Google Ads experiments: test the change before betting the account.

A contractor should use a Google Ads experiment when a proposed campaign change is important, uncertain and testable. Split eligible traffic between the existing setup and one clearly defined alternative, keep other material conditions stable, and judge the result by qualified opportunities, booked jobs and acquisition cost—not clicks or platform conversions alone.

“Let’s test it” often means “we changed it and watched what happened.” That is not a controlled test. Weather changed, competitors moved, demand rose, a dispatcher went on leave and the campaign budget increased at the same time. Performance may improve, but nobody can say why.

Google Ads custom experiments provide a cleaner comparison. The original campaign acts as the control; the experiment carries the proposed change. The two share traffic and budget while results are reported side by side. This does not remove every source of uncertainty, but it is more defensible than replacing a stable setup and comparing this month with last month.

What can a contractor test?

Google says custom experiments can test changes to Search, Display, Demand Gen and Video campaigns. For local lead generation, the most useful Search tests involve a decision that could materially alter reach, bidding or customer response.

The experiment should answer one important question. Testing a new bid strategy, broad match, landing page and advert simultaneously creates one treatment package. If it wins, you still do not know which ingredient earned the result. If it loses, you do not know which good idea was buried by a bad one.

What does a 50/50 traffic split actually mean?

For Search custom experiments, Google offers search-based splitting that randomly assigns a search to the control or experiment. The same person can encounter both arms across different searches. The experiment shares the original campaign’s traffic and budget.

A 50/50 split is a common starting point because both arms have a better chance to collect evidence. But Google explicitly notes that a 50/50 split controls auction eligibility; it does not guarantee equal impressions or spend. Ad Rank, bidding strategy and auction dynamics can still produce different delivery.

That matters when an agency presents a simple cost comparison. Review the actual split, impressions, costs and conversion volume before deciding that an apparent efficiency gain is like-for-like.

Write the experiment before building it

A useful experiment brief can fit on one page. It should state:

• Business question: What decision will this result change?
• Hypothesis: What should improve, for which campaign and why?
• Single treatment: What exact setting, asset or strategy differs?
• Primary outcome: Qualified opportunity, booking, won job or another reliable result.
• Guardrails: Service-area fit, missed-call rate, wrong-job rate, spend and job value.
• Traffic split and dates: How demand and budget will be shared.
• Decision rule: What evidence justifies applying, extending or rejecting the change?

For example: “Adding broad match to the non-brand emergency plumbing campaign will increase qualified local calls without raising cost per booked dispatch beyond the company’s allowable acquisition cost.” That is testable. “Broad match will improve performance” is too vague. Before launching, use the broad-match readiness framework to check conversion goals, bidding, service boundaries and lead feedback.

How long should the experiment run?

Google’s setup guidance says the initial end date for Search custom experiments must be between two and twelve weeks after the start date. Google also recommends allowing seven to fourteen days for the treatment arm to stabilize. Those are platform parameters, not a promise that two weeks produces a trustworthy business decision.

Duration depends on three clocks:

1. Delivery: Both arms need enough relevant auctions, cost and enquiries.
2. Conversion delay: A drain-unblocking call may book immediately; a reroof estimate may take weeks to become won revenue.
3. Operational variation: Include comparable weekdays, weekends, staffing patterns and ordinary demand where possible.

A low-volume contractor may finish with an “undecided” result. That is not failure and it is not permission to choose the arm someone preferred before the test. It means the account did not collect enough evidence to separate the alternatives confidently.

Do not keep extending forever to chase a desired answer. If the required sample is unrealistic, test a larger practical difference, use a dependable upstream result such as qualified opportunities, or label the conclusion as directional.

How should owners read confidence intervals?

Google’s scorecard shows the estimated difference between treatment and control and a confidence interval—a range of plausible differences. It also marks statistically significant results. Google currently lets advertisers select confidence intervals, with 80% shown as the default in its monitoring guidance.

Imagine the treatment shows a 12% conversion-rate improvement, but the range includes both a meaningful decline and a meaningful increase. The headline lift is not a firm result; the data still supports several realities. A narrower interval entirely on the beneficial side is stronger evidence.

