Google Ads Experiments: How to Test Campaign Changes Properly

Google Ads experiments let you test a change on part of a campaign's traffic while the rest keeps running unchanged, so you can see whether the change actually helped before you roll it out. You set them up from the Experiments page in the Campaigns menu, choose how traffic or budget is split, let the test run for several weeks, and then apply the winning version or end the test without changing anything.

That sounds simple, and the setup screens are short. The part that takes judgement is everything around them: which experiment type fits the question, how the split is assigned, what experiment sync quietly does to your trial, how long to wait, and how to read a result that Google marks as significant or not. This guide walks through each of those using the mechanics Google documents in its own Help Center.

What are Google Ads experiments?

An experiment compares two versions of something under the same conditions at the same time. The control is your current campaign or ad. The treatment (Google also calls it the trial or the experiment arm) is the version with your change. Because both arms run in the same auctions over the same days, things like seasonality, competitor activity and news events hit both sides equally. That is the whole point, and it is why an experiment is more trustworthy than comparing last month to this month.

Google groups its tests on the Experiments page, which it now describes as part of the Experiment Center alongside lift studies. The experiment types listed there include:

• Custom experiments for testing bidding, match types, landing pages, audiences and ad group changes on a copy of a campaign.
• Ad variations for testing text changes across responsive search ads in one or many Search campaigns.
• Video experiments for comparing video ads or settings against the same audience on YouTube.
• Performance Max experiments for measuring uplift from Performance Max or testing a PMax setting.
• AI Max experiments for testing AI Max for Search features inside an existing Search campaign.

There are also App asset experiments and Demand Gen experiments, which follow their own flows. Campaign drafts are a related but separate tool: a draft lets you stage changes to a Search or Display campaign and apply them later, without splitting any traffic.

Custom experiments

A custom experiment creates a trial campaign that copies your base campaign, then lets you change whatever you want to test in the copy. The two share the original campaign's traffic and budget for the dates you choose. Google's Help pages list Search and Display as supported, along with some Video setups (auction-based campaigns only, not YouTube Select or Masthead). Shopping campaigns are not supported for custom experiments; to compare Standard Shopping with Performance Max, you use a Performance Max experiment instead.

The setup flow is: go to Experiments, click the plus button, choose Custom experiment, pick the campaign type, name the experiment, select the base campaign, and add a suffix so the trial campaign is easy to spot. You then pick up to 2 success metrics (for example Clicks with the goal "Increase"), set the split, choose the advanced split option for Search, and set the dates. You can also start one from a campaign you are editing by choosing "Save as experiment".

Some campaigns can't be used. Google lists shared budgets, Portfolio bid strategies and campaigns still holding deprecated text ads or expanded text ads among the blockers, and the base campaign has to stay active for the experiment to run. You can schedule up to 5 experiments for one campaign, but only one can run at a time.

Ad variations

Ad variations are the lighter option for ad copy. Instead of copying a campaign, you describe a change and Google applies it to the matching responsive search ads for the test period. There are three variation types: find and replace (swap one piece of text for another, for example "Book now" for "Call now"; Google notes the matching is case sensitive), update text (add, remove or pin headlines and descriptions), and update URLs (final URLs, display paths and final mobile URLs).

You can run a variation across all campaigns or a single campaign, and filter which ads it touches by headline, description, display path or final URL. The split is a percentage of the campaign's budget, and users are assigned by cookie, so a given person tends to see only one version however many times they search. Once a variation has started, its start date is locked. When you apply it, Google creates new ads with the change. Google's own guidance is to use ad variations for creative tests across a whole campaign and custom experiments when you want to test specific ad groups.

Video experiments

Video experiments compare video ads or settings against the same audience on YouTube. The basic path is an A/B test of video assets: two arms, where creative is the only thing that differs, available for Video Reach and Video View campaigns. The advanced Custom path lets you test other variables across more arms. The treatment arm duplicates the control's audiences, bids and formats, and budget or date changes to the control flow through to the treatment automatically, except for the creatives being tested. With two arms, the default is an even 50% split per arm.

The metrics include Absolute Brand Lift (which needs a Brand Lift study and isn't available to every advertiser), CTR, conversion rate, cost per conversion, CPC, CPM, CPV and view rate. Video experiments have to be ended manually.

Performance Max experiments

Performance Max experiments come in three families. Uplift experiments measure the extra conversions or conversion value that Performance Max adds next to your other campaigns. Upgrade experiments simulate moving budget from an existing campaign into PMax, for example Standard Shopping versus Performance Max, or a Dynamic Search Ads or Display campaign versus PMax. Optimization experiments test settings inside a PMax campaign, such as final URL expansion or asset testing. Google may also suggest these through recommendation cards when a campaign has enough conversion volume to be eligible.

