Google Ads Experiments: Running Valid A/B Tests (2026 Guide)

"I think we should switch to Target CPA bidding."

"I feel like Broad Match might work better now."

"The new landing page looks stronger."

"The old headline feels tired."

These are common conversations in Google Ads.

They sound harmless.

But opinions are expensive.

A wrong change can double CPA.

A rushed test can kill a strong campaign.

A landing page swap can reduce conversion rate.

A bid strategy change can reset learning.

A broad match test can open the door to poor search terms.

A new ad can get more clicks but worse leads.

That is why serious advertisers do not guess.

They test.

Google Ads Experiments allow you to test meaningful changes without gambling the whole account.

Instead of changing a live campaign and hoping for the best, you split traffic between a control and a variant.

One campaign keeps the current setup.

The experiment tests the new idea.

Then you compare the results.

This is how good decisions are made.

Not by opinion.

Not by pressure.

Not by platform recommendations alone.

By evidence.

In this "Mega-Authority" guide, we cover:

1. The Methodology: Science applied to marketing.
2. The Setup: Creating a Cookie-Split test.
3. What to Test: Bidding, Match Types, Creative.
4. Interpreting Results: Statistical Significance and confidence.

The goal is simple.

Protect the downside.

Measure the upside.

Scale what proves itself.

---

Part 1: The Financial Impact of Testing

Imagine you switch your main campaign from Manual CPC to Maximize Conversions.

It performs badly.

CPA doubles.

You lose $5,000 in a week.

The sales team complains.

The client loses confidence.

You panic and switch back.

Now imagine you tested it on 50% of traffic.

It performs badly.

You lose less.

You stop the test.

The original campaign keeps running.

You learn without breaking the whole account.

Or the experiment succeeds.

CPA drops by 30%.

Conversion volume holds.

Lead quality stays strong.

Now you can roll it out with confidence.

That is the value of experiments.

They give you a controlled way to make bigger changes.

Without full-account risk.

Experiments are your sandbox.

They protect your downside while unlocking upside.

This matters because many Google Ads changes are not small.

Changing one headline is small.

Changing bidding strategy is not.

Changing match type is not.

Changing landing pages is not.

Changing final URLs across a campaign is not.

Changing conversion goals is not.

Changing budget strategy is not.

These changes can affect the entire learning system.

They can affect traffic quality.

They can affect search terms.

They can affect conversion rate.

They can affect cost per lead.

They can affect sales quality.

An experiment lets you ask a better question:

"Does this change improve performance when tested fairly against the current setup?"

That is a much stronger question than:

"Do we like this idea?"

The financial impact is simple.

A good test can find growth.

A bad test can prevent damage.

Both are valuable.

If a test loses, it still helped.

It stopped you from rolling out a bad idea.

That is not failure.

That is risk control.

---

Part 2: Theory - Cookie Split vs Search Split

A valid experiment needs a fair split.

Google Ads experiments can use different split methods depending on setup and availability.

Two important concepts are cookie-based split and search-based split.

Cookie-Based Split

With a cookie-based split:

• User A falls into the "Control" bucket. They consistently see the original campaign.
• User B falls into the "Experiment" bucket. They consistently see the test campaign.

This helps data integrity.

It prevents the same user from seeing Ad A in the morning and Ad B in the evening.

That matters because user behaviour is not independent.

A person may click one ad, think about it, return later and convert.

If they see both variants during the journey, attribution becomes harder to read.

Cookie-based splitting is usually preferred when you want cleaner user-level comparison.

It is especially useful for:

1. Landing page tests.
2. Ad copy tests.
3. Bid strategy tests.
4. Audience tests.
5. Longer consideration journeys.

Search-Based Split

With a search-based split, each eligible search can be assigned to either the control or the experiment.

This can gather data faster.

But the same user may be exposed to both variants across different searches.

That can make the test less clean in some cases.

