Google Ads Experiments: A/B Testing Your Campaigns

  • Google Ads

A badly designed experiment is worse than no test, because it sends you confidently the wrong way. How to change one variable, size the test, wait long enough and read the result without fooling yourself.

Last updated: November 10, 2026

In This Guide

  1. Executive Summary
  2. Benchmarks & Numbers at a Glance
  3. Why Test
  4. Experiments, Drafts & A/B Tests
  5. Designing a Valid Experiment
  6. Sample Size & Significance
  7. What to Test
  8. Reading Results
  9. Applying or Rolling Back
  10. Testing with Automation
  11. Common Mistakes to Avoid
  12. What Changed Recently
  13. References

1. Executive Summary

Google Ads experimentation has matured significantly by November 2026. The platform now consolidates campaign experiments, ad variations, AI Max tests, Performance Max asset experiments, and incrementality lift studies inside a unified Experiment Center. Five principles define best practice:

  • One variable, one hypothesis. Every experiment must isolate a single material change—bidding strategy, landing page, ad copy, or audience—against a stable control. Testing multiple variables simultaneously produces a result that cannot be attributed to any one cause.[1]
  • Volume before launch. A campaign needs sufficient conversion history before an experiment can produce a credible result. Google’s vendor guidance points to at least 100 conversions per variation as a working minimum, with 300 or more preferred.[2] Run experiments only on campaigns that meet this bar.
  • Duration beats speed. The default minimum is four weeks for Search experiments and four to six weeks for Shopping, with additional time when conversion delay is long or automated bidding needs to recalibrate.[2][3] Stopping early because a result first crosses a significance threshold is a recognized source of false positives.
  • Statistical significance is necessary but not sufficient. Google identifies 95% confidence (p ≤ 0.05) as the conclusive threshold.[2] A result that clears this bar but delivers a commercially trivial lift—or harms a guardrail metric such as lead quality or gross margin—should not be promoted.
  • Apply winners deliberately, not automatically. Before applying an experiment, verify that conversion quality data is complete, that the winning condition still matches current market conditions, and that the change is appropriate for the full campaign budget and target audience.

2. Benchmarks and Numbers at a Glance

Metric Typical range or threshold Applies when Source
Recommended traffic and budget split 50% control / 50% treatment Vendor claim; default for most Search and custom experiments where risk is acceptable [4]
Minimum experiment duration (Search) 4 weeks (30 days) Vendor claim; applies to Search custom experiments; assumes adequate weekly conversion volume [2]
Recommended experiment duration (Shopping) 4–6 weeks, or 1–2 conversion cycles—whichever is longer Vendor claim; extends further when conversion delay exceeds 14 days [3]
Conclusive significance threshold 95% confidence (p ≤ 0.05, two-tailed) Vendor claim; Google’s academic standard for declaring a winner [2][5]
Directional significance threshold (lower bound) 70% confidence Vendor claim; acceptable only for fast directional reads on low-volume or low-budget accounts; label results as directional, not conclusive [5]
Directional significance threshold (upper bound) 80% confidence Vendor claim; for app-uplift experiments and low-volume campaigns where 95% is unattainable in a reasonable timeframe [5][6]
Minimum conversions per variation (working floor) 100 conversions Third-party vendor claim; minimum below which results are highly unreliable regardless of duration [2]
Preferred conversions per variation 300+ conversions Third-party vendor claim; preferred for detecting effects smaller than 10–15% with 80%+ power [2]
Minimum audience list size for cookie-based custom experiments 10,000 users Vendor claim; smaller lists produce less accurate audience-based splits in Search experiments [4]
Ramp-up exclusion for automated bidding experiments 7–14 days from experiment start Vendor claim; first 7 days excluded for general ramp-up; first 1–2 weeks excluded when Smart Bidding or a new feature needs recalibration [2][3]
Target proximity for Performance Max bidding experiments Within ±15% of historical actual performance Vendor claim; targets set more than 15% away from historical actuals increase learning instability and reduce result reliability [3]
Illustrative conversion-volume multiplier between two video ad creatives Up to 8× difference Vendor claim; illustrates potential creative impact; not a generalizable benchmark across ad types or industries [5]
Advance scheduling lead time before experiment launch 3–7 days Vendor claim; allows ad review and policy approval to complete before measurement begins [2]
Minimum recent conversion history for bidding experiments ~30 conversions in prior 30 days Vendor claim; Google’s recommended minimum for the base campaign before testing a Smart Bidding strategy [1]

3. Why Test in Google Ads

Google Ads auctions are non-stationary: Quality Scores shift, competitor budgets change, Smart Bidding models update, and user behavior evolves across devices and seasons. An account that ran the same settings for twelve months without testing is almost certainly leaving measurable efficiency or revenue on the table. Controlled experiments answer the question that optimization reports, automated recommendations, and historical trend lines cannot: did this specific change cause this specific outcome?[7]

Three business pressures make systematic testing especially valuable in the United States market as of 2026. First, Smart Bidding and AI Max handle micro-decisions automatically, which means the highest-leverage remaining decisions are strategic: bid strategy selection, campaign structure, audience targeting philosophy, and landing-page experience. These are exactly the decisions that campaign experiments are designed to validate. Second, rising cost per click across most US verticals means that a 10–15% improvement in conversion rate or cost efficiency has a larger absolute dollar value than it did several years ago. Third, Google’s shift toward Privacy Sandbox and reduced third-party signal availability has increased conversion uncertainty, making a direct control-versus-treatment comparison more valuable than relying on modeled attribution alone.[1][2]

Experimentation also protects against the most common source of wasted spend: applying account-wide recommendations without evidence. When Google recommends switching from Maximize Conversions to target CPA, or enabling broad match, or expanding to Performance Max, running that recommendation as an experiment first allows the advertiser to measure actual business impact before making an irreversible change.[8][9]

