
Testing ads means changing one element between two versions of an ad, showing each version to a separate part of the audience on the same budget, and judging only once enough results have come in. A trial where you change the image, the copy and the offer at once tells you nothing useful, even if one version wins by a mile, because you cannot tell which change made the difference.
Below is the order in which to test elements, what Meta and Google say in their official help pages about test length and picking a winner (as checked in November 2026), a weekly plan you can start next week, a results log to copy as it is, and a simple way to tell when your numbers are too small to judge.
What to test first in an ad creative test
Start with the angle, the idea the ad uses to persuade, and move to smaller elements in order. This table ranks the variables from the one that usually moves results most to the one that moves them least:
| Variable | What changes | Two versions, for example |
|---|---|---|
| Angle | The reason the ad gives to buy | Saving time versus making fewer mistakes |
| Hook | The first two seconds of a video, or the first line of copy | A question about the customer's problem versus a figure about your product |
| Format | The shape of the ad | A short video versus a carousel |
| Visual | The image or scene within the same format | The product alone versus the product in use |
| Copy | The wording of the message | Short copy versus copy that explains the steps |
| Offer | What the customer gets for responding | A free consultation versus a sample or trial |
In our view most companies start at the bottom of the table, with a button colour or one word in a headline. On a small budget the gap between two colours is usually too small to show up in your numbers, while the gap between two angles can be plain to see. Test an offer only if you can actually deliver it to everyone who takes it up.
The rule that matters most is one variable per test. Meta recommends keeping your ad sets identical except for the variable you are testing, and its page on selecting a variable says the tool accepts creative, audience, placements or several variables together, while advising newcomers to start with one for the clearest read.
Platform tools that split the audience for you
Use the testing tool inside the platform rather than comparing two ads by switching one off and the other on. Meta says it does not recommend informal testing, such as turning ad sets on and off by hand, because it leads to overlapping audiences and unreliable results. Its A/B test shows each version to a segment of the audience so that nobody sees both.
| Platform | Tool | Length the platform suggests | How the winner is picked |
|---|---|---|---|
| Meta | A/B test in Ads Manager or the Experiments tool | At least 7 days, up to 30 | Lowest cost per result, with a confidence percentage |
| Google Ads | Custom experiments | 2 to 3 weeks | Performance difference with a confidence interval, and a blue asterisk when significant |
| Google Ads | Asset report for responsive search ads | Several weeks | Picks no winner, gives statistics per asset |
On Meta you build a test by duplicating a campaign, ad set or ad and changing one variable, or by comparing two existing campaigns. Note that duplicating a campaign duplicates its budget too, so spend doubles unless you adjust it, and Meta advises the same budget for both versions. The winner is found by comparing each version's cost per result: Meta simulates possible outcomes tens of thousands of times to produce a confidence percentage, the chance of similar results if you ran the test again. Results appear in Experiments or under View charts in Ads Manager, and a test without a clear winner may come back with a suggestion to extend it.
Then there is the learning phase. An ad set usually leaves it after about 50 results in the week after its last significant edit, and performance during it is less stable and often costlier. Meta's same page says editing during learning resets it, and that running many ads means the system learns less about each one. So test two versions, and leave them alone until the test ends.
On Google Ads, custom experiments are available for Search, Display, Video and Hotel campaigns and split the budget between the original campaign and the experiment. Google suggests running an experiment for 2 to 3 weeks, shows the difference with an 80% confidence interval by default, and marks a statistically significant result with a blue asterisk.
Responsive search ads take up to 15 headlines and 4 descriptions, and Google picks the combinations it shows. That makes them a poor A/B test, since the system serves the assets it rates highest more often. The asset report gives full statistics per asset for dates from 5 June 2025, but Google advises judging performance at campaign level and treats ratios such as cost per conversion as directional only. An asset with no impressions after several weeks is a candidate to replace.
How long to wait before you judge
Give Meta tests at least 7 days, Google experiments 2 to 3 weeks, and longer if your customers take more than a week to decide. Meta's own example is that if your customer usually takes more than seven days to convert, you would run the test longer, such as ten days, and it warns that tests under seven days may be inconclusive.
Time alone is not enough. What counts is the number of results. Neither Meta nor Google gives a fixed number of results for a creative test, and Meta only asks for a budget that will produce enough results to determine a winner confidently. Our rule of thumb, which is no official figure: do not call a winner before each version has collected a few dozen results on the metric you are judging by. How to work out the budget that gets you there is in how to split your first ad budget across platforms.
