In This Article
- 1. What Google Ads Experiments Actually Are
- 2. How the Split Works: Traffic, Budget, and Isolation
- 3. What's Worth Testing (and What Isn't)
- 4. Setting Up an Experiment Step by Step
- 5. Reading the Results: Statistical Confidence and What It Means
- 6. The Most Common Mistakes That Invalidate Experiments
- 7. Applying the Winner and Building a Testing Program
- 8. Frequently Asked Questions
What Google Ads Experiments Actually Are
Google Ads Experiments (found under Campaigns → Experiments in the left nav) let you run a controlled split test between two versions of a campaign — an original and a variant — with traffic divided between them according to a percentage you choose. Both run simultaneously, under the same auction conditions, competing for the same audience. When the test ends, Google shows you which version performed better across the metrics you care about, with a confidence level attached.
The key word is controlled. When you change a bidding strategy directly on a live campaign, you have no baseline to compare against. You're looking at before-and-after numbers that are confounded by seasonality, competitor activity, search volume shifts, and dozens of other variables. An experiment strips most of that out by running both versions at the same time, in the same market, against the same audience.
Direct Change vs. Experiment
Direct Change
- Before/after comparison
- Confounded by time, season, competition
- No statistical confidence score
- Can't reverse cleanly if it hurts
Experiment
- Side-by-side simultaneous test
- Same conditions for both variants
- Confidence level reported
- Apply winner with one click when done
There are two experiment types in the current interface. Custom experiments let you duplicate a campaign, make any changes you want to the copy, and split traffic between them — this is the most flexible type and what most of this article covers. Optimized targeting experiments and Smart Bidding exploration experiments are Google-managed tests with narrower scope. We'll cover when each makes sense.
Experiments are available for Search, Display, and Shopping campaigns. They are not available for Performance Max, Video, or App campaigns — which is worth knowing, because it means if you're running PMax you don't have a first-party controlled testing mechanism at the campaign level.
How the Split Works: Traffic, Budget, and Isolation
When you create an experiment, you choose what percentage of traffic to send to the variant versus the original. The default is 50/50, but you can run asymmetric splits — for example, 80% to the original and 20% to the variant — if you want to protect most of your volume while still gathering test data.
The split is applied at the cookie level. A user who is bucketed into the original sees the original campaign for the duration of the experiment. A user bucketed into the variant sees the variant. This prevents the same user from seeing both, which would contaminate the data.
How Budget Splits Work
Shared budget (default)
The original campaign's budget is shared with the experiment. If you set a 50/50 split, each arm effectively gets half the budget. This is the standard approach for most tests.
Separate budget
You can assign the experiment its own budget. Useful when you want to test without reducing the volume on your original campaign — but costs more overall.
The practical implication of budget sharing: a 50/50 experiment on a $100/day campaign means each arm gets roughly $50/day. Lower-volume accounts will take longer to accumulate statistically significant data. Plan accordingly.
Why Simultaneity Matters
Running both variants at the same time is what gives the experiment its validity. Consider what happens if you don't: you run your original campaign in March, then switch to the variant in April. April happens to have higher search volume because of a seasonal event. The variant looks like a winner. Was it the change, or the season? You can't know.
With an experiment running in parallel, that seasonal lift applies to both arms equally. The only thing that should differ between them is the variable you changed. That's the entire premise — and it's why experiments are fundamentally more reliable than before/after comparisons, even when the before/after window is very short.
The one caveat: very short experiments (under two weeks) can still be affected by day-of-week patterns. A four-day experiment that happens to run Tuesday through Friday will underrepresent weekend behavior. Give your experiments at least two full weeks to capture a representative range of days.
What's Worth Testing (and What Isn't)
The most valuable experiments are the ones that answer a question you'd otherwise have to guess at. A good test has a clear hypothesis, a measurable outcome, and a change large enough that a real difference in performance is detectable. The weakest tests change something trivial — a single word in a single headline — and expect to see a measurable lift across a campaign with modest volume.