Worked example (illustrative)

Validating a Smart Bidding recommendation before account-wide application

  • Setup: A Chicago B2B software account spends $18,000 per month on Search. Google Ads shows a recommendation to switch from Manual CPC to Maximize Conversions on the account’s top campaign, which generates 85 conversions per month at an average CPA of $212.
  • Numbers: 85 conversions ÷ 2 arms = 42–43 conversions per arm per month. At a 30-day minimum run[2] and 50/50 split[4], both arms fall below the preferred 300-conversion threshold[2] but are above the 100-conversion working floor. The team schedules a 60-day experiment (June 1–July 31, 2027) to accumulate approximately 85 conversions per arm. Budget allocated: $9,000 per arm per month × 2 months = $36,000 total.
  • Decision: Run a custom Search experiment with a 50/50 cookie-based split, Maximize Conversions in the treatment arm and Manual CPC at current bids in the control arm, for exactly 60 days. Primary metric: cost per qualified lead confirmed in CRM. Significance threshold: 95% confidence (p ≤ 0.05).[2]
  • Why: Google’s vendor guidance identifies p ≤ 0.05 as the conclusive threshold[2], and the campaign needs 60 days to accumulate sufficient conversions in each arm because monthly volume (85) is below the preferred 300-conversion bar.[2]

4. Experiments, Drafts and A/B Tests Explained

Google Ads in 2026 offers four distinct mechanisms for structured testing. Understanding which tool to use—and when—prevents the most common setup errors.[1][10]

Campaign Experiments (Custom Experiments)

A campaign experiment creates a treatment copy of an existing campaign and runs it alongside the original, splitting eligible traffic and budget according to a predefined ratio. The control campaign continues to run unchanged. This is the correct tool for testing bidding strategies, match types, audience settings, landing pages, and campaign-level structural changes. Custom experiments support Search, Display, Demand Gen, and Video campaigns as of November 2026. App and standard Shopping campaigns use separate experiment pathways.[1][11]

Eligibility requirements for custom experiments include: the base campaign must be active, must use an individual daily budget rather than a shared budget, and must use a standard bidding strategy rather than a portfolio bid strategy.[11]

Drafts

A draft is a working copy of a campaign used to stage proposed changes before applying or testing them. It is not a controlled test on its own. A draft has no treatment-versus-control measurement until it is converted into an experiment. Use drafts to review and QA proposed changes—ads, URLs, tracking templates, bid adjustments, audience exclusions—without disturbing the live campaign. Apply a draft directly only for operational corrections or approved expansions where causal evidence is not needed.[1][12]

Draft creation through Google Ads Scripts is restricted to Search, Search Network with Display Expansion, and Display campaigns. The base campaign cannot use a shared budget.[13]

Ad Variations

Ad variations allow advertisers to test a single text or responsive-search-ad change consistently across multiple campaigns simultaneously. This tool is appropriate when the hypothesis is narrow—such as testing one headline across all brand campaigns—and when campaign-level structural variables are not changing. Ad variations report performance differences and allow the modified ads to be applied after the test.[1]

AI Max Experiments

AI Max experiments are a one-click test type introduced for Search campaigns. They use a 50/50 split of the existing campaign to compare current settings against AI Max features (search-term matching expansion, asset optimization) without requiring the advertiser to create a separate copied campaign. This reduces setup complexity and preserves a direct control comparison.[4][11]

Performance Max Asset Experiments

As of November 2026, Google has announced the ability to test creative assets within Performance Max campaigns, comparing new assets against existing assets or testing additional creative types. Availability is rolling out progressively and may not yet be enabled in all accounts.[8]

Experiment Center

The Experiment Center is the unified hub that consolidates campaign experiments, ad variations, lift studies, and incrementality measurement into one reporting interface. It supports testing campaign settings, assets, and the impact of adding campaign types. Up to two primary success metrics can be designated per custom experiment, with a summary comparison view between original and experiment campaigns.[10][11]

Tool Best use case Supported campaign types (Nov 2026) Traffic split method
Custom experiment Bidding, match types, landing pages, audience settings, campaign structure Search, Display, Demand Gen, Video Cookie-based or search-based[4]
Draft (operational) Staging approved changes; QA before applying Search, Search with Display Expansion, Display N/A—not a controlled test
Ad variations Consistent ad-copy change across multiple Search campaigns Search (RSA and text ads) Google-managed rotation
AI Max experiment Testing AI Max features on an existing Search campaign Search 50/50 of existing campaign[4]
Performance Max asset experiment Creative asset comparison within PMax Performance Max (rolling availability)[8] Google-managed within PMax
Experiment Center / Lift Study Incrementality measurement; cross-campaign validation Multiple (depends on account eligibility)[10] Holdout-based

Worked example (illustrative)

Choosing the right experiment tool for a landing-page test

  • Setup: A San Diego home-security ecommerce account spends $22,000 per month on Search. The marketing team wants to test whether a new dedicated landing page (Page B, emphasizing a 30-day money-back guarantee) outperforms the current product page (Page A) in conversion rate. The base campaign uses an individual daily budget of $733 ($22,000 ÷ 30) and Maximize Conversion Value bidding—not a portfolio strategy.
  • Numbers: Current CVR = 3.2% on 4,100 sessions per month. At a 50/50 split, each arm receives approximately 2,050 sessions per month. To detect a 0.6 percentage-point improvement (from 3.2% to 3.8%)—a commercially meaningful threshold given $85 average order value and $1.19M annual revenue—at 80% power and p ≤ 0.05 requires approximately 2,900 sessions per arm. Expected time to reach 2,900 sessions per arm: 2,900 ÷ 2,050 = 1.41 months ≈ 43 days. Minimum runtime set to 45 days (March 1–April 15, 2027).
  • Decision: Launch a custom Search experiment with a 50/50 cookie-based split and final URL set to Page B in the treatment arm. Primary metric: conversion value per session. Guardrail: average order value must not fall below $80 in the treatment arm.
  • Why: A custom experiment is the correct tool because the change is a campaign-level setting (final URL), not an ad-copy change; cookie-based splitting ensures each user sees only one page version, preventing contamination.[4]

Done for you

Everything on this page, done for you before you launch

We build the campaign properly the first time — so you are not unpicking it three months from now.