It may still be useful where cookie-based testing is not ideal or not available.

The key point is this:

Do not ignore split method.

It affects the quality of your test.

A good experiment is not just about what you test.

It is also about how you split the traffic.

For most important tests, use the cleanest split available.

Then run the test long enough for meaningful data.

---

Part 3: Framework - The Testing Hierarchy

Do not test random things.

Test high-impact levers.

Many advertisers waste time testing tiny changes while major issues remain untouched.

They test one description line while the bidding strategy is wrong.

They test a CTA while the landing page is weak.

They test a headline while the match type is pulling junk traffic.

That is backwards.

Start with the changes that can materially affect performance.

This does not mean ad copy is unimportant.

It is very important.

But if the campaign is using the wrong conversion goal, ad copy will not fix it.

If the landing page is poor, ad copy will only send more users into a weak experience.

If the match type is uncontrolled, better headlines may attract more poor-fit clicks.

So test in the right order.

A mature testing roadmap might look like this:

1. Conversion tracking audit.
2. Bid strategy test.
3. Match type test.
4. Landing page test.
5. Ad copy test.
6. Audience observation or targeting test.
7. Budget scaling test.
8. Value-based bidding test.
9. New customer acquisition test.
10. Offer test.

You should also test based on the business problem.

If CPA is too high, test bidding, match types, landing page or value proposition.

If volume is too low, test broader match, higher targets, new keywords or new landing pages.

If CTR is weak, test ad copy and assets.

If lead quality is poor, test qualification, landing pages, conversion goals or offline conversion imports.

Do not test because you are bored.

Test because there is a decision to make.

---

Part 4: Execution - Setting Up a Test

Let's test Manual CPC vs Target CPA.

This is one of the most useful experiments for a campaign that has enough conversion data.

The question is:

"Can Target CPA improve efficiency or volume compared with our current manual bidding setup?"

1. Campaigns → Experiments → All Experiments.
2. Click + -> Custom Experiment.
3. Base Campaign: Select your current "Search - Generic".
4. Suffix: - Experiment - tCPA.
5. Configuration:
   • Change Bidding Strategy to Target CPA.
   • Set Target CPA using your historical 30-day average.
6. Split: 50% where volume allows.
7. Schedule: Start Date (Tomorrow). End Date (None - manually end it).

Setting the Target CPA matters.

Do not set it too low.

If your historical CPA is £80, do not set the experiment target at £25.

That is not a fair test.

You are not testing Smart Bidding.

You are starving it.

Start close to recent performance.

Then let the experiment show whether the strategy can improve.

A fair setup should keep everything else the same.

Same keywords.

Same ads.

Same landing page.

Same location.

Same schedule.

Same conversion action.

Same budget split.

Only the bidding strategy changes.

That is how you isolate the effect.

Before launch, check:

1. Is conversion tracking clean?
2. Does the campaign have enough conversions?
3. Is the base campaign stable?
4. Are there major seasonal events coming?
5. Is the budget large enough for a 50/50 split?
6. Are there enough clicks and conversions to measure?
7. Is the test start date sensible?
8. Is there a clear success metric?

A weak experiment setup creates weak conclusions.

Take the setup seriously.

---

Part 5: The "Don't Touch" Rule

Once an experiment is live, DO NOT TOUCH IT unless there is a serious issue.

Do not change the budget.

Do not add keywords.

Do not change ads.

Do not change the landing page.

Do not change conversion settings.

Do not add new locations.

Do not change the target after three bad days.

If you change variables mid-test, you invalidate the results.

This is one of the most common mistakes.

People launch a test.

Then they get nervous.

They start editing.

Then they do not know what caused the result.

Was it the bid strategy?

The new ad?

The added negative keywords?

The landing page tweak?

The budget change?

Nobody knows.

The test is now muddy.

Patience is the skill here.

The "Learning Phase" may take several days.

The "Data Collection" phase often takes 14-30 days or longer.