High-Value Experiments
Bidding Strategy Changes
Switching from Manual CPC to Target CPA, or from Maximize Clicks to Maximize Conversions, is one of the highest-impact changes you can make — and one of the riskiest to apply directly. An experiment lets you evaluate whether the new strategy actually improves your cost per conversion before committing. This is arguably the single most valuable use of campaign experiments.
Landing Page Variants
Sending all traffic from one campaign to two different landing pages — same ad, different destination — lets you measure conversion rate in a controlled way that standalone A/B testing tools often can't match for paid traffic. The experiment controls the ad quality; the only variable is the page. This is clean, high-signal testing.
Audience Signal Changes
Adding or removing audience layers, testing bid adjustments for specific segments, or evaluating the effect of customer match lists — all of these are well-suited to experiments because the impact is campaign-wide and can be significant. "Observation" mode sounds safe but doesn't give you a controlled test; an experiment does.
Match Type Structure
Testing whether consolidating to broad match (with Smart Bidding) outperforms a tightly controlled phrase/exact match structure is a question that's been contentious in the industry since Google's broad match improvements. An experiment is the only honest way to answer it for your specific account. The data almost always surprises you in one direction or the other.
Ad Copy Overhauls
When you want to test a completely different messaging approach — not just one headline swap, but a wholesale rewrite of the ad set — an experiment is more appropriate than asset-level CTR testing. You're measuring the effect of the entire new creative direction against the existing one, and you want campaign-level conversion data rather than just CTR.
What's Not Worth Experimenting On
Not every change warrants the overhead of a formal experiment. Single-headline swaps, small bid adjustments, and adding one new keyword are better handled through direct changes and monitored via the normal performance dashboard. The effort of setting up and running an experiment pays off when the change is significant enough that getting it wrong would hurt — and when you have enough volume to accumulate meaningful data within a reasonable timeframe.
As a rule: if the potential performance impact is large (bidding strategy, landing page, match type structure), run an experiment. If the potential impact is small and reversible (one new keyword, a minor copy edit), just make the change.
Not sure which changes in your account are high enough impact to warrant an experiment? We can review your campaign structure and identify the best testing opportunities.
Talk to us →Setting Up an Experiment Step by Step
The setup process in the current Google Ads interface is straightforward but has a few configuration decisions that affect the quality of your results. Here's the full process.
Navigate to Experiments
In Google Ads, go to Campaigns in the left nav → Experiments. Click the blue plus button and choose "Custom experiment." Select the original campaign you want to test against.
Name and schedule the experiment
Give it a descriptive name that records what's being tested (e.g. "Target CPA vs Manual CPC — June 2026"). Set a start date (usually today or tomorrow) and an end date. Plan for at least 2–4 weeks; longer if your campaign is lower volume. You can end early if results reach significance sooner.
Set the traffic split
Choose cookie-based split (the default and correct option for most tests) and set your percentage. 50/50 generates data fastest. Use 80/20 only if you genuinely can't afford to reduce volume on the original — it will take proportionally longer to reach significance on the variant arm.
Make your change in the variant
Once the experiment is created, Google creates a draft copy of the campaign. Navigate to it and make exactly the change you're testing — and only that change. Change the bidding strategy, swap the landing page URL, update the ads, modify the audience layer. One variable. Do not add extra changes "while you're in there."
Set your primary metric
In the experiment settings, choose the metric you care most about — cost per conversion, conversion rate, ROAS, CTR. This is what Google will use to declare a winner and report confidence levels. Choose the metric that directly reflects the business goal you're trying to improve.
Launch and leave it alone
Start the experiment and resist the urge to make changes to either arm during the test. Any mid-experiment change introduces a new variable and contaminates your results. Check in weekly to see if results are trending toward significance, but don't act until the planned end date (or until confidence reaches your threshold).
Reading the Results: Statistical Confidence and What It Means
When you open an active or completed experiment in Google Ads, you'll see a results panel showing performance metrics for both the original and variant arms, and a confidence level expressed as a percentage. This is the part most advertisers misread — either overconfidently acting on 70% confidence or dismissing a 95% result as "just statistics."