✓  Keywords validated against real search volume, grouped by intent

✓  Responsive search ads written to the asset counts that actually perform

✓  A negative keyword list built before you spend, not after

✓  Delivered in 24 hours as a spreadsheet you keep

See what you get — $129

Human-reviewed · No subscription · We never touch your Google Ads account

5. Designing a Valid Experiment

Experimental validity depends on decisions made before the experiment launches, not after results arrive. The following framework applies to all Google Ads experiment types.[1][2]

Write a falsifiable hypothesis first

Every experiment must begin with a written hypothesis in the form: “Changing [one variable] from [control] to [treatment] will improve [primary business metric] by at least [minimum worthwhile effect], without breaching [guardrail metrics], because [mechanism].”

The minimum worthwhile effect is not optional. It forces the team to calculate whether the expected sample size is achievable given campaign volume and budget, and it prevents declaring a win on a commercially irrelevant improvement.[7]

Isolate one variable

Change one major variable in the treatment arm. If the experiment tests a new bid strategy, the landing page, ad copy, and audience settings must remain identical to the control. If the experiment tests a landing page, the bid strategy, ad copy, and audiences must remain identical. Multiple simultaneous changes make the result impossible to attribute.[1]

The only exception is a bundled test where the business decision is specifically whether to adopt an entire package—for example, the complete AI Max feature set—rather than individual components.[14]

Select the correct split method and ratio

Use a 50/50 split as the default. Equal allocation provides the highest statistical efficiency—it reaches significance fastest for a given total traffic volume—and simplifies interpretation.[4] Use an unequal split (70/30 or 80/20) only when the treatment carries meaningful downside risk or the control is the primary revenue source. A 90/10 split can function as a safety check but is generally too small a treatment arm to estimate a modest effect within a reasonable timeframe.

Use a cookie-based split when repeated user exposure affects the outcome—ad messaging, landing-page experience, audience behavior. Use a search-based split only when volume is severely limited and faster statistical accumulation outweighs user-level consistency. Do not change the split method after launch.[4][15]

Define primary metric and guardrails before launch

Designate one primary metric that will determine the outcome: qualified conversions, conversion value, CPA, ROAS, or profit. The updated Experiment Center allows up to two primary success metrics per custom experiment[11]; if using two, specify in advance which takes precedence when they conflict. Treat CTR, CPC, impression share, and Quality Score as diagnostic guardrails, not decision metrics.

Avoid mid-test changes

Do not edit the original campaign, the experiment campaign, conversion actions, attribution settings, or tracking templates while the test is running. Any change to either arm can restart Smart Bidding learning, introduce auction volatility, and contaminate the comparison.[1][3] Record any unavoidable changes (tracking outage, budget emergency) in a log and exclude the affected dates from the primary analysis.

Schedule in advance

Schedule the experiment start date three to seven days in the future. This allows Google’s ad review system to approve new ads and assets before measurement begins, preventing the first days of data from reflecting an arm with disapproved or partial creative.[2]

Worked example (illustrative)

Writing a complete hypothesis for a bidding-strategy experiment

  • Setup: An Atlanta legal services account spends $14,000 per month on Search. The campaign currently uses Maximize Conversions and generates 48 conversions per month at a CPA of $292. The account team wants to test whether adding a $275 target CPA cap maintains conversion volume while reducing CPA.
  • Numbers: Historical CPA = $292. Proposed target = $275 (5.8% reduction). Minimum worthwhile effect: ≥5% CPA reduction while maintaining ≥90% of conversion volume (floor: 43 conversions per month in treatment arm). At a 50/50 split, each arm receives approximately 24 conversions per month. To detect a 5.8% CPA improvement at 80% power and p ≤ 0.05 with this conversion volume, the experiment needs approximately 310 total conversions per arm—requiring 310 ÷ 24 = 12.9 months, well beyond practical scope. The team recalibrates: test for 60 days (accumulating ~48 conversions per arm), accept 70% confidence as a directional read[5], and document this as a directional experiment rather than a conclusive one.
  • Decision: Launch a 60-day custom experiment (August 1–September 30, 2027), 50/50 split, cookie-based, with target CPA = $275 in the treatment arm. Primary metric: cost per qualified consultation confirmed in CRM. Confidence threshold: 70% (directional). Guardrail: treatment arm must not drop below 43 conversions over 60 days.
  • Why: Google’s vendor guidance designates 70% confidence as an acceptable directional threshold for lower-volume accounts[5]; the result will be labeled directional and used to inform a larger-scale test rather than triggering an immediate account-wide rollout.

6. Sample Size, Duration and Significance

Sample size planning is the step most commonly skipped in Google Ads experiments, and skipping it is the primary cause of inconclusive results and false positives. The required number of conversions per arm depends on three inputs: the baseline conversion rate (or CPA/ROAS variability for value-based metrics), the minimum detectable effect (MDE), and the chosen power and significance levels.[2]

Minimum conversions per arm

Google’s vendor guidance identifies 100 conversions per variation as a working minimum, with 300 or more preferred for detecting effects smaller than 10–15%.[2] For ROAS and conversion-value experiments, the variance of revenue is substantially higher than the variance of a binary conversion event, so conversion count alone overstates available power—plan for more conversions or a wider MDE when testing value-based outcomes.

