Creative Testing Sample Size for Meta & Google Ads: India Guide

Which inputs determine creative-test sample size?

Four inputs drive the planning estimate:

| Input | Meaning | Planning mistake to avoid | | --- | --- | --- | | Baseline rate | Current conversion rate for the chosen outcome | Using a blended account rate from a different audience or funnel | | Minimum detectable effect | Smallest lift worth acting on | Asking the test to detect any difference, however commercially trivial | | Significance level | Tolerance for a false positive | Lowering the threshold after seeing weak results | | Statistical power | Probability of detecting the planned effect if it exists | Ignoring false negatives in low-volume accounts |

This guide uses a two-sided 5% significance level and 80% power for planning. That convention is not a rule imposed by Meta or Google. It is a disclosed choice that makes the worked examples comparable.

Use the Conversion Rate Calculator to establish the baseline with the same numerator and denominator you will use in the experiment. The conversion-rate glossary explains why changing either definition invalidates a comparison.

How do baseline rate and detectable lift change the answer?

For a binary outcome such as purchase or qualified lead, an approximate two-proportion calculation shows how quickly the requirement grows when the baseline is low or the desired lift is small.

| Baseline | Target | Relative lift | Approx. observations per arm | Approx. total | | ---: | ---: | ---: | ---: | ---: | | 1.0% | 1.2% | 20% | 42,693 | 85,386 | | 2.0% | 2.2% | 10% | 80,682 | 161,364 | | 2.0% | 2.4% | 20% | 21,109 | 42,218 | | 2.0% | 2.6% | 30% | 9,798 | 19,596 | | 5.0% | 6.0% | 20% | 8,158 | 16,316 |

Method: estimates use the normal approximation for two independent proportions with equal allocation, a two-sided 5% significance level and 80% power. They assume independent observations and a stable binary outcome. They are planning estimates, not replacements for native platform analysis.

The table exposes the real constraint for many Indian D2C accounts: detecting a 10% relative improvement in purchase rate may require far more traffic than the available weekly budget can buy. The answer is not to stop early. Choose a more material creative change, a longer test, a higher-frequency outcome or a directional learning objective with appropriately cautious language.

How does Google Ads decide whether an experiment is significant?

Google's statistical methodology says its experiments use bucketed data, jackknife resampling and two-tailed testing. Eligible auctions are assigned to control and treatment before targeting, and the interface can report a confidence interval around the estimated difference.

The experiment scorecard guide lists common reasons for a non-significant result:

• the experiment has not run long enough;
• the campaign lacks enough traffic;
• the split sends too little traffic to the experiment; or
• the real performance difference is too small to distinguish from noise.

The native scorecard is the operational source of truth for a Google Ads experiment. The sample-size table above is a preflight estimate that helps decide whether the test is feasible before budget is committed.

What should a clean creative experiment hold constant?

Test one material hypothesis while keeping the rest of the system as stable as practical.

For example:

• Hook: problem-first versus outcome-first opening;
• Proof: demonstration versus customer evidence;
• Format: native vertical video versus resized brand film;
• Offer framing: percentage saving versus bundle value; or
• Creator delivery: founder voice versus creator voice.

Do not change the hook, audience, landing page, offer and optimization event in the same two-cell test. Even if the treatment wins, the team will not know what caused the difference.

Google advises that an A/A test should keep the original and experiment identical, including changes and approvals. Its experiment FAQ also warns that simultaneous experiments can interfere with one another and recommends sequential testing where overlap would contaminate results.

HML's creative-fatigue guide explains when to refresh assets after launch. The short-form video strategy helps turn the winning hypothesis into platform-native variants without assuming that every resize is a new idea.

What does an India D2C worked example look like?

Consider a fictional home-care brand with a stable 2% purchase rate. The team wants to compare a demonstration-led video against its current testimonial-led creative. A 20% relative lift—to 2.4%—is the smallest improvement that would justify new production and a wider rollout.

| Planning field | Decision | | --- | --- | | Primary metric | Net purchase conversion rate | | Baseline | 2.0% | | Minimum detectable effect | +20% relative, or +0.4 percentage points | | Confidence and power | 95% two-sided; 80% power | | Planned sample | 21,109 eligible observations per arm | | Expected purchases at baseline | About 422 per arm | | Allocation | 50:50 | | Minimum calendar coverage | Two full weekly cycles plus conversion lag | | Guardrails | Refund rate, contribution per visitor and new-customer mix |

Suppose each arm reaches the planned observation count. Control converts at 2.02% and treatment at 2.37%. The treatment is commercially promising, but the team should still use the platform confidence interval and backend transaction reconciliation before applying it. If refund rate or contribution worsens, a purchase-rate win may not be a profit win.

Disclosure: all figures are synthetic. No client result or universal benchmark is implied.

What if the account cannot reach the required sample?

Use one of five honest alternatives:

1. Test a larger creative contrast. A genuinely different proposition is easier to detect than a button-colour change.
2. Use a higher-frequency primary metric. Qualified landing-page view or add-to-cart may be appropriate if purchase volume is too low, but label the conclusion as upstream.
3. Pool only comparable traffic. Combine campaigns or time periods only when audience, offer, placement and measurement definitions are sufficiently stable.
4. Extend the test. Cover full business cycles and conversion lag, but stop if seasonality, stock or promotion changes make the cells incomparable.
5. Treat the result as directional. A low-confidence signal can inform the next test without being called a proven winner.

Do not solve low power by checking more often, adding variants mid-test or switching the success metric after results appear.

How should the winner be approved for scale?

Use a three-part decision gate:

| Gate | Evidence | | --- | --- | | Statistical | Platform experiment status, interval and planned primary metric | | Commercial | Net orders, contribution, refunds and new-customer economics | | Operational | Creative can be produced, refreshed and served without policy or brand risk |

The performance marketing furniture case study shows why account structure and commercial outcomes need to be read together; its results are not evidence for the synthetic test above.

If the team is running many variants but cannot explain the hypothesis, stopping rule or minimum sample, request a creative experimentation audit from HML's performance marketing team. The deliverable should be a ranked test backlog, feasibility estimate and decision register—not a weekly spreadsheet of premature winners.

Sources

1. The statistical methodology behind experiments — Google Ads Help; accessed 24 September 2026.
2. Monitor your experiments — Google Ads Help; accessed 24 September 2026.
3. About the Experiments page — Google Ads Help; accessed 24 September 2026.
4. Experiments FAQs — Google Ads Help; accessed 24 September 2026.
5. Instagram and Facebook Reels ads — Meta for Business; accessed 24 September 2026.