Google Ads Experiments: Test Bidding, Ads and Search Campaigns

Learn how our clients are crushing it like never before! Start with a Growth Plan!

   +1 (480) 908-8609  625 W Adams St, Chicago IL 60661 

Google Ads Experiments: How to Test Bidding, Ads and Search Campaigns

“We switched to Target ROAS and revenue went up 20%.”

I’d want to see that account. I’d also ask what else happened that month.

Did the store launch a collection? Did someone raise the budget? Were the best sellers finally back in stock? A better month is good news, but I’d need more evidence before recommending the same bid strategy across the rest of the account.

That’s where I’d use a Google Ads experiment. Run the current setup and the proposed change over the same period, decide what success means in advance, and give both enough traffic to produce a useful result.

The setup is only part of the job. Choosing a question the test can answer is where I’d start.

Key takeaways

  • Use the experiment type that supports the change you want to test. Search custom experiments, ad variations, and dedicated AI Max or Performance Max tests have different workflows.
  • Define the business question, primary metric, spending limits, and evaluation period before launch.
  • Keep purchase goals and measurement consistent between the two arms of a bidding test. Check data volume and campaign eligibility first.
  • Account for ramp-up and conversion delay. Choose the confidence level before looking for a winner.
  • A statistically significant lift can still be too small or too expensive to justify rollout. An inconclusive result also doesn’t establish that the versions perform equally.

What Google Ads experiments can tell you

A custom experiment compares a proposed campaign change with the original setup using an allocated share of traffic and budget. Running them concurrently helps reduce the seasonal and promotional differences that complicate a before-and-after comparison.

It still takes planning. A stockout, broken tag, or promotion available to only one arm can change the question you’re answering. Record disruptions and keep unrelated edits under control.

The result applies to the campaigns, traffic, conditions, and period tested. A successful experiment in a high-volume brand campaign doesn’t establish what will happen in a smaller non-brand campaign.

Choose the experiment type before editing campaigns

For a Search bidding or campaign-setting change, inspect the available custom or guided experiment options. Google’s custom experiment setup lets you select eligible campaigns and a supported variable. Availability and the exact flow depend on the test selected.

For a repeatable text change across responsive search ads, use ad variations. They let you define a campaign scope and test a change such as a call to action without manually rebuilding every ad.

For AI Max, review the dedicated AI Max experiment. Its workflow can divide traffic within an existing campaign. Don’t assume it behaves like a separate copied campaign.

Performance Max has dedicated experiment options, including tests designed for different comparison questions. Choose the relevant design before claiming the result measures additional business sales. This article focuses on Search bidding and messaging tests.

Write the test brief first

I’d write a short brief that another account manager could understand without a handover call.

The question. State the decision you want to make. For example, “Can value-based bidding increase purchase value while meeting our required return?”

The proposed change. Identify exactly what differs. If you change the bidding goal, offer, landing page, and keywords together, you’re testing that package. You won’t know which part produced the result.

The primary metric. Pick the measure that answers the question. For a value-based bidding test, that may be purchase conversion value, with ROAS as the required efficiency condition. For a messaging test, choose the purchase outcome you want to improve and use CTR to help interpret shopper response.

The boundaries. Record the acceptable spend, minimum return, or operational conditions that would justify stopping. A broken checkout needs attention even during an experiment.

The evaluation plan. Decide which dates count, how you’ll handle delayed purchases, the confidence level, and what happens if the result remains uncertain.

I’d save the brief alongside the campaign settings and change log. It stops the team from quietly replacing “more profitable sales” with “higher CTR” when that’s the only number that improves.

Check whether the campaign can support the test

Verify purchase tracking and values

Inspect the purchase action, bidding goals, revenue values, currency, and duplicate-event handling. Both arms need the same measurement basis for a bidding comparison.

If one arm optimizes toward purchases and the other toward a mixture of purchases and newsletter sign-ups, the result won’t answer a straightforward bid-strategy question. Resolve measurement changes before starting.

Our conversion tracking guide covers the measurement work. URL tagging matters too, but a correct UTM doesn’t establish that purchase values are accurate.

Look at volume after splitting

A campaign reporting 20 purchases per month may leave each arm with very little evidence. A six-week end date doesn’t solve that automatically.

Estimate how many purchases each arm could receive at the intended split. Consider the size of improvement the business would need to justify a change. A small effect generally needs more data to distinguish from ordinary variation.