Duration floors

Even when conversion volume is adequate, run for at least four weeks (30 days) for Search experiments to capture full weekly cycles.[2] Shopping experiments require four to six weeks, or one to two full conversion cycles, whichever is longer.[3] When Smart Bidding or a new feature is in the treatment arm, exclude the first seven to fourteen days from the primary analysis to allow the bidding model to recalibrate before measurement begins.[2][3]

Statistical significance thresholds

Google’s platform reports experiment confidence using three labeled bands:[5]

  • 95% confidence (p ≤ 0.05): conclusive result; appropriate for high-stakes decisions such as bid-strategy changes on campaigns spending more than $10,000 per month.[2]
  • 80% confidence: directional; suitable for low-volume accounts or app-uplift experiments where 95% is unattainable in a reasonable timeframe.[6]
  • 70% confidence: weak directional signal; label explicitly as preliminary and use only to inform the design of a follow-up experiment, not to justify a rollout.[5]

These are vendor claims, not independently validated statistical standards. The conservative approach is to treat 95% confidence as the minimum for any consequential account change and to report lower-confidence results only as hypotheses for future testing.

Avoiding optional stopping bias

Do not check results daily and stop when a threshold is crossed. Repeated interim looks inflate the Type I error rate significantly above the nominal 5%. Set a fixed end date before launch, or use a predeclared sequential-testing protocol with adjusted thresholds. Document the stopping rule in the pre-launch checklist.[2]

Worked example (illustrative)

Calculating required duration for a Search conversion-rate experiment

  • Setup: A Houston roofing lead-generation account spends $9,500 per month on Search. The campaign generates 160 conversions per month (form submissions) at a 4.0% conversion rate across 4,000 sessions per month. The team wants to test a new landing page and needs to know how long to run the experiment at a 50/50 split.
  • Numbers: Baseline CVR = 4.0% (p₁ = 0.040). Target CVR = 4.8% (p₂ = 0.048; MDE = +0.8 percentage points, or +20% relative lift). Using the standard two-proportion z-test formula at α = 0.05 (two-tailed, z = 1.96) and 80% power (z = 0.842): n ≈ 2 × (1.96√(2 × 0.044 × 0.956) + 0.842√(0.040 × 0.960 + 0.048 × 0.952))² ÷ (0.008)² ≈ 2 × (1.96 × 0.290 + 0.842 × 0.288)² ÷ 0.000064 ≈ 2 × (0.569 + 0.243)² ÷ 0.000064 ≈ 2 × 0.659 ÷ 0.000064 ≈ 20,594 sessions per arm. At 50/50 split, each arm receives 2,000 sessions per month. Required duration: 20,594 ÷ 2,000 = 10.3 months—not practical. The team widens the MDE to 1.6 percentage points (from 4.0% to 5.6%), reducing per-arm sessions to approximately 5,100, or 2.55 months ≈ 77 days. Planned runtime: May 1–July 17, 2027 (77 days). Ramp-up exclusion: days 1–7 (May 1–7) excluded from primary analysis.[3]
  • Decision: Run the experiment for 77 days with the first 7 days excluded, targeting a minimum 1.6 percentage-point CVR improvement as the MDE. If the treatment arm does not reach 95% confidence at the end of 77 days, record as inconclusive and do not apply the new page.
  • Why: The standard four-week minimum[2] is insufficient here because the MDE requires 5,100 sessions per arm at the available traffic rate; widening the MDE is preferable to shortening the test below statistical adequacy.

Worked example (illustrative)

Applying the Shopping experiment duration rule

  • Setup: A Portland outdoor-gear retailer runs a Shopping campaign spending $31,000 per month. The account’s average conversion lag (time from click to purchase) is 18 days based on the attribution report. The team wants to test a switch from Maximize Conversion Value to target ROAS = 450%.
  • Numbers: Shopping experiment minimum = 4–6 weeks[3]; conversion lag = 18 days = 2.57 weeks. One conversion cycle ≈ 18 days; two cycles = 36 days. Required duration: maximum of (6 weeks = 42 days) and (2 conversion cycles = 36 days) = 42 days. Add 14-day Smart Bidding ramp-up exclusion[3]: total experiment window = 42 + 14 = 56 days. Planned dates: April 1–May 27, 2027. Measurement window: April 15–May 27 (after ramp-up). Performance Max asset experiments are not yet available in this account, so the team uses a custom experiment framework.
  • Decision: Run a 56-day Shopping experiment with a 50/50 split. Target ROAS = 450% in treatment arm (within 15% of the campaign’s trailing 30-day actual ROAS of 420%[3]). Primary metric: conversion value per $1 spent. Significance threshold: 95% confidence. Ramp-up days 1–14 excluded from analysis.
  • Why: Google’s vendor guidance specifies 4–6 weeks for Shopping experiments and 1–2 conversion cycles[3]; the 18-day conversion lag means the team must plan for 56 total days to allow both ramp-up exclusion and two full conversion cycles of mature data.

Not ready to change anything today?

We post the sharpest bits of these guides — what a change is worth, how long it takes to show up, and when to leave it alone. Short enough to read on your phone.

Follow on LinkedIn Follow on Instagram

7. What to Test: Bidding, Copy, Landing Pages, Audiences

Not all tests are equally high-leverage. Bidding-strategy experiments typically have the largest CPA or ROAS impact because they affect every auction. Landing-page experiments affect conversion rate after the click and require CRM integration to measure quality. Ad-copy experiments affect click quality and quantity. Audience experiments answer incrementality questions—whether the targeted users would have converted without the ad—and require careful interpretation.[1]

Bidding strategy experiments

Test: Manual CPC versus Maximize Conversions; Maximize Conversions versus target CPA; Maximize Conversion Value versus target ROAS; different target CPA or target ROAS levels; standard versus portfolio bid strategies; conversion-based versus value-based bidding when reliable revenue data is available.[1]

Requirements: the base campaign should have at least approximately 30 conversions in the prior 30 days.[1] Set the treatment target within ±15% of historical actual performance to avoid excessive learning instability.[3] Evaluate qualified conversions or revenue, not raw lead count, because automated bidding can hit volume targets by lowering lead quality.