For lower-volume accounts, a valid test may need more time.

For high-volume accounts, you may get directional learning faster.

But do not judge too early.

A good experiment should run long enough to cover:

1. Weekdays and weekends.
2. Normal business cycles.
3. Conversion lag.
4. Learning period.
5. Sufficient clicks.
6. Sufficient conversions.
7. Sales quality review where relevant.

If your sales cycle is 14 days, do not judge the test after 7 days.

The conversions have not finished arriving.

If your leads take 30 days to close, platform CPA alone is not enough.

You need CRM data.

The more valuable the decision, the more patient the test should be.

---

Part 6: Interpreting Results

After enough data has collected, check the experiment dashboard.

Google may show confidence indicators or performance comparisons depending on experiment type and available data.

Do not look only at the metric that flatters the experiment.

Decide the success metric before the test starts.

Common success metrics include:

1. Cost per conversion.
2. Conversion rate.
3. Conversion volume.
4. Conversion value.
5. ROAS.
6. Revenue.
7. Lead quality.
8. Qualified lead rate.
9. Cost per qualified lead.
10. Profit per impression.

• Metric: Conv. / Cost (ROAS) or Cost / Conv. (CPA).
• Result: "Experiment outperformed Base by +15% with strong confidence." -> Consider applying.
• Result: "No significant difference." -> End or keep the simpler setup.
• Result: "Experiment underperformed." -> End. The test protected the account.

Be careful with statistical significance.

A test can look better by chance.

A test can look worse during learning, then improve.

A test can produce cheaper leads but worse customers.

That is why you need both platform data and business data.

For lead generation, do not apply a test only because CPA dropped.

Ask:

1. Did lead quality hold?
2. Did contact rate hold?
3. Did qualified leads increase?
4. Did sales accepted leads improve?
5. Did close rate change?
6. Did search terms remain relevant?
7. Did the campaign attract weaker traffic?

For ecommerce, ask:

1. Did revenue increase?
2. Did ROAS improve?
3. Did average order value change?
4. Did margin improve?
5. Did new customer acquisition improve?
6. Did return rate change?

A good test winner should improve the business.

Not just the ad platform.

How to Apply:

Click "Apply Experiment." You can choose to:

1. Update Original: Converts the base campaign to the new settings. This is usually preferred when the experiment proves a better setup.
2. Convert to New: Creates a new campaign from the experiment. Use this only when you have a reason to keep the original separate.

In many cases, Update Original is cleaner.

It preserves the base campaign structure and avoids unnecessary duplication.

But if the experiment created a genuinely different campaign type or strategy, converting to a new campaign may make sense.

Choose based on account structure.

Not habit.

---

Part 7: Summary & Checklist

If you aren't testing, you aren't growing.

But if you test badly, you are just guessing with extra steps.

A proper experiment has:

1. A clear hypothesis.
2. One main variable.
3. A fair split.
4. Enough time.
5. Enough data.
6. A defined success metric.
7. A calm decision process.

Your Action Plan:

1. Identify a campaign that has plateaued.
2. Hypothesize a change, such as "Broad Match with Smart Bidding will get more qualified volume."
3. Launch a 50/50 experiment where volume allows.
4. Wait 4 weeks, or long enough to collect meaningful data.

Be the scientist.

Here is the deeper checklist:

1. Choose one business problem.
2. Write one clear hypothesis.
3. Select one variable to test.
4. Choose the right campaign.
5. Check conversion tracking before launch.
6. Set the right traffic split.
7. Use cookie-based split where suitable.
8. Set a realistic test duration.
9. Avoid major edits during the test.
10. Allow for conversion lag.
11. Review confidence and significance.
12. Check lead or sales quality.
13. Apply winners carefully.
14. End losers without regret.
15. Document what you learned.

A losing experiment is not a failure.

It is an expensive mistake avoided.

---

The "Big Three" Growth Tests

Most accounts only need to run three types of experiments regularly.