AI Max experiments

An AI Max experiment splits an existing Search campaign: one share runs with the AI Max toggle off (the control), the remaining share runs with it on. It tests features such as search term matching and asset optimization. The flow is blocked if the campaign has text customization turned on, targets the Display network, belongs to a Portfolio bid strategy, uses a shared budget, uses bidding exploration, or already has an active experiment. If you end it without applying, the AI Max settings go back to how they were before the test.

How do traffic split and cookie-based vs search-based split work?

Two separate decisions hide behind the word "split": how much goes to each arm, and how users are assigned to an arm.

For the amount, Google uses a budget split for some Search, App and Display experiments and a traffic split (a share of eligible auction entries) for Performance Max, Demand Gen and Video experiments. Google recommends a 50% split for custom experiments because it gives the cleanest comparison. A smaller treatment share protects the control, but it collects data more slowly. The split can't be changed after setup; a different split means a new experiment.

A split decides auction eligibility, not results. Google is explicit that a 50/50 split does not guarantee equal impressions or spend, because Ad Rank, bids, budget limits and the bid strategy's learning all move exposure around. For Performance Max, Google's FAQ says a spend gap of under 20% between arms counts as normal auction behaviour.

For assignment, Search custom experiments offer two options under the advanced settings:

• Cookie-based (Google's recommendation) randomly assigns each user to one arm, and that user then sees only the control or only the trial.
• Search-based randomly assigns each search. A person who searches several times can land in both arms, which can reach significance faster but mixes exposure.

Display custom experiments always use cookie-based assignment. Use cookie-based when the change affects a user over several visits (landing pages, bidding toward a conversion that happens days later). Search-based can make sense for changes that act on a single query, such as a match type test, when you need the extra speed.

What does experiment sync do?

Experiment sync exists for custom Search and Display experiments, and it is on by default. With sync on, changes you make to the base campaign after creating the trial are copied into the trial, usually within about 15 minutes. It only works one way: edits to the trial never flow back to the base.

You can only turn sync off while creating the trial. It can't be switched on or off later for a trial that is scheduled or running. Two details catch people out. Ads are shared between the arms, so editing an ad in the control changes it in the treatment whatever the sync setting. And bulk ad uploads change both campaigns even with sync off. If a sync error happens, syncing stops for the rest of the trial and any later base changes have to be copied by hand.

The practical rule: leave sync on if you want routine maintenance (new negatives, paused keywords) to hit both arms, and turn it off only when the change you are testing is something the sync would overwrite. If you do turn it off, every base-campaign change during the test needs a matching manual change in the trial, or the comparison drifts.

How long should a Google Ads experiment run, and how do you read significance?

Google's experiments FAQ recommends at least 4-6 weeks, and longer if your conversions take a while to happen after the click. Its monitoring page mentions 2-3 weeks as a minimum to gather data, and the custom experiments page says to allow 7-14 days for the treatment arm to stabilize. The FAQ also says the first 7 days of data are discarded from some experiment results to cover ramp-up. Put together: plan for a month or more, and treat the first one to two weeks as a ramp-up period in which Google says data can be volatile.

The results scorecard shows, under each metric, the experiment's own total and then the estimated difference against the control, for example +10%, followed by a range such as [+8%, +12%]. That range is the confidence interval at the confidence level you have chosen. The default is 80%, and dynamic confidence reporting lets you pick a different level. A "--" means there isn't enough data yet. When a result is statistically significant, Google adds a blue asterisk. Hovering over the line explains the interval and shows the underlying numbers for both arms.

Video experiments use a different scale. Directional results can appear at 70% confidence or 80%, and Google calls a result conclusive at 95% and recommends waiting for that.

If a result is not significant, Google lists four possible reasons: the test hasn't run long enough, the campaign has too little traffic, the split was too small, or the change made no real difference. The last one is a valid outcome. A flat result suggests the change made no measurable difference at that traffic level, which is useful in itself.

Applying vs ending an experiment

When a custom experiment finishes, you open it and click Apply. You then choose between two options:

• Update your original campaign. The experiment ends and its changes are written into the base campaign.
• Convert to a new campaign. The trial becomes a normal campaign with the original's budget and dates, and the original campaign is paused.

Either way, the performance data for both the original and the experiment is kept. Google's FAQ also describes auto-apply, which is on by default for new experiments: if the results are favourable, the changes are applied to the control automatically. What counts as favourable depends on the bid strategy (with Maximize conversions and a target CPA, the treatment needs more conversions at a lower CPA). Auto-apply only acts on experiments that end on the end date you set, so a test you stop early is not applied. You can opt out before the experiment ends.