Landing-page experiments

Test: message match between keyword/ad and page headline; short versus long forms; number of form fields; CTA specificity (e.g., “Get a Free Quote” versus “Learn More”); pricing transparency; social proof elements (reviews, trust badges, case studies); page load speed; dedicated campaign page versus generic product page; different offers (consultation, demo, trial, estimate, download).[1]

Measure conversion rate and downstream quality. A page that increases raw form submissions but reduces the qualified-lead rate is not a winner. Connect CRM outcome data—accepted leads, opportunities, closed revenue—before declaring the test complete.

Ad-copy and creative experiments

Test one message dimension at a time: core value proposition; benefit versus feature framing; specific versus general CTA; price or promotional claim; trust or expertise claim; local relevance signal; urgency element. For responsive search ads, test adding specific new headline or description assets.[1]

Use ad variations when the objective is to test one consistent change across multiple Search campaigns. Use a custom experiment when the copy change is campaign-specific or coupled with a landing-page or audience change.[1]

Measure qualified conversion rate and conversion value first. CTR is a diagnostic metric: stronger copy can attract low-intent clicks, while a lower-CTR ad can produce more profitable conversions at better CPA.

Audience and targeting experiments

Test: observation versus targeted audience settings; first-party customer match lists; remarketing windows (7, 30, or 90 days); in-market segments; broad targeting versus specific audience signals in automated campaigns; new-customer acquisition goals; targeting expansion settings; geographic or demographic exclusions with a documented business rationale.[1]

Assess incremental conversions, new-customer rate, and total qualified revenue—not audience-segment CPA in isolation. A remarketing audience may show a CPA of $45 while the incrementality study shows that 80% of those users would have converted organically, making the true incremental CPA $225.

Worked example (illustrative)

Bidding experiment: Maximize Conversions versus target CPA on a lead-gen campaign

  • Setup: A Minneapolis financial planning firm runs a Search campaign spending $7,200 per month. Trailing 30-day performance: 54 conversions (consultation requests), CPA = $133, conversion rate = 6.1% on 885 clicks. The team hypothesizes that adding a $120 target CPA will reduce cost per qualified consultation without materially reducing volume.
  • Numbers: Historical CPA = $133. Treatment target CPA = $120 (9.8% reduction; within the ±15% historical-performance guidance[3]: 15% of $133 = $20, so the allowable range is $113–$153). At a 50/50 split, each arm receives $3,600 per month in budget and approximately 27 conversions per month. Minimum duration: 4 weeks[2], accumulating ~27 conversions per arm. This falls below the 100-conversion working floor[2], so the team extends to 90 days (June 1–August 30, 2027), targeting ~81 conversions per arm. Ramp-up exclusion: days 1–14 excluded from analysis.[3] Effective measurement window: June 15–August 30 (76 days), targeting ~69 conversions per arm. Confidence threshold: 70% directional (insufficient volume for 95%).[5]
  • Decision: Launch 90-day custom experiment, 50/50 cookie-based split, target CPA = $120 in treatment arm, Maximize Conversions in control arm. Primary metric: cost per CRM-confirmed qualified consultation. If the treatment arm achieves ≥70% confidence and CPA ≤ $125 with ≥90% of control conversion volume (≥62 conversions over 76 measurement days), apply and schedule a follow-up 90-day test targeting 95% confidence.
  • Why: The $120 target is within ±15% of $133 historical CPA[3]; the 70% directional threshold is appropriate because the 90-day window produces only ~69 measurable conversions per arm, below the 100-conversion minimum for a conclusive result.[2][5]

Worked example (illustrative)

Ad-copy experiment: benefit versus feature framing for a SaaS account

  • Setup: A Boston SaaS project-management company spends $26,000 per month on Search. Current top responsive search ad uses feature-framing headlines (“Unlimited Projects,” “Real-Time Gantt Charts,” “API Access Included”). The team hypothesizes that benefit-framing headlines (“Ship Projects 30% Faster,” “Never Miss a Deadline,” “Teams Scale Without Extra Headcount”) will increase qualified trial starts.
  • Numbers: Current campaign: 3,900 clicks per month, CVR = 5.1% (199 trial starts), CPA = $131. At a 50/50 split, each arm receives 1,950 clicks per month. To detect a 1.0 percentage-point CVR improvement (from 5.1% to 6.1%, a 19.6% relative lift) at 80% power and p ≤ 0.05 requires approximately 3,200 sessions per arm. Time to reach 3,200 sessions at 1,950 per month: 3,200 ÷ 1,950 = 1.64 months ≈ 50 days. Planned runtime: September 1–October 21, 2027 (50 days). Expected conversions per arm: 1,950 × (50/30) × 0.051 = 165 conversions—above the 100-conversion working floor.[2]
  • Decision: Launch an ad variation experiment applying benefit-framing headlines to the treatment arm across all ad groups in the Search campaign. Primary metric: cost per CRM-confirmed trial that completes onboarding (not raw trial start). Guardrail: treatment-arm average CPC must not exceed $14.00 (current average = $13.33). Significance threshold: 95% confidence (p ≤ 0.05).[2]
  • Why: With 165 expected conversions per arm over 50 days, the experiment exceeds the 100-conversion working floor[2] and the 50-day duration exceeds the 30-day minimum[2]; the downstream onboarding-completion metric prevents a CTR-driven false positive.[1]

8. Reading Results and Avoiding False Positives

Google’s Experiment Center reports results as a confidence level alongside a point estimate and confidence interval for each primary metric. Reading these outputs correctly is critical to avoiding both Type I errors (declaring a winner that is not real) and Type II errors (discarding a treatment that actually works).[2]

