 responses per group: 136 of 400 against 120 of 400. The lift is still 4.0 points. The standard error is 3.30 points, the interval runs from -0.2 to 8.2, and the certainty is 88.8%. The interval includes zero, so the platform prints "No lift detected".

The dot is the measured lift, the band is the 80% interval. The effect is identical; only the number of responses changes. Scale in percentage points.

"No lift detected" is not "no lift"

Same creative, same media, same real effect. The only thing that changed is how many people answered. Google draws the line itself: "Not enough data" when responses are below the minimum, "No lift detected" when there were enough responses and no statistically significant difference.³ Neither status says the lift was zero.

Put your own report through the same arithmetic:

Brand lift calculator

Starts with the worked example. Replace the numbers with the ones from your report.

Exposed group: responsesExposed group: positive answersControl group: responsesControl group: positive answersCampaign reach, peopleSpend

Absolute lift4.0 pts80% interval: 2.1 … 5.9

Relative lift13.3%

Headroom lift5.7%

Certainty (1 - p)99.7%z = 2.71

Lifted users48,00025,335 … 70,665

Cost per lifted user12585 … 237

Lift detected. There is a 90% chance the true lift is above 2.1 points.

How many responses a study needs

You can work out the required sample before launch. For a one-sided 90% test with 80% power (z_α = 1.2816, z_β = 0.8416), the responses needed in each group are:

n = \frac{( z _{α} 2 p ˉ ( 1 - p ˉ ) + z _{β} p _{c} ( 1 - p _{c} ) + p _{e} ( 1 - p _{e} ) ) ^{2}}{( p _{e} - p _{c} ) ^{2}}, \overset{p}{ˉ} = \frac{p _{c} + p _{e}}{2}

At a 30% baseline:

Lift you want to detect	Responses per group	Responses in total
1 point	19,111	38,222
2 points	4,821	9,642
3 points	2,161	4,322
4 points	1,226	2,452
5 points	791	1,582

30% baseline, one-sided test at 90%. Each dot is 2,000 simulated studies. Half the lift takes four times the responses: 1,226 against 4,821.

Halving the lift you want to see quadruples the sample. That is why small flights come back empty.

The curves are the formula. The dots check it against 2,000 simulated studies at each sample size, with the lift really there: at 1,226 responses per group 81.8% of them detected the 4-point lift, and at 4,821 per group 79.6% detected the 2-point one. Both sit on the 80% the formula promises.

Google's own planning figures are in the same range. Google Ads Help says a strong campaign needs about 2,000 responses per metric, a campaign at the recommended minimum budget about 4,100, and that if nothing shows after 16,800 the lift may be undetectable.³ The Display & Video 360 table gives 1,200 to 2,800 responses for a lift above 4 points, 5,000 to 11,000 for 2 points and 20,000 to 45,000 for 1 point.⁴

Responses are bought with reach, and reach with budget. This is where the platform minimums come from, and it is why every extra question costs more: each question needs its own responses.

A study with no lift in it

A test earns trust by staying quiet when nothing happened. The way to check that is the one Google's measurement team applies to its own geo-experiment framework: run the analysis many times on data where the true effect is zero and look at the p-values. With no effect they should be spread evenly between 0 and 1, so a 10% threshold is crossed 10% of the time.¹²

I ran that check on the test from this article: 4,000 simulated studies in which both groups answer "yes" at 30.0%, with 143 responses arriving per group per day for 14 days.

Left: one of 4,000 simulated studies. There is no real lift in it, yet on day 4 the measured lift crossed the threshold, and by day 14 it was back near zero. Right: all 4,000 of them. Read once, the test is wrong as often as it should be; read daily, it nearly triples the false lifts.

How the study is read	Studies that show a lift	What it should be
Once, on day 14	10.2%	10%
Every day from day 2, stopping at the first "lift"	29.4%	10%

Read once, the test is wrong as often as a one-sided 90% test is allowed to be. Read daily, the same test on the same data nearly triples its false lifts, and nothing about the campaign changed. Thirteen looks at accumulating answers are thirteen chances for noise to cross the line, and in the first days the measured figure sits far from zero even where there is no lift at all.