If volume is thin, choose a better-supported campaign, extend a reasonable test period, or simplify the question. Use eligible multi-campaign options only when the campaigns and design suit the same hypothesis. Don’t combine unrelated tests merely to create a bigger total.

Confirm practical eligibility

Open the chosen experiment flow and inspect which campaigns are eligible. Google’s setup guidance documents workflow-specific restrictions and notes that the base campaign must remain active.

Allow time for ad review before the planned start. Check that an automated rule won’t pause the base campaign or end the sale halfway through the test.

How I’d set up a Search experiment

  1. Open Campaigns → Experiments and choose the available custom or guided test matching the question.
  2. Select the eligible base campaign or campaigns and define the treatment.
  3. Choose the success metric, traffic allocation, dates, and any supported split settings.
  4. Review synchronization and automatic application settings before scheduling.
  5. Save the original settings, name the test clearly, and assign an owner to monitor its health.

The setup screen can vary by experiment type. Follow that flow and check the final summary rather than assuming every test requires a manual campaign copy.

For an important bidding test, I’d review the result myself before applying it. Google’s current setup includes automatic application options based on directional results or statistical significance. Inspect which one is enabled. A platform-selected winner still needs to meet the business conditions in your brief.

Cookie-based or search-based split?

Google documents two Search split approaches in its custom experiment guidance.

A cookie-based split keeps a recognized user in one arm. It can help when the shopper returns to search several times before buying. It shouldn’t be described as a guarantee that a person across every browser and device sees only one version.

A search-based split assigns searches between the arms, so the same shopper may be exposed to both. That changes how you interpret the test, particularly when messages or offers could influence later searches.

For a supported ecommerce bidding comparison, I’d usually consider a 50/50 split so neither arm is starved of traffic. Confirm the selected experiment’s requirements. Google also notes a user-list size consideration for cookie-based custom experiments using audience lists.

Choose the split before launch. Google’s experiment FAQs say the split percentage cannot be changed after setup. Changing the design means planning a new test.

How I’d test value-based bidding

Suppose the campaign currently uses Maximize conversions and reliably measures purchase value. The question is whether Maximize conversion value produces a better commercial outcome with the same purchase goal.

Keep the offer, products, destinations, and measurement consistent. If you also change conversion values or primary actions during the experiment, it becomes harder to interpret the bidding result.

Google’s value-based bidding experiment guidance recommends sufficient recent conversion volume, including at least 50 conversions in the prior 30 days for this experiment design. That’s guidance for producing a useful comparison, rather than a universal requirement for using value-based bidding.

Its other recommendations include a 50/50 allocation, comparable targets where applicable, and a campaign that isn’t budget-constrained when testing Target ROAS. Plan two weeks or one to two conversion cycles for ramp-up, whichever is longer, followed by at least 30 days of evaluation. Exclude recent days with materially incomplete conversion reporting.

Enable supported experiment synchronization as recommended for this design, but avoid major edits. Verify which changes actually synchronize. An urgent sitewide correction may still be necessary; document it and decide whether the test remains interpretable.

Choosing a ROAS target

A very demanding new target can restrict volume and produce an unhelpful comparison. Start from historical performance and what the business can afford.

In the brief, distinguish “increase value within the available budget” from “reduce spend while maintaining an efficiency threshold.” Both can be reasonable goals, but the winning result may look different.

If the campaign spends much less after a higher target is introduced, inspect revenue and contribution alongside ROAS. Better efficiency on a much smaller business outcome may or may not meet the plan.

How I’d test ad messaging

Ad variations work well for a specific messaging hypothesis across eligible responsive search ads. Select the scope, define the text change, set the test details, and inspect which ads it affects.

I might test “Find Your Size” against “Shop Now” on a footwear campaign. Check the relevant headlines and descriptions before launch so the replacement makes sense wherever it appears.

“Free Shipping” against “Free 2-Day Shipping” needs more care. If only one message reflects the real service, the test is invalid as advertising. If the service itself differs, you’re testing the offer as well as its wording.

The same applies to a first-order discount. Confirm eligible customers, landing-page terms, code behavior, and margin. A conversion lift can cost more than it earns when the treatment gives away additional contribution.

Responsive search ads can serve combinations of assets. Review the variation as a treatment to the eligible ads rather than claiming that one headline was shown identically on every impression.

A worked ecommerce test decision

Here’s a hypothetical example, separate from Google’s reported benchmarks.

A store tests a bidding change over a completed evaluation period after ramp-up. Both arms use the same purchase action and revenue definition. The control spends $10,000 and records $40,000 in purchase value. The treatment spends $10,500 and records $44,100.