These are the tests that usually move the account forward.

1. "Brain Transplant" — Manual CPC → tCPA

This is one of the biggest possible efficiency tests.

You are testing whether Smart Bidding can outperform manual control.

Set experiment to 50/50 where possible.

Run for at least 30 days or until you have enough conversions in each arm.

Use this when:

1. The campaign has stable conversion tracking.
2. There is enough conversion volume.
3. CPA is plateauing.
4. Manual management is becoming inefficient.
5. You want more auction-time optimisation.

Do not use it when tracking is weak.

Smart Bidding will only optimise towards the signal you give it.

2. "Net Cast" — Phrase Match → Broad Match

This tests whether broad match can find more valuable volume.

Broad match is better than it used to be.

But it still needs controls.

Use it with:

1. Smart Bidding.
2. Strong negatives.
3. Clean conversion tracking.
4. Search term reviews.
5. Budget limits.
6. Lead quality checks.

Do not assume broad match fails.

Do not assume it works.

Test it.

The best setup is often to test broad match on top-performing themes, not the whole account at once.

3. "Landing Page Duel" — Page A vs Page B

This is one of the cleanest commercial tests.

In experiment settings, use Find and Replace in the Final URL field to swap Page A for Page B across selected ads.

One change.

Clean data.

Clear question.

This is useful when:

1. You have a new landing page.
2. The current page has weak conversion rate.
3. You want to test short vs long page.
4. You want to test form layout.
5. You want to test proof, pricing or offer structure.
6. You want to test speed or mobile improvements.

Do not judge only by conversion rate.

Check lead quality.

A shorter form may increase leads but reduce quality.

A longer page may reduce leads but improve close rate.

The business result decides the winner.

Naming Convention That Keeps You Sane

Every experiment must follow a naming convention.

Use:

EXP - [Date] - [Hypothesis]

Examples:

• EXP - Jan 2026 - Test tCPA vs Manual
• EXP - Feb 2026 - Broad Match on Core Services
• EXP - Mar 2026 - Landing Page B vs A
• EXP - Apr 2026 - RSA Proof Hook Test

Without this convention, 6 months later you will not know what "Experiment 3" was testing.

Also document:

1. Start date.
2. End date.
3. Campaign tested.
4. Hypothesis.
5. Variable changed.
6. Success metric.
7. Result.
8. Decision.
9. Notes.
10. Follow-up action.

This turns testing into institutional knowledge.

Not a memory game.

The "Testing Too Much" Failure Mode

The single most common experiment mistake is changing multiple variables at once.

You change the bid strategy.

And the headlines.

And the landing page.

And the match type.

CPA improves 30%.

Great.

But why?

Was it the bid strategy?

The headlines?

The landing page?

The match type?

You will never know.

One experiment = one main variable.

If you want to test three things, run three consecutive experiments.

This feels slower.

It is actually faster in the long run because the learning is clean.

Bad testing creates confusion.

Clean testing creates reusable knowledge.

"Apply" vs "Convert" — End-Game Decision

When an experiment wins, you have two options:

• Apply to Base Campaign — Merges the winning variant into the original. Recommended for many bid strategy and match type tests.
• Convert to New Campaign — Creates a brand-new campaign from the winner. The original stays unchanged. Use this when you want to maintain the original as a separate control or when the experiment deserves its own budget and structure.

Default recommendation:

Apply to Base Campaign unless you have a specific reason to preserve the original.

But do not apply blindly.

Before applying, check:

1. Did the test have enough data?
2. Did it run long enough?
3. Did it cover normal business cycles?
4. Did lead quality hold?
5. Did the result match the hypothesis?
6. Are there any external factors?
7. Will applying disrupt other campaigns?
8. Does the business agree with the trade-off?

The final rule is simple.

Experiments do not remove judgement.

They improve it.

They give you evidence.

You still need to make the decision.