Most Google Ads accounts run “tests” the same way: make a change, watch the numbers for two weeks, decide it worked or didn’t. That’s not testing. That’s superstition with a spreadsheet.
The native Google Ads Drafts and Experiments tool has existed since 2016 and the majority of advertisers have never used it once. Meanwhile, they’re making bid strategy decisions on 200 conversions and calling ad copy “proven” after a three-day spike. We’ve audited hundreds of accounts and the pattern is nearly universal: the instinct to act fast destroys the data quality that would actually justify acting.
This guide is about doing it right — using Google Ads experiments to get answers you can actually trust, without doubling your budget or waiting six months for results.
- The Google Ads Drafts and Experiments tool lets you split traffic between a control and a variation at the campaign level — cleanly, natively, without third-party tools.
- Statistical significance at 95% confidence is your minimum threshold. Anything below that is noise you’re mistaking for signal.
- Run one variable per experiment. Testing a new bid strategy and new ad copy simultaneously guarantees you’ll never know which one moved the needle.
- Most experiments need at least 4–6 weeks and 100+ conversions per variant to produce reliable results — plan your test calendar accordingly.
- Applying a winning experiment takes two clicks. Rolling back a losing one takes one. There is no good reason not to be testing something right now.
What Google Ads Drafts and Experiments Actually Does (And Why Most People Ignore It)
Here’s the setup: you create a Draft of an existing campaign, make your changes to the draft, then launch it as an Experiment. Google splits your campaign’s auction traffic between the original (control) and the draft (experiment) at whatever percentage you define — typically 50/50.
Both variants run in the same auction, targeting the same keywords, competing for the same impressions. That’s what makes this a real test. You’re not comparing January performance to February performance and pretending seasonality doesn’t exist. You’re comparing two versions of reality running simultaneously.
So why does almost nobody use it? Partly because it lives buried under “Campaigns → Experiments” in the left nav and Google has never pushed it aggressively. Partly because it requires patience, and patience is the rarest resource in a PPC account. And partly because if you run a proper test and the variant loses, you have to admit your hypothesis was wrong — which is harder than just never getting a clean answer.
The Three Things Worth Testing (And the One Thing Everybody Wastes Tests On)
Not everything deserves an experiment. Some things are just best practices you should implement and move on. Here’s where campaign experiments earn their keep:
1. Bid Strategies
This is the highest-value use case and it’s not close. Switching from Manual CPC to Target CPA, or from Target CPA to Target ROAS, is a bet with significant consequences. A bad transition can collapse conversion volume for weeks while Smart Bidding re-learns. An experiment lets you test the new strategy on 50% of traffic while the other 50% holds steady — so if the wheels fall off, you still have a functioning campaign.
Run bid strategy experiments for a minimum of 4–6 weeks with a goal of at least 100 conversions per variant. Smart Bidding needs roughly 30–50 conversions just to exit the learning phase, so anything shorter is measuring the learning phase, not the strategy.
2. Ad Creative
Responsive Search Ads already do internal asset rotation, but that’s not the same as a controlled test. If you want to know whether a pain-point headline outperforms a benefit-led headline across your entire campaign — not just for the subset of impressions Google chose to test it on — an experiment gives you that control.
Set your Ad Rotation to “Do Not Optimize” during creative experiments. Otherwise Google’s own optimization will contaminate your data by disproportionately serving the ad it predicts will win before you have the sample size to actually know.
3. Landing Pages
Changing the Final URL in an experiment variant is one of the cleanest split testing Google Ads setups you can run. Same keywords, same ads, same bid — the only variable is where people land. If your new landing page beats your control by 20%+ CVR with statistical significance, that’s a legitimate finding worth acting on.
The Waste: Testing Tiny Creative Tweaks
Using a full campaign experiment to test whether “Get a Free Quote” beats “Request a Free Quote” in a headline is overkill and will almost certainly produce no significant result. RSA asset testing handles micro-copy variations. Save experiments for changes that could genuinely move your CPL or ROAS by 15% or more.
Setting Up an Experiment: The Exact Steps
Navigate to Campaigns → Experiments in the left-hand nav. Click the blue plus button, select “Custom experiment,” then choose the campaign you want to test. You’ll create a Draft first — an editable copy of your campaign where you make your single change.
Key settings to configure intentionally:
- Traffic split: 50/50 is standard. If you’re nervous about risk (testing a radically different bid strategy, for example), go 70/30 in favor of the control. Just know a smaller experiment slice means you need more time to reach significance.
- Cookie-based vs. search-based split: Cookie-based keeps the same user in the same variant across sessions — better for measuring full-funnel behavior. Search-based randomizes per query — better for high-volume campaigns measuring click-through rate.
- End date: Set one. An experiment without an end date is a second campaign you forgot you were running.
- Primary metric: Pick it before you start. Choosing your success metric after you see the data is how confirmation bias disguises itself as analysis.
Reading Your Results Without Fooling Yourself
Google will show you a significance level next to each metric in your experiment results — typically represented as stars or a percentage. Your threshold is 95% confidence. Below that, you do not have an answer. You have a trend. Trends don’t justify permanently changing your campaign.