Control ROAS is 4.0×. Treatment ROAS is 4.2×. Purchase value is 10.25% higher, with spend 5% higher.

Those figures tell us what happened in the observed sample. Before rolling out, I’d inspect the experiment’s interval for the chosen primary metric and whether the efficiency condition was met. I’d also check the orders behind the added value.

If both groups had a 40% contribution rate before ad spend, illustrative contribution after advertising would be $6,000 for control and $7,140 for treatment. If the treatment sold a lower-margin mix, the conclusion could change. Use actual economics when judging the result.

A confident platform result is useful evidence. The final decision still needs the commercial target and the limits of the test.

How to read significance and “No clear winner”

Google’s experiment scorecard shows estimated differences and confidence intervals, with 80% as the default confidence level and options for other levels. A blue asterisk indicates statistical significance for the metric.

Choose the confidence level before interpreting results. Moving it after seeing the numbers makes it too easy to select the answer you wanted.

I’d read the estimated effect and its interval together. A small positive point estimate with an interval spanning negative and positive outcomes leaves meaningful uncertainty. “No clear winner” can reflect insufficient evidence; it doesn’t establish that the versions are equivalent.

Likewise, a significant effect can be too small to matter commercially. Check whether the plausible improvement justifies rollout costs and risk.

What I’d do with an inconclusive result

Check traffic, purchases, approvals, and delivery. Confirm that both arms actually ran as planned and that the evaluation dates exclude the periods specified in the test brief.

If the test is healthy and more time could reasonably help, consider an extension within the agreed budget. If volume is too low or the business would only act on a much larger improvement, record what you learned and choose the next step.

I wouldn’t automatically raise the budget until the report produces a winner. More spend needs a business justification.

Monitor the test without repeatedly changing it

During the run, inspect tracking health, serving, disapprovals, stock, and spend against the agreed boundaries. Record incidents such as a payment failure or a high-selling variant going out of stock.

Keep the final result review separate from daily health checks. A few good days don’t replace the evaluation plan. Google’s experiment FAQs also warn that overlapping experiments can interfere with interpretation.

If a serious problem affects one arm differently, assess whether restarting is more defensible than presenting the remaining dates as a clean test.

Apply the result and verify the rollout

Google’s monitoring guidance describes applying an experiment, ending it, or using it to create a new campaign where supported. The available action depends on the experiment type.

Before applying, save the result, evaluation dates, settings, confidence level, and business decision. Confirm what will change and whether any items require manual follow-up.

After rollout, check budgets, bidding targets, ads, goals, tracking, and serving. Don’t leave two campaigns spending unintentionally. Continue monitoring the new full-traffic state because the result is now operating at a different scale.

Should you run an experiment during BFCM?

A holiday experiment can answer a holiday question, provided it’s planned, adequately supported, and operationally manageable. It’s harder to generalize that outcome to ordinary trading.

If the sale lasts four days and the bidding test needs weeks of ramp-up and evaluation, the timing doesn’t fit. I’d test durable bidding changes earlier and use the holiday to execute the agreed plan.

For seasonal messaging tests, decide whether the expected learning is worth the opportunity cost during a high-spend period. Our Google Ads sale strategies cover the wider event planning.

Google Ads experiments FAQ

Is four to six weeks always enough?

Duration depends on the design, volume, conversion cycle, and effect you need to detect. Follow the guidance for the selected experiment. Value-based bidding has a specific ramp-up and evaluation recommendation.

Should I use CTR to pick the winning ad variation?

CTR can explain click response. If the objective is profitable purchases, inspect purchase performance and the economics of the offer before applying the result.

Can I test a new landing page?

Use a supported campaign experiment design and verify the destination change. If the change is made to a shared URL used by both arms, both may receive the new page. Arrange the implementation so the intended difference actually exists.

Does a significant experiment prove incremental store revenue?

It compares the tested Google Ads treatments on the measured outcomes. Establishing additional total-business revenue requires a design and measurement suitable for that incrementality question.

What should I do if neither arm wins?

Review whether the experiment ran correctly and whether the remaining uncertainty matters to the decision. Continue, redesign, or retain the current setup according to the evidence and budget.

Want help choosing what to test?

Bring the current campaign, purchase data, proposed change, and the decision you need to make. If you’d like help designing the experiment and reading the result, talk to Solutions 8 about Google Ads management.

Author

Get even more value!

  • Services
  • About Us
  • Resources
  • Case Studies
  • Contact