How to set up Google Ads experiments

What a custom experiment is, what you can test, how traffic is split, how long to run it, how to read the result and its confidence interval, and how to apply or end it.

· 12 min readAdvertising

Short answer

A Google Ads custom experiment tests one change on a copy of the original campaign, over the same period and with part of the traffic. Pick the variable and success metric, split the traffic, wait, read the confidence interval, then apply the experiment, convert it to a new campaign or end it.

In one week a campaign changes three things: the bidding strategy, two keywords and the headline of the landing page. The week after, conversions are up and everyone is pleased. But nobody can say which change worked, or even whether the rise came from the season. Knowing how to set up Google Ads experiments exists to remove that doubt: you test the change in a separate arm, over the same period, without touching the live campaign.

Google’s custom experiments page describes it like this: an experiment campaign is created from the original, you choose the length and how much of the original traffic (and budget) it uses, and then you compare the two. Here we walk through one experiment from start to finish. We give no amounts; the numbers in the examples are made-up values. We used the interface names as they appear on Google’s English Help pages, and they may look slightly different in your account.

An experiment moves through four stops and one decision:

The path of a Google Ads experiment: set up, split, wait, read; then apply it, convert it to a new campaign or end it 1 Set up One variable Success metric Start, length 2 Split Traffic and budget Half and half Cookie or search 3 Wait To settle: 7 to 14 days Hands off 4 Read Confidence interval Asterisk Log flat results Decide: apply to the original, convert to a new campaign or end the experiment The path of a Google Ads experiment: set up, split, wait, read; then apply it, convert it to a new campaign or end it 1 Set up One variable Success metric Start, length 2 Split Traffic and budget Half and half Cookie or search 3 Wait To settle: 7 to 14 days Hands off 4 Read Confidence interval Asterisk Log flat results Decide: apply to the original, convert to a new campaign or end it
Figure: An experiment moves through four stops; the variable and metric are chosen at set-up, and at the end the result is read and applied, converted or ended.

1. Choose one question

An experiment answers one question: “If we change the bidding strategy, do conversions go up?” or “Does this landing page bring more form submissions than the old one?” Put two changes in one experiment and you cannot tell which one mattered. Write the question together with its success metric: conversions, cost per conversion or clicks. Conversion metrics need conversion tracking that is set up correctly; if the measurement is wrong, the experiment compares the wrong number.

2. Know what you can test

According to Help’s Experiments page, custom experiments are typically used to test Smart Bidding, keyword match types, landing pages, audiences and ad groups. The custom experiments page recommends ad variations if you want to test your Search ad creative, and custom experiments if you want to test specific ad groups rather than the whole campaign. The table below pairs that list with the related guides on this site.

What to test What you change in the experiment More detail
Bidding A different Smart Bidding strategy in the experiment arm Bidding strategies guide
Keywords The match type Match types guide
Ads Text and assets; ad variations are often the better fit Assets guide
Landing page A different page in the experiment arm (this is our suggestion; Help only says landing pages are typically tested) Note below

One distinction on landing pages. Splitting the traffic that comes from your ads between two pages is a Google Ads experiment. Showing two versions to every visitor on your own site is the subject of how to run an A/B test. The two are not interchangeable.

3. Set up the experiment

Help’s set-up page lists the steps like this:

  1. Go to Experiments within the Campaigns menu. Click the plus button in the “All experiments” table and choose Custom.
  2. Select an experiment and the variable to test, then go to Next.
  3. Select the original campaign. Only eligible campaigns are listed.
  4. In the “Success metrics” section, choose a metric.
  5. If you like, turn on “Auto-apply settings” and choose how to decide the winner: directional results, or a requirement for statistical significance.
  6. Enter the start date, the length and a name for the experiment, then select Schedule.

There is a shortcut too: if you edit an active campaign, change the settings and choose Save as experiment, an experiment reflecting the new settings is created for you. The set-up page also lists a few limits: you can schedule up to 5 experiments and one experiment can include several campaigns, but only one experiment runs at a time. If the original campaign has a shared budget, remove it first.

On campaign types, Google’s pages do not say the same thing. The set-up page says custom experiments are only available for Search, Display, Video and Hotel Ads campaigns and cannot be created for App or Shopping campaigns. The custom experiments page lists Search, Display, Demand Gen and Video campaigns; it also says App and Shopping are not supported. The Experiments page, however, says custom experiments are available for App campaigns and Search and Display campaigns. The campaign selection list in your own account shows which applies to you.

There is one more note on eligibility. According to the campaign selection step on the set-up page, only eligible campaigns are shown: the campaign must be active, use Smart Bidding and include exact or phrase match keywords. The same note says the campaign cannot take part in another experiment or run on the Google Display Network; that last condition does not agree with the list of campaign types at the top of the same page.

Reviewing the ads can take a little while, so the page suggests scheduling the experiment for a future date. That way it does not start before the ads are approved.

4. Choose how to split the traffic

According to the set-up page, the experiment shares the original campaign’s traffic and budget, and you decide the percentage it gets. In Search campaigns you can use a cookie-based or a search-based split. In Display campaigns Google always uses a cookie-based split, so a user sees either the experiment or the original, never both.

Even an equal split does not produce equal impressions. The custom experiments page puts this down to the auction: every ad space is a new auction, so one side can win more impressions. The page advises waiting 7 to 14 days for the experiment arm to settle. According to the set-up page, if you use audience lists with the cookie-based split, the list should have at least 10,000 users; with fewer, the results may be less accurate.