A few things that will undermine your results even when the confidence level looks right:
Novelty effect: New landing pages and new ad copy often see a short-term lift simply because they’re new. A one-week result that looks incredible frequently regresses by week three. This is another reason your minimum test duration is four weeks, not ten days.
External events: A competitor going dark, a promotional email blast, a news event in your industry — any of these can skew one variant’s window. If something major happens mid-experiment, either extend the runtime to wash out the noise or restart the test.
Looking too early: Checking your experiment results every day and mentally “calling it” based on day-four data is the most common mistake. Commit to your end date before you start, and don’t peek with intent to act until you hit it.
What to Do When the Experiment Ends
You have three options when a Google Ads experiment concludes:
- Apply the experiment: Google merges the winning variant’s settings into your original campaign. The experiment campaign is removed. This is instant and clean.
- Keep both running: Now you have two campaigns splitting traffic indefinitely. This is almost never the right call — it’s just avoiding a decision.
- End without applying: The control survives unchanged. The experiment is deleted. The right move when the variant lost or the result was inconclusive.
An inconclusive result is not a failure. “We tested moving to Target ROAS and it didn’t beat manual CPC at this budget level” is genuinely useful information. Document it, note the conditions (spend level, conversion volume, time of year), and revisit in six months when your account has more data.
Building a Test-and-Learn Calendar That Actually Gets Used
The accounts that compound results over time aren’t running more tests simultaneously — they’re running one test at a time, consistently, quarter after quarter. Here’s a structure that works:
Aim for two to three completed experiments per quarter per campaign. That means your test calendar should have a new experiment launching every three to four weeks. Keep a simple log: hypothesis, start date, end date, primary metric, result, action taken.
That log becomes one of the most valuable documents in your account. When a new team member joins, when your agency does a QBR, when you’re trying to explain why you’re running Target ROAS at this specific target — the log shows the reasoning, tested and documented. It also shows clients and stakeholders that decisions aren’t arbitrary, which is worth more in a retainer relationship than almost any single optimization.
Prioritize your test queue by potential impact × confidence in the hypothesis. A landing page change you’ve seen lift CVR by 25% at other clients in your vertical belongs in the first slot. Testing whether to include your phone number in the ad description does not.
Frequently Asked Questions
How long should I run a Google Ads experiment?
Four weeks is the practical minimum for most campaigns. For bid strategy experiments involving Smart Bidding, six weeks is safer — you need time for the algorithm to exit its learning phase before you start measuring real performance. The harder rule is conversion volume: aim for at least 100 conversions per variant before drawing conclusions.
What’s the difference between Google Ads Drafts and Experiments?
A Draft is a staging copy of your campaign where you make changes without them going live. An Experiment is what you launch when you’re ready to test — it takes your draft and runs it against your original campaign with split traffic. You must create a draft before you can launch an experiment.
Can I test two different landing pages with Google Ads experiments?
Yes, and this is one of the cleanest uses of the tool. Create a draft of your campaign, change the Final URL at the ad or ad group level to your new landing page, and launch the experiment. Both variants will run with identical keywords and bids, isolating landing page performance as the only variable.
What traffic split should I use for a Google Ads experiment?
50/50 is the standard and gives you results fastest. If you’re testing something higher-risk — like a completely new bid strategy on a campaign that’s currently hitting revenue targets — a 70/30 or even 80/20 split in favor of the control limits your downside exposure. The trade-off is it takes longer to reach statistical significance.
How do I know if my experiment results are statistically significant?
Google displays a confidence indicator directly in the Experiments dashboard. Look for 95% confidence or higher before acting on a result. If you’re not at 95% when your planned end date arrives, either extend the experiment by two weeks or call it inconclusive — do not split the difference and declare a “likely winner.”
Can I run experiments on Smart Shopping or Performance Max campaigns?
The Drafts and Experiments tool currently has limited support for Performance Max campaigns. Google has been rolling out native PMax experiments for things like final URL expansion and bidding, but the functionality is less mature than Search campaign experiments. For PMax testing, check what’s available in your account under the Experiments tab directly — it varies by account.
Is Your Agency Actually Testing — Or Just Optimizing on Gut Feel?
There’s a fast way to find out. Ask your agency to show you the last three experiments they ran in your account: what they tested, what the hypothesis was, what the result was, and what they changed as a result.
If they can’t answer that in five minutes, decisions in your account are being made without evidence. That’s not necessarily malicious — plenty of capable practitioners just never built a testing habit. But it means your account’s growth is capped by whoever happens to be right more often than they’re wrong.
A disciplined test-and-learn process — two to three experiments per campaign per quarter, documented and reviewed — is one of the clearest separators between accounts that plateau and accounts that compound. If you want a second set of eyes on your current setup, we offer free account audits that specifically look at experiment history, bid strategy logic, and whether your current structure can support controlled testing. No pitch, just a honest read of what’s working and what isn’t.