Use the confidence interval, not just the point estimate

A point estimate of “+12% CPA improvement” is meaningless without the interval. If the 95% confidence interval is [+2%, +22%], the result is statistically significant and practically meaningful. If the interval is [−3%, +27%], the result is not conclusive at 95%—it includes the possibility of a 3% worse CPA. Report the interval in every experiment debrief.[2]

Check integrity before reading results

Before interpreting any metric, verify: (1) the traffic or budget split executed as planned; (2) no arm was budget-constrained, policy-paused, or eligibility-limited during the test; (3) conversion settings and attribution remained unchanged; (4) no major mid-test edits occurred; (5) the ramp-up period has been excluded from the analysis window.[1][3]

Evaluate primary metric and guardrails in sequence

Assess the primary metric first. If it clears the predeclared threshold (95% confidence for conclusive, 70–80% for directional), check each guardrail metric. A winning primary metric with a materially deteriorated guardrail—for example, CPA improves 12% but qualified-lead rate in the CRM drops 18%—is not a genuine win. The guardrail failure overrides the primary metric result.[1]

Distinguish statistical significance from practical significance

A statistically significant result at high conversion volume can represent a commercially trivial absolute improvement. A campaign spending $20,000 per month where the treatment saves $400 per month (2% CPA improvement at 95% confidence) may not justify the operational cost of applying and monitoring the change. Assess absolute dollar impact, not percentage alone.[2]

Treat “inconclusive” as a valid outcome

When results are inconclusive—the confidence interval spans both meaningful gains and meaningful losses—do not apply the treatment. Record the result, refine the hypothesis, increase the volume requirement, and design a follow-up experiment. Treating “no significant difference” as evidence that the treatment is better is a logical error with potentially expensive consequences.[1]

Worked example (illustrative)

Identifying a guardrail failure that overrides a positive primary metric

  • Setup: A Phoenix insurance leads account runs a 45-day Search experiment (February 1–March 18, 2027) testing a new landing page against the control. Primary metric: cost per form submission. Guardrail: CRM-confirmed qualified-lead rate must not fall below 38% (current rate = 42%).
  • Numbers: After 45 days, treatment arm: 218 form submissions at CPA = $38.20. Control arm: 194 form submissions at CPA = $44.50. Primary metric improvement: ($44.50 − $38.20) ÷ $44.50 = 14.2% CPA reduction. Confidence: 96% (clears 95% threshold[2]). Guardrail check: CRM data shows treatment-arm qualified-lead rate = 29% (63 of 218 submissions). Control-arm qualified-lead rate = 41% (80 of 194 submissions). Qualified lead CPA in treatment: $38.20 ÷ 0.29 = $131.72. Qualified lead CPA in control: $44.50 ÷ 0.41 = $108.54. The treatment arm’s qualified-lead CPA is $23.18 (21.3%) higher than control despite the lower raw CPA.
  • Decision: Do not apply the treatment page. Record the experiment as a guardrail failure. The new landing page generated more raw submissions at lower cost but attracted a substantially lower-quality audience. Redesign the page to include a qualifying question before the form submit.
  • Why: The predeclared guardrail (qualified-lead rate ≥ 38%) was breached at 29%[1]; the downstream qualified-lead CPA worsened by 21.3%, meaning the apparent primary-metric win was entirely explained by attracting lower-quality traffic rather than improving efficiency.

“So what do I actually have to hand over?”

Nothing. No access to your Google Ads account, no login, no agency link request. You tell us the business and the offer; we send back a spreadsheet.

You import it yourself, or hand it to whoever runs the account. It is yours to keep and re-use — there is no ongoing fee and nothing to cancel.

How it works — $129

9. Applying or Rolling Back a Winner

Applying an experiment winner in Google Ads is a one-way operation: once the experiment is applied to the original campaign, the original campaign adopts the treatment settings. Rolling back requires manually reverting every changed setting, which is operationally complex at scale. Treat the apply decision with the same rigor as the launch decision.[1][7]

Pre-application checklist

  • Confirm the primary metric clears the predeclared significance threshold (95% confidence for conclusive results).[2]
  • Confirm all guardrail metrics are within acceptable ranges.
  • Confirm conversion tracking, offline conversion imports, and attribution settings are complete and consistent.
  • Verify that the winning condition still matches current market conditions: a target CPA that worked during a February promotion may not be appropriate in July.
  • Document the hypothesis, dates, split, settings, primary result, guardrails, limitations, and the person who approved the application decision.
  • Check that the experiment campaign’s eligibility is still valid: individual daily budget, standard bidding strategy (not portfolio), and active status.[11]

Application options

Google Ads offers two application pathways for a winning experiment:[1][7]

  • Apply to the original campaign: the original campaign adopts the treatment settings and the experiment campaign is ended. Use this when the volume split is acceptable and there is no reason to maintain both campaigns.
  • Graduate the experiment campaign: the experiment campaign becomes an independent campaign. Use this when the team wants to retain the original campaign in its current state for a specific segment or as a continuing control.

Rolling back

If an experiment result was applied incorrectly—or if post-application performance deteriorates materially—roll back by manually reverting each changed setting in the now-modified original campaign. There is no one-click revert function in Google Ads. Maintain a configuration record (bidding strategy, targets, match types, URLs, audiences, exclusions, and tracking templates) before every application so that the rollback checklist is specific rather than reliant on memory.