A lift that shows up early is not a result

Fix the day you will read the study before it starts, and read it then. The same arithmetic applies to slices: every extra look and every extra cut is another draw from the same noise.

What each platform requires

	Google Ads / YouTube	Meta	TikTok	Yandex Direct
How to get it	Through a Google account representative¹	Brand survey test in Experiments, or a managed study through a Meta representative; certain regions only⁷	Through an account manager only⁸	In the settings of a media campaign⁹
Minimum budget	Kazakhstan is in the lowest tier: $5,000 for 1 question, $10,000 for 2, $20,000 for 3, spent within the first 10 days¹	By country. Kazakhstan is not on the list; the default for unlisted countries is $15,000⁵	Not published	₸4,000,000 for the period and ₸285,500 per day before VAT, reach of at least 3 million users⁹
Questions	Up to 3: ad recall, awareness, association, consideration, favorability, purchase intent, custom¹	Up to 3. Ad recall is mandatory; plus awareness, favorability, familiarity, recommendation⁷	Not published in the help centre	3 of 6: awareness, ad recall, purchase intent, consideration, favorability, ad message recall⁹
Duration	Standard study completes in 3 to 14 days¹	At least 2 weeks, up to 90 days⁷	Not published	At least 14 full days; 2 to 4 weeks recommended⁹
How certainty is shown	Certainty = 1 - p-value; 80% interval³	Confidence percentage; above 90% counts as reliable⁶	Not published	No confidence level stated¹⁰

Google

The budget is the sum across all campaigns in the study and is converted to the account currency. If an earlier study ended in "No lift" or "Not enough data", Google offers an Enhanced study: three times the minimum budget, 9 to 14 days, roughly three times the surveys.¹

Setting it up

The order matters, because most of these cannot be fixed after launch. Tick them off as you go; the marks stay in this browser. 0 of 7 done

Decide the question firstOne question, tied to what the campaign is for. A launch flight for an unknown product moves ad recall and awareness. Consideration and purchase intent move for brands people already know.
Put the whole flight in the studyAll campaigns for the brand that run in the same period belong in it. Meta states plainly that other active campaigns for the same brand contaminate the holdout.⁵
Create the study before the first impressionGoogle counts the budget over the first 10 days. Yandex wants the Top of Mind study attached within 3 days of launch.
Check the reach forecast against the sample tableIf the forecast buys 1,500 responses and the lift you can expect is 2 points, the study will end as "no lift detected". Know it now.
Freeze what you canChanging creatives, audiences or bids mid-study changes what the exposed group was exposed to.
Write down the decision rule"If the lower bound of awareness lift is above 2 points, we repeat the format in Q1." Written before the result exists.
Fix the reading dayOne look, at the planned end of the study. In the simulation above, daily looks turned a 10% false-lift rate into 29.4%.

Where the reading goes wrong

The same 2-point lift, three brands:

Baseline	Absolute lift	Relative lift	Headroom lift
5%	2 points	40.0%	2.1%
30%	2 points	6.7%	2.9%
60%	2 points	3.3%	5.0%

A new brand with 5% awareness reports "+40%" for the same 2 points that a market leader reports as "+3.3%". Neither number is wrong. They are not comparable, and a benchmark deck that mixes them compares nothing. For a small brand read absolute lift. For a brand near the ceiling read headroom lift.

Slicing by segment

A report cut by 6 age-and-gender groups and 3 creatives has 18 cuts. At 90% confidence, with no real effect anywhere, expect 1.8 of them to show a "significant" lift by chance. The probability that at least one does is 85%.

A lift in one segment is a hypothesis for the next flight

Ad recall as awareness

Ad recall asks whether the person remembers seeing the ad. It is the easiest question to move and the least connected to the brand.

Report it as ad recall

Estimated ad recall lift

The "estimated ad recall lift (people)" column in Meta Ads Manager is a model calibrated with polls. Meta labels it as estimated and says a brand lift test is what you need for the most accurate measurement.⁶ It has no holdout.