Statistical confidence does not establish commercial value. A significant increase in platform conversions can be harmful if they are unqualified calls. A modest lift may be valuable if booked jobs shift toward higher-margin replacement work. Use Google’s report to understand uncertainty, then use CRM and call outcomes to decide whether the difference matters.

Choose a platform metric and a business metric

The cleanest account has a dependable action close to revenue: a qualified lead, booked appointment or won job imported into Google Ads. Many contractors are not there yet. Their campaigns optimize toward calls and forms while the useful result lives in the CRM.

Evaluate two linked layers:

• Platform: Search terms, connected calls, valid forms, cost and conversion rate.
• Business: Qualification, booking, dispatch or estimate, won work, job mix and complete acquisition cost.

Use the same definitions for both arms. A “qualified lead” cannot mean a 30-second call in control and a dispatcher-approved opportunity in treatment. Our primary versus secondary conversion guide explains how weak actions distort automated decisions.

Do not rewrite the test halfway through

Google recommends avoiding changes during the initial learning or stabilization period. Mid-test edits also change the question. If an agency replaces ads, adjusts negatives, moves the target CPA and changes the landing page while the experiment runs, the early and late results no longer represent the same treatment.

Experiment sync is turned on by default for new custom Search and Display trials. Google says it copies optimizations from the original campaign into the experiment. That can keep shared improvements aligned, but it does not remove the need for a change log. Record every edit and decide whether it preserved the test or contaminated it.

Necessary safety fixes are different. Repair a broken phone number, disapproved advert or tracking failure promptly. Document the interruption and consider restarting if it materially affected one arm.

When an experiment should not launch

Do not use an experiment to make a weak campaign look scientific. Repair the foundation first when:

• Page views, button taps or duplicate events are primary leads.
• Calls are frequently missed and no one reviews outcomes.
• Location targeting or search terms already attract obviously wrong demand.
• Budget is too small for both arms to collect evidence.
• The business cannot keep capacity, pricing or availability reasonably stable.
• The treatment bundles too many changes to interpret.

If the proposal comes from Google’s Recommendations page, treat it as a hypothesis—not proof. Our auto-apply recommendations guide shows which changes deserve manual review before testing.

A practical experiment workflow

1. Fix measurement. Confirm phone, form and CRM outcomes and remove duplicate primary actions.
2. Select one consequential question. Make it large enough to matter and narrow enough to interpret.
3. Choose a stable base campaign. Avoid a campaign already undergoing a rebuild or seasonal shock.
4. Create the treatment. Change only the planned variable and check copied settings, exclusions and destinations.
5. Set dates and split. Allow for stabilization, conversion delay and operating cycles.
6. Write decision rules. Define a win, loss, guardrails and an undecided outcome before results arrive.
7. Monitor without meddling. Watch delivery, approvals, tracking and service quality; log unavoidable changes.
8. Wait for mature outcomes. Do not judge a replacement campaign while estimates remain open.
9. Apply deliberately. Google can update the original campaign or convert the experiment into a new campaign. Review the complete setup first.

What should an agency report?

An agency should show the written hypothesis, control, treatment, dates, traffic split, exact differences and success metrics. The closeout should include Google’s result and the contractor’s downstream outcome.

• What single decision is this experiment meant to settle?
• Which conversion actions and CRM stages define success?
• What are the confidence range and actual sample sizes?
• Which changes occurred while the test was live?
• Did qualified rate, booking rate, job mix or close rate change?
• What happens if the result remains undecided?
• Will anything apply automatically when the experiment ends?

Google supports automatic application for some favorable results. Manual review is usually safer for a contractor. “Favorable” inside the platform may not account for crew capacity, service profitability, answer rates or unwanted work.

US, Canada and UK considerations

Run experiments within an operationally coherent market. A treatment tested across several US states, Canadian provinces or UK regions may mix different competition, weather, terminology, licensing and dispatch economics. A result dominated by one metro should not automatically become policy everywhere.

Seasonal trades need additional care. Do not compare a treatment receiving the first heat wave with an earlier control period from mild weather. Simultaneous control and treatment helps, but local shocks, storms and capacity constraints still belong in the experiment record.

Frequently asked questions