Every Google Ads account eventually reaches the same crossroads: a change looks promising, but nobody wants to gamble the whole budget on it. Switching a campaign from Maximise Clicks to Target CPA, rewriting every headline or moving traffic to a new landing page can lift results — or quietly drain a month of spend before anyone notices.
Google Ads experiments solve that problem. Instead of editing a live campaign and hoping, you run the change on a share of the campaign's traffic, compare it with the untouched original and only roll it out when the numbers back it up. It is the closest thing paid search has to a controlled trial.
In this guide you will learn how custom experiments work, which changes are worth testing, how to set one up step by step, how long to run it and how to read the results without fooling yourself.
Key Takeaways
- An experiment copies a campaign, changes one thing and splits the original campaign’s traffic and budget between the two versions.
- Custom experiments are available for Search, Display, Video and Hotel campaigns; App and Shopping campaigns are not supported.
- Use a 50/50 split and cookie-based assignment on Search when you want the cleanest comparison.
- Test one variable at a time and decide your success metric before you launch, not after.
- Let an experiment run long enough to cover full weekly cycles and conversion lag before you apply it.
What Is a Google Ads Experiment?
A Google Ads experiment is a controlled copy of an existing campaign. You start from the original (often called the base campaign), make one change in the experiment arm and let Google divide the traffic between the two. Because both versions run at the same time, in the same auction conditions, seasonality and competitor activity affect them equally. That is what makes the comparison fair.
“The experiment shares your original campaign's traffic (and budget).”
— Google Ads Help, Set up a custom experiment
This is the key difference from a simple before-and-after comparison. If you change a bid strategy on the first of the month and compare it with the previous month, you cannot separate the effect of your change from everything else that moved: holidays, a competitor pausing, a news story, a price change on your site. An experiment removes most of that noise.
Experiments vs drafts vs ad variations
Google Ads has several related tools, and they are easy to mix up:
- Drafts let you stage changes to a campaign without making them live. On their own they do not test anything; they are a workspace.
- Custom experiments take a campaign (or a draft of it) and run it against the original with a traffic split. This is the tool for testing bid strategies, keywords, match types, audiences and landing pages.
- Ad variations are designed for testing ad text changes across many campaigns at once, such as swapping one phrase in headlines account-wide.
| Tool | What it does | Best for |
|---|---|---|
| Campaign draft | Stages edits privately without serving them | Preparing complex changes before testing |
| Custom experiment | Runs a modified copy against the original with a traffic split | Bid strategies, keywords, audiences, landing pages |
| Ad variations | Applies a text change to a share of ads across campaigns | Headline or description wording at scale |
| Before/after comparison | Compares two time periods | Quick sanity checks only — not proof |
Which Campaigns and Changes Can You Test?
According to Google's help documentation, custom experiments are available for Search, Display, Video and Hotel campaigns, while App and Shopping campaigns are not supported. You can schedule up to five experiments per campaign, but only one runs at a time. If a base campaign still contains legacy text ads or expanded text ads, Google asks you to remove them before the experiment can be created.
Within those campaign types, the tests that usually earn their keep are:
- Bid strategy changes — for example manual CPC vs Target CPA, or Maximise Conversions vs Maximise Conversion Value. Our guide to Google Ads bidding strategies explains what each option optimises for.
- Match type and keyword changes — moving phrase match keywords to broad match paired with smart bidding is one of the most common experiments today.
- Landing page tests — sending the experiment arm to a new page while the ads stay identical. Pair this with the advice in our PPC landing pages guide.
- Ad copy and assets — testing a new set of responsive search ad headlines against the current set.
- Targeting changes — adding audience segments, widening a radius in your location targeting or adjusting an ad schedule.
- Automation features — for instance, comparing a standard Search campaign with one that adds dynamic search ads coverage.
How to Set Up a Custom Experiment: Step by Step
The interface changes from time to time, but the workflow has been stable. Here is the process we follow for client accounts:
- Check that conversion tracking is reliable. An experiment is only as good as the data it measures. Confirm that your primary conversion actions are firing correctly — see our conversion tracking setup guide if you have any doubt.
- Write a hypothesis. For example: “Switching from Maximise Clicks to Target CPA at £45 will lower cost per lead by at least 15% without cutting lead volume by more than 10%.” A written hypothesis stops you from cherry-picking a metric later.
- Open the Experiments section (under Campaigns) and choose to create a custom experiment.
- Select the base campaign and give the experiment a clear name that says what changed, such as “Brand-Search | tCPA 45 test | Nov”.
- Make one change in the experiment arm. Edit only the setting you are testing. If you change bids, ads and keywords together, you will not know which change caused the result.
- Choose the split. Google recommends a 50% split for the most useful comparison. On Search you can choose a cookie-based split, which randomly assigns each user to one version, or a search-based split, which assigns each individual search.
- Set dates. Because new ads go through review, consider scheduling the start a day or two ahead.
