Drafts and experiments: safe test environment for live campaigns in 2026

Every change made directly to a live campaign is a calculated risk. You adjust a setting, and only discover afterwards whether the impact was positive or negative. For smaller advertisers this risk is limited, but for campaigns with substantial daily budgets, unvalidated changes can cost significant performance. That is exactly why Google Ads built Drafts and Experiments: a fully integrated testing environment that allows you to safely validate campaign changes before they go live. This article explains what Drafts and Experiments are, how to use them effectively, and how AdBrains uses its own AI technology to automate and accelerate this testing process.

What are Drafts and Experiments in Google Ads?

A Draft is a concept version of an existing campaign. You create a copy of your live campaign, make changes within that copy, and then decide whether to apply those changes directly to the live campaign or first validate them through an experiment. A Draft does not affect the live campaign in any way; it is purely a safe editing space without consequences.

An Experiment is the next step: you activate the Draft as a parallel campaign that receives a portion of the live traffic. That traffic is divided through a traffic split, where you determine what percentage of users sees the experimental variant and what percentage continues to see the original campaign. The most commonly used split is 50/50, though 70/30 or 80/20 are also possible depending on how cautious you want to be about protecting existing performance.

The major advantage of this approach is the purity of the test environment. Because the traffic split happens at the cookie level, each user consistently sees the same variant throughout their entire session and any subsequent retargeting period. This prevents results from being skewed by the same user seeing different variants at different times, making outcomes statistically more reliable than a simple week-on-week comparison.

What can you test with an Experiment?

Google Ads Experiments are applicable to a broad range of campaign elements. The most valuable practical applications include the following:

• Smart Bidding strategy: Considering switching from Target CPA to Target ROAS, or from manual CPC to an automated Smart Bidding strategy? Test it via an experiment before fully committing to the switch.
• Ad copy (RSA): Run two sets of Responsive Search Ads side by side with different headlines or descriptions to determine which delivers the highest CTR and conversion rate.
• Landing page URL: Test whether a specific product page, category page, or dedicated landing page converts better for the same keywords.
• Keyword match type: Compare the performance of broad match versus phrase match or exact match for the same keywords, including the effect on search volume, CPC, and conversion quality.
• Audience targeting: Add an RLSA audience to the experimental variant and measure whether bid adjustments based on previous site visits improve ROAS.
• Campaign structure: Test a different ad group organisation, such as more specialised ad groups per theme versus broader, combined ad groups.

For every variable tested, the same principle applies: the live campaign continues running undisturbed, while the experimental version runs alongside it on a shared or split budget. Once the experiment has collected statistically significant data, you can easily decide whether to apply the winning variant to the live campaign or stop the experiment and retain the current settings.

When and for whom are experiments most valuable?

Experiments are particularly powerful when the stakes are high and uncertainty is significant. Think of campaigns with substantial daily budgets where a wrong decision immediately affects cost per lead or ROAS. They are also essential when introducing new Google Ads features, such as transitioning to a Performance Max campaign alongside an existing Search campaign, where an experiment provides the confidence needed to scale responsibly.

For e-commerce advertisers like ToetsJeKennis.nl, testing Smart Bidding strategies via experiments is a highly effective approach. For instance, if the campaign currently runs on Target CPA with stable conversion costs, but there is a question of whether Target ROAS might better suit a product mix with varying margins, an experiment can answer that question without endangering existing stable performance. For lead generation campaigns such as those run by Clima-Active.nl for air conditioning and heat pump installations, experiments are valuable for finding the right bidding strategy for high-quality quote requests. With higher CPL and significant per-installation margins, even a small improvement in conversion rate can have a major impact on profitability.

How to correctly set up an Experiment

Setting up a reliable experiment requires a structured approach. The following steps help you create a trustworthy test:

1. Define one test variable: Never test multiple elements simultaneously in one experiment. If you change both the bidding strategy and the ad copy, you cannot determine afterwards which change drove the results.
2. Choose a representative duration: An experiment needs at least two to four weeks to collect sufficient data, depending on the campaign's conversion volume. Shorter tests produce statistically unreliable results.
3. Set a relevant primary metric: Decide in advance which KPI determines the winner: CPA, ROAS, CTR, or conversion rate. Do not change this criterion midway through the test.
4. Use a fair traffic split: A 50/50 split is the most equitable and achieves statistical significance fastest. If you want to limit risk to the live campaign, use 70% original and 30% experimental variant, but note that the test will take longer to reach significance.
5. Ensure accurate conversion tracking: Experiments are only reliable when conversion tracking is correctly set up. Use Enhanced Conversions and consider server-side tracking to capture all conversions as completely as possible.
6. Document your hypothesis upfront: Record why you expect the variant to perform better. This forces clear thinking and helps you learn from the outcome regardless of the result.