How to Interpret Confidence Levels
| Confidence Level | What It Means | What to Do |
|---|---|---|
| 95%+ | Strong evidence of a real difference. Only a 5% chance the result is due to random variation. | Apply the winner confidently. |
| 80–94% | Moderate evidence. The trend is real but not conclusive. May be underpowered — needs more data. | Extend the experiment if possible, or apply with caution and monitor closely. |
| 60–79% | Weak evidence. The result could easily be noise. Don't act on directional trends at this level. | Extend significantly or accept the null result and move on. |
| Below 60% | No meaningful signal. The change had no detectable effect, or the test was too small to detect one. | End the experiment. Stick with original or redesign the test. |
The Difference Between Statistical and Practical Significance
A result can be statistically significant without being practically meaningful. If your experiment runs for eight weeks, accumulates tens of thousands of impressions, and shows that the variant has a 0.4% lower cost per conversion at 96% confidence — that's a statistically real result, but the size of the improvement may not justify the operational overhead of making the change.
Conversely, an experiment with limited volume might show a 25% improvement in cost per conversion at only 78% confidence. The magnitude of the signal is large; the certainty is moderate. In that case, the right move is often to extend the experiment rather than act on it — or to apply the change and flag it for close monitoring, accepting that you're making a probability-weighted decision rather than a certainty.
The framework: read statistical confidence as the probability the result is real, and the magnitude of the difference as the reason to care. Both need to clear a threshold before you act.
Don't end experiments early just because you like what you see
Early experiment results are almost always noisy. The first week of a 4-week experiment often shows exaggerated swings in both directions as Google's algorithm learns the new bidding or targeting configuration. Ending on day 10 because the variant looks great — or terrible — is a common mistake. Give the experiment the full duration you planned unless confidence clearly exceeds 95% and the magnitude is large enough to act on.
The Most Common Mistakes That Invalidate Experiments
A poorly run experiment produces data you can't trust, which is arguably worse than no data at all — because it creates false confidence. These are the mistakes most likely to contaminate results.
Testing more than one variable
If you change the bidding strategy and the landing page in the variant, you won't know which change drove the result — or if they interacted in unexpected ways. One variable per experiment, always. The temptation to bundle changes is strong ("we're already running a test, let's throw in the new landing page too") but it makes the data uninterpretable.
Making changes to the original campaign mid-experiment
If you add a new keyword, adjust bids, or change ad copy on the original campaign while the experiment is running, you've introduced a new variable into the baseline. The comparison is now between "original + change" and "variant" — not what you set up. Treat the original as frozen for the duration of the test.
Running the experiment over a non-representative period
If your experiment runs entirely over a holiday shopping period, a product launch, or a major news event that affected search behavior in your category, the results may not generalize to normal operating conditions. Be aware of what's happening in your market during the test window and flag any major external events in your notes.
Underpowering the test with too small a traffic split
An 80/20 split on a campaign that converts 15 times per month means the variant arm is seeing roughly 3 conversions per month. You cannot draw any conclusions from 3 conversions, let alone statistically significant ones. If your volume is limited, either accept that tests will take longer or reserve experiments for changes that are likely to have large enough effects to detect with limited data.
Treating Smart Bidding learning periods as experiment noise
When you test a new Smart Bidding strategy (Target CPA, tROAS, Maximize Conversions), the algorithm needs a learning period — typically 1–2 weeks — to calibrate. During this period, performance is often worse than steady state. If you read results before the learning period ends, the variant will look worse than it actually is. Factor in at least 2 weeks of learning time when planning the experiment duration.
Applying the Winner and Building a Testing Program
When your experiment reaches a confident result, Google gives you two options: apply the experiment (replace the original with the variant) or end the experiment and keep the original. Applying is a single click — Google handles the swap without disrupting campaign history or resetting the algorithm's learning where possible.
The more important question is what happens next. A single experiment is useful. A systematic program of experiments — one running at all times, each building on the results of the last — is what actually moves the needle on account performance over time.
1
Test one thing
Bidding strategy, landing page, match type, audience layer, or ad copy direction.
2
Read the result
Apply the winner at 95%+ confidence. Extend or discard below that threshold.
3
Queue the next test
Each result reveals the next question worth asking. Build a backlog and work through it.