If your budget cannot reach that many leads in two weeks, judge on an earlier step, such as landing page views or click-through rate. That signal is weaker because a click is no lead, so treat the winner as a candidate to confirm with leads later.
In the end, judge by what matters to the business: leads your sales team confirmed and the cost per sale, rather than the platform's figure alone. An ad that brings cheap leads from people who never buy is worse than one with dearer leads that close. How platform metrics tie back to profit is covered in why ROAS alone is not enough.
What small numbers cannot tell you
A gap of a few results does not make a winner. Take a hypothetical example: you spend the same on two versions for a week, version A brings four leads and version B brings seven. B looks far better. Move just two leads from B to A and it reads 6 to 5, and the winner flips. Two leads can turn up or go missing by chance in any week.
This flip test is a quick rule of thumb, not a statistical calculation: if moving one or two results from the winner to the loser reverses the verdict, you do not have a result yet. Extend the test, or log it as "no clear difference" and move on to a variable with more weight.
Other conclusions that come up often and do not hold:
- "It won in the first two days." Early days usually fall inside the learning phase, which Meta describes as unstable, and the order on day two often changes by day seven.
- "Its click-through rate is higher, so it is the better ad." A provocative headline can bring more clicks and fewer leads. Judge by the metric you set before launch.
- "The platform named a winner, so we are done." Check the confidence percentage. Meta notes that a winning version can have higher costs when the test was too short or had too few results.
- "It lost during Ramadan, so it is weak." Buying behaviour shifts in seasons and around paydays in Egypt and Saudi Arabia. Compare two versions over the same period, never a season week against an ordinary one.
- "It worked in Saudi Arabia, so it will work in Egypt." Dialect, prices and delivery differ between the two. Test each country in its own campaign, or hold the country fixed within the test.
A weekly test plan
The plan is a seven-day cycle for one Meta test and takes about two and a half hours a week. On Google, make the cycle 2 to 3 weeks as Google suggests, and run one experiment at a time.
| Day | What you do | Rough time |
|---|---|---|
| Day 1 | Write the hypothesis as: if we change X, cost per lead will fall because Y. Pick the variable, the metric you will judge by, and when you would extend | 30 minutes |
| Day 1 | Prepare both versions and create the test in the platform's tool with one budget per version | 45 minutes |
| Day 2 | Check the ads were approved and are spending, and send a test lead to confirm tracking works | 15 minutes |
| Days 3 to 6 | Five minutes a day for faults only: spend, link, comments. No edits to the ads | 5 minutes a day |
| Day 7 | Read the results, run the flip test, and decide: winner, extend to day 14, or no difference | 20 minutes |
| Day 7 | Log the test and pick the next variable from the first table | 15 minutes |
Keep testing to a small share of the budget and the rest on ads that have proven themselves. We explain that split and the loss limit in what media buying is and how it works for companies. If an ad wins and leads still do not rise, the problem may sit on the page visitors land on, which needs a test of its own.
A results log for your ad tests
Log every test on one line, including tests that ended with no difference. Copy this template into a sheet your team shares. The last column is a hypothetical example of how to fill it in:
| Field | What you write | Hypothetical example |
|---|---|---|
| Test number and dates | A running number, start and end day | Test 7, the 2nd to the 9th |
| Platform and campaign | The campaign name as it appears in the account | Meta, service leads campaign |
| Variable | One element only | Angle |
| Hypothesis | If we change X, Y will happen, because Z | A time-saving angle will lower cost per lead because customers keep asking how long it takes |
| Version A and version B | A short description and a link to the design | A: fewer mistakes. B: saves time |
| Spend per version | From the platform report | Roughly equal |
| Results and cost per result | For each version separately | A: 31 leads. B: 44 leads |
| Leads confirmed by sales | From the sales record, not the platform | A: 12. B: 19 |
| Confidence | As the test tool shows it | As shown in Experiments |
| Decision | Winner, extend, or no difference | B wins, pause A |
| Next test | The next variable and hypothesis | B's hook in the first two seconds |
After three months this log becomes the most useful file in your ad account. A new designer sees what has been tried, and whoever runs the account can see why certain ads were stopped. When you audit your ad account yourself, ask for this log first.
Your next step
This week, pick one campaign that is running now, write a hypothesis about its angle, and prepare a second version that differs in the angle alone. Create an A/B test in Ads Manager with one budget for both versions, put day seven in your calendar, and open the log before you launch. If you would like someone to run this cycle and read the results with you, see PlanKit's ad campaign management service.