The AdBrains AI approach: automated experiment management

At AdBrains, we do not manage Google Ads campaigns manually but through our own AI technology that makes and executes dozens of optimisation decisions daily. Experiments are no exception to this. Our AI integrates testing through Drafts and Experiments as a structural component of the campaign lifecycle, delivering demonstrably faster and more reliable results than ad-hoc testing by a human campaign manager.

The process starts with our multi-agent verification system. Every optimisation decision, including whether an experiment should be launched, is evaluated by four independent AI agents before anything is executed. This prevents experiments from being set up on insufficient data or an incorrect hypothesis. The agents verify that conversion volume is high enough for a statistically reliable experiment, that the planned duration is realistic, and that the test variable is genuinely isolated from other campaign changes.

Our automated tCPA/tROAS optimisation actively uses experiments to safely validate bidding strategy switches. When our AI detects that a campaign is ready to scale from Target CPA to Target ROAS, it does not directly modify the production campaign. Instead, a Draft is automatically created, an experiment is activated with a predetermined traffic split, and results are monitored daily. Once statistical significance is reached and the experimental variant meets the set performance standards, our AI automatically applies the winning variant to the live campaign. Large strategic switches are thus executed on data, not intuition.

For the Keyword Incubator, a similar logic applies. New keywords are first safely tested in a separate incubator campaign before being promoted to the production campaign. This is essentially a keyword-level experiment: if new keywords perform sufficiently in terms of CPA or ROAS, they are automatically transferred. If a keyword underperforms, it is paused or added as a negative keyword. This significantly reduces the risk associated with introducing new keywords.

For clients like E-4motion.com, the webshop for new electric folding bikes, and LeroyBrouwer.nl, this means campaign experiments are not dependent on the availability or expertise of a human campaign manager. The AI runs continuously, identifies testing opportunities, executes them in a structured way, interprets results, and applies or rejects them based on data. This creates a faster iteration cycle and therefore faster campaign improvement, without the risks of unvalidated live changes.

Common mistakes when using Experiments

Despite the power of this tool, the same mistakes appear repeatedly in practice, undermining the value of experiments. The most common are listed below:

• Duration too short: A five-day experiment rarely yields statistically significant results, especially for campaigns with modest conversion volume. Patience is a prerequisite.
• Testing multiple variables at once: If you change both the bidding strategy and the landing page in the experimental variant, it becomes impossible to determine which element drove the difference.
• No predefined success criteria: Without predetermined KPIs and thresholds, there is a strong temptation to stop the experiment as soon as the desired direction appears, even if it is not statistically supported.
• Forgetting to monitor experiments: An experiment that continues running for months after reaching significance wastes budget and delays the implementation of improvements.
• Failing to document: Without recording the hypothesis, results, and conclusions, the team learns nothing from the test, and the risk exists that the same test is run again in the future.

Overview: what can you test with Drafts and Experiments?

The table above provides a realistic overview of the most common experiment applications in Google Ads Search and Shopping campaigns. The durations are guidelines; campaigns with high conversion volume reach statistical significance faster, while campaigns with fewer conversions need more time.

Frequently asked questions about Drafts and Experiments

What is the difference between a Draft and an Experiment in Google Ads?

A Draft is a concept version of an existing campaign, serving as an editing space without any impact on the live campaign. You can create, edit, and delete a Draft at any time without affecting your running campaigns. An Experiment is the activation of a Draft as a parallel campaign version, diverting a portion of live traffic to the experimental variant. The experiment version actually runs live, but only for the configured percentage of traffic. Together, both tools form the structured testing infrastructure of Google Ads.

How much budget does an experiment use?

With a 50/50 split, the experiment shares the budget of the original campaign. Google distributes the total available budget across both variants based on the configured traffic split. This means you do not need to reserve additional budget for an experiment, but each variant will on average receive half the normal budget. For campaigns with a limited budget, this may extend the time needed to reach statistical significance. Consider a temporary budget increase during the experiment, or opt for a shorter but more focused test with a clearly measurable primary KPI.

Can I automatically apply the winning variant?

Yes. After completing an experiment, Google Ads offers the option to directly apply the winning variant to the original campaign. This works smoothly for most element types tested via the Draft framework. For complex changes such as a full restructuring of ad groups, manual implementation may still be required. AdBrains AI continuously monitors experiment results and initiates the application of the winning variant once statistical significance is reached, always running an additional verification round through our multi-agent system before the change is permanently applied.

How long does an experiment need to run to produce reliable results?

A reliable experiment needs a minimum of two weeks, and for campaigns with lower conversion volume, four to six weeks is more realistic. The industry rule of thumb is that you need at least thirty conversions per variant to reach statistical significance. Google Ads itself shows an indication of statistical reliability within the experiment interface, expressed as a percentage. Only when this percentage exceeds 95% should you consider the results statistically significant and draw a well-founded conclusion. Never stop an experiment early based on an interim trend, as this leads to false-positive conclusions.