Staged rollout for high-risk changes

For campaigns spending more than $30,000 per month or for changes that substantially alter the bidding strategy or audience targeting, consider a staged rollout: apply the winning settings to one geographic or audience segment, monitor for seven days, then apply broadly. This reduces the blast radius of an application error.[1]

Worked example (illustrative)

Staged application of a target-ROAS bidding experiment winner

  • Setup: A Seattle ecommerce outdoor-furniture account spends $47,000 per month on Search across three geographic campaigns: Pacific Northwest ($19,000), Mountain West ($16,000), and Southwest ($12,000). A 60-day experiment (January 15–March 16, 2027) tested target ROAS = 500% against Maximize Conversion Value. The Pacific Northwest campaign served as the experiment base.
  • Numbers: Pacific Northwest experiment result: treatment ROAS = 532% versus control ROAS = 481% over 60 days (excluding 14-day ramp-up[3]). Improvement: (532 − 481) ÷ 481 = 10.6%. Confidence: 97% (clears 95% threshold[2]). Guardrail: conversion volume in treatment arm = 184 purchases versus control arm = 191 purchases (3.7% volume difference—within predeclared ±10% guardrail). Application decision: apply to Pacific Northwest immediately; stage Mountain West and Southwest rollout over two weeks (March 20 and April 3, 2027 respectively).
  • Decision: Apply target ROAS = 500% to the Pacific Northwest campaign on March 17, 2027. Apply target ROAS = 500% to the Mountain West campaign on March 20, 2027, after a 3-day monitoring window. Apply to the Southwest campaign on April 3, 2027, after a 14-day monitoring window. Document all pre-application settings in a shared configuration log.
  • Why: The staged rollout limits total monthly spend at risk during application to $19,000 (Pacific Northwest only) in the first 72 hours; the ±10% conversion-volume guardrail was predeclared and met, making the application defensible.[1][2]

10. Testing with Smart Bidding and Automation

Smart Bidding experiments require additional design care because the treatment arm’s bidding model is learning throughout the experiment. Performance during the first seven to fourteen days reflects model initialization, not steady-state behavior. Evaluating the full experiment window including this ramp-up period produces systematically biased results that understate the treatment arm’s true performance.[2][3]

Smart Bidding ramp-up and evaluation windows

Exclude the first seven to fourteen days from the primary analysis of any experiment where the treatment arm uses a new Smart Bidding strategy or a meaningfully different target. The specific exclusion period depends on campaign volume: high-volume campaigns (more than 50 conversions per week) may stabilize in seven days; lower-volume campaigns may require fourteen days or more.[3] Report the full period for transparency but designate the post-ramp window as the primary evaluation period in the pre-launch plan.

Target proximity

Set the treatment arm’s target CPA or target ROAS within ±15% of the campaign’s actual historical performance over the prior 30 days.[3] Targets set outside this range cause extended learning instability and produce experiment data that reflects the bidding model’s struggle to reach an unrealistic target rather than the target’s inherent value.

AI Max experiments

AI Max experiments use a 50/50 split of the existing campaign and test AI Max features—including search-term matching expansion and asset optimization—against current settings. Because both arms share the same campaign structure and bidding strategy (only the AI Max features differ), the ramp-up exclusion is typically shorter than for a full bidding-strategy change. The key evaluation metric is conversion value or CPA, not CTR or search-term coverage alone, because AI Max can expand to new search terms that inflate impressions without proportional conversion value.[4][11]

Performance Max asset experiments

As of November 2026, Performance Max asset experiments are announced but rolling out progressively.[8] When available, use them to compare new asset sets against existing assets before scaling new creative. The primary metric should be conversion value or ROAS, not asset-strength rating (which is an internal Google quality signal, not a business outcome measure).[10]

Portfolio bidding experiments

Custom experiments are not supported on campaigns using portfolio bid strategies.[11] To test a portfolio strategy change, the team must either (1) move the campaign to a standard individual bid strategy, run the experiment, and then convert the winner back to a portfolio; or (2) compare two separate but structurally similar campaigns using an observational rather than controlled design—and label the result accordingly as non-causal.

Worked example (illustrative)

AI Max experiment design for a national retail Search campaign

  • Setup: A national US home-goods retailer runs a Search campaign spending $55,000 per month. The campaign uses exact and phrase match keywords with Maximize Conversion Value bidding and generates 310 purchases per month at an average order value of $177 and ROAS of 390%. Google recommends enabling AI Max features. The team wants to test AI Max before enabling it account-wide.
  • Numbers: Current ROAS = 390%. At a 50/50 split, each arm receives $27,500 per month in budget and approximately 155 purchases per month. AI Max experiments use a 50/50 split of the existing campaign[4]; this produces 155 conversions per arm per month, exceeding the 100-conversion working floor[2] and reaching the preferred 300-conversion bar within 2 months (310 conversions per arm over 60 days). Planned runtime: July 1–August 30, 2027 (60 days). Ramp-up exclusion: days 1–7 (July 1–7).[3] Measurement window: July 8–August 30 (53 days). Expected conversions per arm in measurement window: 155 × (53/30) = 274 conversions. Confidence threshold: 95% (p ≤ 0.05).[2]
  • Decision: Launch AI Max experiment via the Experiments page, 50/50 split, July 1, 2027. Primary metric: ROAS (conversion value ÷ cost). Guardrail: search-term quality—if the new-customer rate drops below 15% or branded search cost increases more than 10% in the treatment arm during the measurement window, flag for review before the end date. Significance threshold: 95%.
  • Why: AI Max experiments require no separate campaign copy[4] and the 60-day window produces 274 conversions per arm in the measurement period—above the 100-conversion working floor[2] and approaching the 300-conversion preferred threshold, making a 95% confidence result achievable.[2]

11. Common Mistakes to Avoid

Even well-resourced accounts repeat the same experimental errors. The mistakes below are listed in approximate order of frequency and business cost. Each one has a specific corrective action.