5. Set the length and the waiting rule

Three Help pages give different guidance on length, so we report all three. The edit page says the initial end date of a custom experiment must be 2 to 12 weeks after the start date, and you can later extend it to at most 12 weeks from the current date. The monitoring page recommends letting an undecided experiment run for 2 to 3 weeks to gather data. The Experiments page recommends at least 4 to 6 weeks if the results cannot yet be determined.

We suggest counting the settling time within those ranges: a two-week experiment is too short once the 7 to 14 day settling period is taken out. Work in full weeks, because the days of the week behave differently. And stay still while it runs: the set-up page says changes made to the original are not reflected in the experiment, and changes made while it runs make the results harder to interpret. The edit page also says to keep the bid type and rules the same in both arms and not to change them mid-test.

6. Read the result and its confidence

According to the monitoring page, the scorecard gives two values for each metric. The first is the estimated difference of the experiment against the original; for “Clicks”, +10% means the experiment is estimated to get 10% more clicks. The second is the confidence interval: the range the difference may fall in at the level you chose. The default level is 80% and you can pick your own. If the result is statistically significant, a blue asterisk appears. The “Results” column says whether the control campaign or the treatment campaign performed better, or that there is no clear winner yet.

The table shows three possible outcomes. All three cards are examples we made up; the percentages do not come from a real account.

What the card shows What it means What you do
Conversions +12%, interval [+4%, +20%], blue asterisk The whole interval is above zero; the difference is unlikely to be chance Apply it and check a guardrail metric, such as cost per conversion
Conversions +3%, interval [-6%, +12%], no asterisk The interval includes zero; the difference could be positive or negative Do not call a winner; there was not enough time or traffic, or the change had no effect
Conversions -9%, interval [-15%, -3%], blue asterisk The experiment is clearly worse Do not apply; end it and write down why

Help lists four possible reasons why a result is not statistically significant: the experiment has not run long enough, the campaign does not get enough traffic, the traffic split is too small, or the change did not make a significant difference. A flat result is still a result: this change, over this period, with this traffic, produced no difference. Put it on record.

7. Apply, convert or end

If you are happy with the result, the set-up page gives two routes: apply the experiment to the original campaign, or convert it into a new campaign. Applying moves the changes you tested into the original campaign. With the second route, according to the custom experiments page, you can pause the original campaigns and convert the experiment into a new campaign with the same dates and budget as the original. According to the set-up page, the experiment’s performance data is kept in both cases. After converting, open the campaign and check that every change came across correctly.

The auto-apply option can be switched on at set-up. According to the monitoring page, the feature is only available for some experiment types, is enabled by default and can be turned off from the “Report” page. We suggest keeping it off in a small account and making the call yourself.

To end an experiment, find it and choose End experiment. According to the edit page it finishes on its end date, or you can end it early with End now; after it ends, you cannot resume it or edit the end date. So note the figures in the table before you end it.

The first tests for a small account

In an account with few conversions it is hard to catch fine differences, and with little traffic the experiment is more likely to end without a significant result. We suggest starting with changes that are large:

  1. A bidding strategy move. If conversions are measured, compare the current strategy with the next stage (for example Maximize conversions against Target CPA). Seeing it in an experiment is safer than switching a live campaign.
  2. A landing page. Send the same ad to two different pages, and make the pages clearly different from each other.
  3. A match type. Try it on a single campaign with a single change.

Two experiments cannot run at the same time, so go in order and write every result in a notebook.

Common mistakes

  • Changing the bidding and the landing page in the same experiment.
  • Closing the experiment in the first week and calling it a win.
  • Setting up a conversion-metric test without checking conversion tracking.
  • Changing budget, target or audience settings on the original campaign while the experiment runs.
  • Not recording a flat result.
  • Not checking, after converting, that the settings came across correctly in the new campaign.

Common questions

What is a Google Ads custom experiment?

It is a test that creates an experiment campaign from the original, tries a change with part of the traffic and budget, and compares the result with the original over the same period. Help says it lets you measure the effect of changes before applying them to campaigns.

How long should an experiment run?

According to the edit page, the initial end date must be 2 to 12 weeks after the start. The monitoring page recommends 2 to 3 weeks and the Experiments page at least 4 to 6 weeks if the result cannot yet be determined. We suggest counting the settling period too and working in full weeks.

Is data lost when the experiment ends?

No. The set-up page says the experiment’s performance data is kept, whether you apply the experiment to the original or convert it into a new campaign. You just cannot resume an experiment that has ended.

One thing you can do today: open the Experiments page in your account, check whether an experiment is running or scheduled, then write the single change you want to test and its success metric in one sentence. If you cannot write the sentence, the experiment is not ready to be set up. If you would like us to run experiments for you, our advertising service does this with a written test log.

Sources

This article is general information, not legal, financial or technical advice. Platforms, their rules and the law change over time, so some details may have changed since 15 November 2026 or contain small errors. Before you make a decision for your business, check the current official sources or write to us.

Tell us about
your project.

Fill in the form or write to us directly, and we will agree on a day to talk.

Your first meeting is with Deniz Sezer, founder of Detmore.

First meeting free · Reply within one working day · Written proposal in 2 to 3 days · NDA ready

What do you need? pick as many as you like

Name, email and phone are required.

Contact

We reply within one working day.