To stop a custom experiment early, change its end date or choose Edit and then End now. For AI Max experiments you can pick "Apply your experiment changes if results are favorable" at setup. In Performance Max upgrade experiments, applying makes the PMax campaign primary with 100% of the traffic and pauses the control.

What you can and cannot test

Custom experiments are the broad tool. Google names bidding, match types, landing pages, audiences and ad group changes as typical tests. Ad variations handle copy and URL changes in responsive search ads. Video experiments handle creative and settings on YouTube. Performance Max and AI Max experiments handle the automated campaign features those products control.

Some things are off the table or come with gaps. Shared budgets and Portfolio bid strategies block custom and AI Max experiments. Custom experiments don't support several reports, including auction insights, search terms reports, bid landscapes and display placement reports, so plan to judge the test on the scorecard metrics. Shopping campaigns can't run custom experiments. And only one custom experiment can run on a campaign at a time.

Experiments vs before/after comparisons

The common alternative to an experiment is to make a change and compare the weeks after with the weeks before. It is quick, but it can easily mislead. Anything else that changed between the two periods (a holiday, a competitor's sale, a price change on your site, a shift in search demand, a bid strategy still learning) ends up attributed to your change. You can't separate them afterwards.

An experiment fixes this because both arms live through the same period. The cost is that each arm gets only part of the traffic, so you need patience. If a campaign is too small to reach significance in a reasonable time, a before/after comparison is the fallback, but treat it as a weak signal and write down everything else that changed in the period.

Mistakes that invalidate a test

• Editing during the test. Google advises against changes while an experiment runs, and especially during the 7-14 day learning phase, when changes can restart learning. Budget changes are allowed but should be rare.
• Forgetting what sync does. With sync on, a base-campaign change you meant only for the control lands in the trial too. With sync off, a base change you forgot to copy makes the arms differ in two ways instead of one.
• Testing several things at once. If the trial changes the bid strategy and the landing page together, a win can't be attributed to either. Test one variable per experiment.
• Overlapping experiments. Google's FAQ recommends running experiments one after another rather than in parallel, because they can interfere with each other.
• Calling it early. A big swing in week one, before the ramp-up period is over, is not a result. Wait for the asterisk, or for the planned end date.
• Reading spend differences as performance. Uneven spend between arms is normal under a split. Compare the metrics you chose as success metrics, not raw cost.
• Too small a split on a small campaign. A 10% treatment on a low-volume campaign may never reach significance.

What this means for advertisers in Thailand

The mechanics are the same in every market, but a few local factors matter when you plan a test. If you run separate campaigns for Thai-language and English-language keywords, each one has less volume than a combined campaign would, so an experiment on one of them takes longer to reach significance, and you may need a 50% split and a longer run than the 4-6 week minimum.

The calendar matters too. If a test window includes a holiday such as Songkran in April or the year-end period, a before/after comparison across those weeks mixes the holiday with your change. A proper experiment running through a holiday still compares both arms under the same conditions, but the result describes holiday traffic, so check whether it holds after the period ends. If your conversions come through LINE chats or phone calls that you import later, run the test long enough for those delayed conversions to arrive before judging it.

FAQ

Can I change the split after an experiment starts?

No, the traffic split can't be changed once the experiment is set up; Google's FAQ says a different split needs a new experiment. Decide the split before launch, keeping in mind that a smaller treatment share takes longer to reach significance.

Should I use cookie-based or search-based splits?

Use cookie-based for most tests, which is what Google recommends, because each user sees only one version. Search-based can reach significance faster, but the same person can see both arms, so it suits changes that act on a single search.

How long should I run a Google Ads experiment?

Plan for at least 4-6 weeks, which is Google's recommendation, and longer if conversions are delayed. Google discards the first 7 days of some results for ramp-up, so short tests have even less usable data than the calendar suggests.

What does the blue asterisk mean in experiment results?

The blue asterisk means the difference between the arms is statistically significant at the confidence level you selected. The default level is 80%, and you can choose another one through dynamic confidence reporting.

What happens to my original campaign when I apply an experiment?

It depends on the option you choose: "Update your original campaign" writes the changes into it, while "Convert to a new campaign" pauses it and keeps the trial running instead. Performance data for both is preserved either way.

Experiments give the cleanest answer to whether a bidding change, a new landing page or AI Max is worth adopting, but they reward planning more than speed. If you want help choosing what to test and reading the results, talk to our Google Ads management team, or see our Relevant Search Google Ads service for Search campaigns. For YouTube creative tests, see our YouTube ads service, and for Shopping or Performance Max upgrade tests, our Google Shopping service.