Maintaining a Testing Backlog
The best testing programs keep a simple log: what was tested, what the hypothesis was, what the result was, and what the next question is. Over six months, this produces a genuine institutional understanding of what drives performance in your account — not instinct or industry convention, but empirical evidence from your own campaigns, with your specific audience, in your specific market.
Most of the accounts we audit have never run a formal experiment. They've made dozens of changes over the years — bidding strategy overhauls, landing page redesigns, match type consolidations — with no way to evaluate which changes actually helped. The accounts that run systematic experiments make better decisions faster, waste less budget on changes that don't work, and compound improvements quarter over quarter.
Want a managed testing program for your Google Ads?
We build structured experiment programs into all of our Google Ads engagements — one active test at all times, monthly reviews, and results documented so you always know what changed and why. At 8% of ad spend, no long-term contracts.
8%
of Ad Spend — Flat Fee
Always On
Experiment Running
Monthly
Results Review
No
Long-Term Contracts
Frequently Asked Questions
How long should a Google Ads experiment run?
At minimum, two full weeks — to capture day-of-week variation and account for the Smart Bidding learning period if you're testing a new bidding strategy. For lower-volume campaigns, four to six weeks is more appropriate. The right duration depends on how quickly you accumulate conversions (or whatever metric you're testing). If you have 100+ conversions per month per arm, two weeks is usually enough. If you have 20 conversions per month per arm, you'll need closer to eight weeks. Don't end early just because results look promising — early data is almost always noisier than the final picture.
Can I run multiple experiments at the same time across different campaigns?
Yes. Each experiment is tied to a specific campaign and is independent of experiments on other campaigns. Running a bidding strategy experiment on your top Search campaign while running a landing page experiment on your brand campaign is perfectly valid — they don't interfere with each other. What you shouldn't do is run two experiments on the same campaign simultaneously, as the interactions between variants would make results uninterpretable.
Does running an experiment hurt campaign performance during the test?
Potentially, yes — especially if the variant turns out to perform worse than the original. A 50/50 experiment means half your traffic is going to the thing you're testing, and if the test fails, that half underperformed for the duration. This is the fundamental cost of running experiments, and it's why you should focus experiments on changes where the potential upside is large enough to justify it. It's also why the experiment cadence matters: a well-designed experiment running for four weeks costs you less than a bad campaign change applied directly and not reversed for three months.
What's the difference between a Google Ads Experiment and a Google Optimize test?
Google Optimize was deprecated in September 2023. For landing page testing specifically, you'll need to use an alternative (VWO, AB Tasty, Optimizely, or a custom implementation). A Google Ads Experiment testing landing pages works by sending different traffic segments to two different URLs — the test controls the traffic split, but both pages need to exist separately. This is less flexible than a proper A/B testing platform that can serve variants of the same URL, but it gives you cleaner Google Ads conversion data and doesn't require any third-party tool on the landing page.
Can I use experiments to test Performance Max against a standard Search campaign?
Not directly — Google Ads Experiments only work within the same campaign type, and PMax can't be used as a campaign experiment variant. However, you can run a side-by-side comparison manually: run PMax and a standard Search campaign simultaneously targeting the same audience and measure performance over the same period. It's not a controlled experiment in the strict sense (no cookie-level split, no automatic confidence scoring), but it's the closest available proxy for evaluating PMax versus Search for your specific account. Google has been pushing PMax hard, but the data on whether it outperforms well-managed Search campaigns for service businesses is genuinely mixed.
What should I test first if I've never run a Google Ads experiment before?
If you're currently on Manual CPC and you have conversion tracking set up, test switching to Maximize Conversions or Target CPA. This is the highest-impact change most accounts can make, and it's also the one with the most risk if applied directly without a test — Smart Bidding can either dramatically improve efficiency or temporarily crater performance during the learning period. Running it as an experiment lets you evaluate whether it's the right move for your account before committing. If you're already on Smart Bidding, test a landing page variant next — the conversion rate improvement potential is usually larger than any ad copy change.
If you're making campaign changes without controlled experiments, you're flying partially blind. Learn how we manage Google Ads with a structured testing program built in, or request a free audit of your current campaigns.