Testing an ad: how many clicks before you conclude

Two ads have been running for ten days. The second has a better click-through rate, so the first is switched off and everyone moves on. The scene is common, and the conclusion often wrong: on a few hundred clicks, a gap of that size can come from chance alone. Testing an ad is about making a decision you will not regret, and that means knowing in advance how many clicks it will take. This article gives the rule of thumb, the traps that skew a test and what a modest account can still test; it complements our online advertising guides.

Why a test concluded too early is expensive

An ad wrongly declared the winner is not harmless: you may be switching off the better one, and building the next tests on a false certainty. The cost is invisible, since nobody will ever know what the other version would have done.

The benefit of a properly sized test is the reverse: a decision that holds, that you can apply to other campaigns and defend to anyone who questions it. It also saves weeks spent trying to separate two nearly identical texts.

How many clicks does it take?

A simple rule gives the order of magnitude. Evan Miller sets it out in his article on A/B testing mistakes: the number of observations per version is roughly 16 times the variance divided by the square of the gap to detect. For a rate (click-through or enquiry) written p, the variance is p × (1 − p).

In practice, with example values rather than market averages: a page that turns 4% of clicks into enquiries. The smaller the gap you want to see, the more clicks you need:

You need those clicks for each version. To see a move from 4% to 5%, a gain of a quarter all the same, you therefore need about 12,300 clicks in total; at €4 a click, again as an example value, the test costs nearly €49,000. To see a doubling, 768 clicks are enough, around €3,000.

Click-through rate is cheaper to test

Click-through rate is measured on impressions, which are more numerous than clicks. With a 5% click-through rate, the same rule calls for about 7,600 impressions per version to see a move to 6%. That is why ads are so often judged on click-through rate: it is the only figure that arrives quickly. But an ad that attracts more clicks may attract browsers: a click-through test tells you which ad attracts, not which one sells. Keep an eye on qualified enquiries, even if they are not enough to decide on their own.

The trap of checking the result every day

The second trap is subtler. You launch the test, look at the table every morning, and stop on the day the gap becomes “significant”. Evan Miller shows that this habit destroys what the threshold promises:

The fix fits in one sentence: set the duration or the number of clicks before you start, and decide only at that point. Looking along the way is still fine for spotting a breakdown (a disapproved ad, a broken page), not for naming a winner.

What Google Ads does for you, and what it decides alone

Custom experiments

Google Ads lets you set up a custom experiment: a copy of the campaign that shares its traffic and budget with the original. Google recommends a 50% split, offers a cookie-based split (each person sees only one version) or a search-based split (faster, but one person may see both), and warns that changing the original campaign during the test makes the result harder to read.

Its experiment monitoring page shows the gap, a confidence interval and an asterisk when the result is statistically significant. Two defaults are worth knowing: the confidence interval shown is 80%, which you can raise to 95%; and for some experiment types, Google automatically applies the trial version if its results are favourable, unless you turn that option off. A lower threshold concludes faster, with more false winners: choose it knowingly.

Responsive search ads already test combinations

A responsive search ad accepts 3 to 15 headlines and 2 to 4 descriptions, which Google assembles and tests itself over time. Nobody can separate the possible combinations by hand: what you can usefully test is one idea against another (two promises, two proofs), in two ads or two campaigns. Google advises most advertisers against pinning headlines, because pinning reduces the combinations.

What to test on a modest budget?

The tests with a chance of concluding on a modest B2B account are those that can produce a large gap:

• the promise: “quote within 24 hours” against “guaranteed call-out within a week”;
• the landing page: a page about the specific need against the home page, often the biggest gap, as our landing page article shows;
• the proof: a real, quantified customer case against a list of benefits.

Variants of the same sentence are best left to Google, which tests them inside the responsive ad. Generating them with AI is quick; sorting and reviewing them still has to be done, as our article on generated copy and visuals explains.

Judging on enquiries, not only on clicks

A test judged on clicks rewards the ad that attracts; a test judged on qualified enquiries rewards the one that sells. For the second, you need to know which ad brought which enquiry. Since 18 October 2026, every enquiry sent through this site’s form carries the campaign parameters of the visit, kept for the browser tab without a cookie: a variant tagged in its parameters shows up in every enquiry it produced. The rest, through to the sale, goes through our guide to sending CRM sales back to Google Ads.

Testing across the rest of your campaigns

• The budget sets what a test can detect: the minimum budget that gives a signal.
• Useless searches blur a test: exclude them first with negative keywords.
• Performance Max tests its own assets, and Microsoft offers its own experiments: what the automated campaign hides, Microsoft Ads for B2B.
• The result is read as the cost of a customer, not of a click: customer acquisition cost.
• Consent reduces measured conversions in the same way for both versions: Consent Mode explained.
• Small areas produce few clicks, so tests are slow: local advertising.
• Other channels are tested with the same rule, from Meta to LinkedIn, from social media to YouTube, and retargeting is read separately.
• The choice between ads and SEO can be tested too: SEO or paid search, where the first euro goes.
• An agency should say how it concludes its tests: what the contract should contain.
• Ads in AI answers are measured without detailed reports: what we know, and a generated variant that over-promises raises a question of trust (AI and disinformation).

To size your tests before launching them, and judge them on the enquiries they produce, our digital marketing offer starts with measurement.

Frequently asked questions

Sources

1. Evan Miller, How Not To Run an A/B Test (new tab)evanmiller.org
2. Google Ads Help, set up a custom experiment (new tab)support.google.com
3. Google Ads Help, monitor your experiments (new tab)support.google.com
4. Google Ads Help, about responsive search ads (new tab)support.google.com