Not a study, do not present it as one

Cost per lifted user across platforms

Different questions, different audiences, different survey placements.

Compare two flights on one platform with one question

"Typical" lift benchmarks

I looked for a published norm by question and did not find one with a traceable sample. Kantar's figures quoted by WARC put display at +2.7 points of aided awareness and video with banners at +4.8, without the size of the database.¹¹

Treat "4 to 8 points is normal" as unsourced until someone shows the source

When a Brand Lift study is the wrong tool

The budget is under the threshold. A YouTube flight that spends less than $5,000 in its first ten days does not qualify for a Google study. The alternative is your own survey with a control cell, which is a different method with its own arithmetic: see measuring brand awareness with top-of-mind, spontaneous and aided questions.
The campaign is meant to sell. A survey answer is not a purchase. For a campaign that should produce applications or orders, hold out conversions, not opinions: see conversion lift.
You want a trend. A study covers one flight. Awareness over a year is a tracker.
The audience is tiny. A B2B campaign to a few thousand decision-makers cannot produce 2,000 survey responses per group. No platform setting changes that.

The takeaway

A Brand Lift study is a small randomised survey attached to your flight. Its output is one difference between two proportions, with an interval around it. Read the absolute lift and the interval. Check the number of responses before believing "no lift". Ask one question, and decide in advance which result changes the next plan.

To size a study before asking your account manager for one, the logic is the same as for any A/B test: the experiment calculator on this site uses the same two-proportion formula.

¹ Set up Brand Lift - Google Ads Help
² About Brand Lift - Google Ads Help
³ Understand your Brand Lift measurement data - Google Ads Help
⁴ Brand Lift in Display & Video 360 - Display & Video 360 Help
⁵ Meta Business Help Center: minimum budgets for brand survey tests, how holdouts work
⁶ Meta Business Help Center: brand survey test results, estimated ad recall lift
⁷ Meta Business Help Center: about brand survey tests, poll questions, best practices
⁸ TikTok Brand Lift Studies - TikTok Business Help Center
⁹ Yandex Direct Help: Brand Lift conditions, how it works
¹⁰ Brand Lift results - Yandex Direct Help
¹¹ The digital brand effect - WARC, 19 March 2020
¹² Ye, Sinn, Li, Jiao (2026). Meridian GeoX: An Open-Source Framework for Precision and Efficiency in Geo Experiments. Google, 1 September 2026

FAQ

Can a small campaign run a Brand Lift study?

Only if it clears the platform minimum. On YouTube a campaign targeting Kazakhstan needs $5,000 of spend in the first 10 days for one question, $10,000 for two and $20,000 for three. Yandex Direct asks for ₸4,000,000 over the period and a reach of 3 million users. Below that, run your own survey with a control region. Thresholds checked on 5 October 2026.

Why does one platform show a lift and another does not for the same campaign?

They are different experiments: different questions, different audiences, different survey placements and different numbers of responses. A platform with fewer responses will print "no lift detected" for the same real effect. Compare flights within one platform, not platforms with each other.

Is "estimated ad recall lift" in Meta Ads Manager a Brand Lift study?

No. It is a modelled estimate calibrated with polls and has no holdout group. Meta's own documentation says a brand lift test is needed for the most accurate ad recall measurement.

How long does a study take?

A standard Google study completes in 3 to 14 days. Meta recommends at least two weeks and allows up to 90 days. Yandex requires at least 14 full days and recommends two to four weeks.

Does a positive Brand Lift result mean the campaign increased sales?

No. It means more people in the exposed group gave the positive survey answer. Whether that turned into sales is a separate question, answered by a conversion lift test or a geo experiment.

Tags: brand-lift, brand, awareness, experiments, surveys, youtube

Seeing this in your own funnel?

Two weeks, fixed scope, read-only access. You get the findings and a 90-day roadmap whether or not we work together afterwards.

Book a 30-minute call