- Monitor, but do not meddle. Check daily that both arms are serving. Avoid editing the base campaign during the test; changes to one side contaminate the comparison.
- Decide and act. When the test ends, apply the experiment to the original campaign, convert it into a new campaign, or end it and keep the original.
Cookie-based vs search-based splits
Cookie-based splitting keeps the same person in the same arm, which matters when users search several times before converting. Search-based splitting can collect data faster, but one person may see both versions. For bid strategy and landing page tests, cookie-based is usually the better choice.
How Long Should an Experiment Run?
There is no universal number, and you should be wary of any fixed rule. What matters is that each arm collects enough conversions to separate a real difference from random variation. A few practical guidelines:
- Run for at least two full weeks so both arms experience the same weekday and weekend patterns.
- Add your typical conversion lag. If most leads convert several days after the click, results from the final days are incomplete.
- Allow a learning period when testing automated bid strategies. Smart bidding needs time to calibrate, so judge the later part of the test more heavily than the first few days.
- Use the statistical significance indicators Google shows in the experiment report rather than eyeballing the difference.
| Situation | Practical guidance |
|---|---|
| High-volume campaign (many conversions per day) | Two to four weeks is often enough to reach a clear result |
| Low-volume lead generation | Expect a longer run, or test on a higher-funnel metric you agree in advance |
| Smart bidding test | Discount the learning period; judge the later weeks |
| Seasonal peak approaching | Start well before the peak or wait until after it |
| Long sales cycle (B2B) | Import offline conversions so you measure qualified leads, not just form fills |
Reading the Results Without Fooling Yourself
The experiment report compares the two arms on metrics such as clicks, cost, conversions and conversion value, and indicates whether differences are statistically significant. Before you apply anything, ask three questions:
- Did the primary metric move in the direction of the hypothesis? Judge on the metric you chose upfront.
- Is the difference significant? If the confidence interval spans zero, you have not proven anything — the honest answer is “no clear winner”.
- Did anything else break? A lower cost per lead is not a win if lead quality dropped. Check CRM data, not just platform conversions; our colleagues cover this in depth in the guide to incrementality testing.
The principle of isolating one variable is not unique to Google. Meta's split testing documentation makes the same point plainly:
“Select only one variable per test.”
— Meta for Developers, Split Testing
Common Mistakes
- Testing too many things at once. A new bid strategy, new ads and a new landing page in one arm produces a result you cannot explain or repeat.
- Ending the test early. A strong first week is often noise. Stick to the planned duration unless performance is clearly damaging the business.
- Editing the base campaign mid-test. Any change to the original arm invalidates the comparison.
- Testing in a low-volume campaign. If a campaign gets a handful of conversions a month, an experiment may never reach significance. Consider consolidating campaigns first — see our guide to Google Ads account structure.
- Measuring the wrong conversion. Optimising to a micro-conversion such as page views can produce a “winner” that never shows up in revenue.
- Forgetting to apply the winner. Experiments that end without being applied simply stop; the base campaign carries on unchanged.
When Not to Use an Experiment
Experiments are powerful, but not every change needs one. Fixing a broken URL, adding obvious negative keywords or correcting a typo does not require a test. Equally, if you want to know whether paid search adds value at all compared with organic results, you need a different design — a geographic or time-based holdout. Google's own researchers used pause studies for exactly this question; their paper on incremental clicks reported that, on average, 89% of paid clicks were not replaced by organic clicks when ads were paused.
Related Guides
- Google Ads Lead Form Assets: Capture Leads Without a Landing Page
- Facebook Lead Ads: How to Get Quality Leads, Not Junk
- Ad Fatigue: How to Spot It and Refresh Your Creatives
Frequently Asked Questions
How much budget does a Google Ads experiment use?
It does not add budget. The experiment shares the original campaign’s budget according to the split you choose, so a 50% split gives each arm roughly half.
Can I run experiments on Performance Max or Shopping campaigns?
Custom experiments are not available for Shopping or App campaigns according to Google’s help documentation. Google offers separate experiment types for some Performance Max scenarios, so check the Experiments page in your account.
How many experiments can run at once?
You can schedule up to five experiments per campaign, but only one can run at a time for that campaign.
What happens to data when I apply an experiment?
Google says the experiment’s performance data is preserved when you apply or convert it, so you keep the history for reporting.
Should I use a cookie-based or search-based split?
Cookie-based splits keep each user in one arm and are usually better for bid strategy and landing page tests. Search-based splits can gather data faster but may show both versions to the same person.
Conclusion: Test First, Then Scale
Google Ads experiments turn opinions into evidence. Write a hypothesis, change one thing, split traffic evenly, wait long enough and judge on the metric that actually matters to your business. Done consistently, a programme of small, well-run tests compounds into a much more efficient account.
If you would like a team to plan and run a structured testing roadmap for your account, our PPC management service can help — or get a free quote and tell us what you want to test first.