Mistake 1: Testing too many variables at once

Running a single experiment that changes the bid strategy, ad copy, landing page, and audience simultaneously produces a result that cannot be attributed to any single factor. If performance improves, you cannot isolate what drove it. If performance declines, you cannot reverse only the harmful change. Google specifically recommends changing one major variable per experiment so the result is attributable.[1] The only defensible exception is a bundled test where the business decision is explicitly whether to adopt the entire package as a unit—for example, adopting AI Max features as a whole when the account cannot afford to separate them.[4]

Mistake 2: Ending the test too early

Stopping as soon as results look favorable is called optional-stopping bias. Because conversion counts fluctuate daily and bidding algorithms take time to stabilize, early data is noisier than late data. Google recommends a minimum of four weeks for most Search experiments and four to six weeks for Shopping experiments.[3] Tests involving automated bidding require an additional ramp-up exclusion of seven to fourteen days.[3] Stopping early inflates the false-positive rate well above the nominal 5% associated with a 95% confidence threshold.

Mistake 3: Declaring a winner based on CTR alone

A higher click-through rate does not guarantee more qualified conversions, more revenue, or a better return on ad spend. An ad that attracts high-intent users may produce a lower CTR and a substantially higher conversion rate and conversion value. Google’s experiment reporting surfaces CTR and CPC as diagnostic metrics; they should not substitute for the predeclared primary metric—qualified conversions, conversion value, CPA, or ROAS.[2]

Mistake 4: Running overlapping experiments on the same campaign

Google explicitly warns that running multiple experiments simultaneously on the same campaign, or on campaigns that compete in the same auctions, can cause experiments to interfere with one another and make results less reliable.[3] The correct approach is to run experiments sequentially on a given campaign and to stagger experiments across campaigns that share significant auction overlap.

Mistake 5: Changing a live arm during the test

Editing the original or treatment campaign after launch—adjusting bids, adding negative keywords, changing budgets, swapping assets, or altering conversion actions—can restart bidding model learning, introduce auction volatility, and contaminate the comparison. Keep both arms frozen for the duration. If an urgent change is required, document it, consider ending the experiment, and re-run from a clean baseline.[1]

Mistake 6: Using a shared budget

Campaigns using a shared budget are ineligible for custom experiments and draft-based experiments.[11] A shared budget can cause one arm to receive systematically more or less spend than intended, invalidating the traffic split. Move the base campaign to an individual daily budget before creating the experiment. This applies to both the original and the experiment campaign arm.

Mistake 7: Ignoring conversion lag when reading results

If the account’s average conversion lag is 14 days and the experiment ended yesterday, conversions from the final two weeks are still arriving. Reading results before the data matures understates both arms but can favor the arm whose conversions happen to arrive faster—often the control, whose bidding model is more established. Exclude recent dates equal to at least one full conversion lag window before finalizing the analysis.[2]

Mistake 8: Treating “no significant difference” as proof the control is better

An inconclusive result means the data did not accumulate enough evidence to distinguish the arms—it does not confirm that the control outperforms the treatment or vice versa. The correct response to an inconclusive result is to retain the control, assess whether the test was adequately powered, and redesign with a larger expected effect or a higher-volume campaign.[2]

Mistake 9: Setting an unrealistic target CPA or ROAS in the treatment arm

For Performance Max experiments, Google advises keeping targets within roughly 15% of actual historical performance.[3] An aggressive target causes the bidding model to under-spend or chase low-probability conversions during the measurement window, producing an artificially negative result for the treatment that reflects the target, not the strategy itself.

Mistake 10: Skipping the pre-launch checklist

Conversion tracking failures, broken final URLs, missing audience lists, and attribution mismatches discovered mid-experiment force either a flawed result or a full restart. Schedule the experiment at least three to seven days in advance so ads complete review, URLs resolve correctly, tracking fires, and the audience list meets the minimum 10,000-user threshold for cookie-based custom experiments before measurement begins.[2][4]

Worked example (illustrative)

Early-Stop Bias on a Low-Volume Lead-Gen Account

  • Setup: A Charlotte, NC commercial roofing account spending $8,000 per month on a Search campaign. Baseline conversion rate: 4.2% (form submissions). Average order value: $18,500. The account runs a custom experiment testing a new landing page (treatment) versus the existing page (control), 50/50 split, with a 95% confidence threshold and a minimum runtime of 28 days.[2]
  • Numbers: Monthly clicks at $8,000 ÷ $6.40 average CPC = 1,250 clicks per month; 625 per arm per month. At a 4.2% baseline conversion rate, each arm receives 1,250 × 0.042 ÷ 2 = 26.25 conversions per month—below the 100-conversion working floor.[2] At day 7, the treatment arm shows 9 conversions versus 5 in the control (a 55% difference). The account manager flags this as a potential winner. However, with only 14 total conversions across both arms, the standard error of the difference in proportions is approximately ±3.1 percentage points, and 95% confidence intervals overlap completely. Required conversions per arm for 80% power to detect a 2-percentage-point lift (from 4.2% to 6.2%) at p ≤ 0.05: approximately 870 per arm, requiring roughly 33 months at current volume.
  • Decision: Do not declare a winner at day 7. Extend to 28 days (the minimum), collect 105 total conversions across both arms, then reassess. If still underpowered, migrate the test to a higher-volume branded + non-branded combined campaign spending $22,000 per month that generates 92 conversions per month per arm over the same 28-day window.
  • Why: Google’s minimum recommended runtime is 4 weeks (28 days)[2] and the preferred conversion threshold is 300 per variation[2]; stopping at 14 total conversions produces a false-positive rate far exceeding the nominal 5% associated with p ≤ 0.05.

Worked example (illustrative)

Shared-Budget Eligibility Failure Before Launch

  • Setup: A Portland, OR e-commerce account running three Search campaigns under a single shared daily budget of $400 ($146,000 annualized). The account manager attempts to create a custom experiment on the highest-volume campaign (110 conversions per month) to test Maximize Conversion Value versus target ROAS at 500%.
  • Numbers: Shared budget of $400/day across three campaigns. Custom experiment eligibility requires an individual daily budget.[11] The highest-volume campaign historically consumes approximately 58% of the shared budget, or $232/day ($6,960/month). To run the experiment, the account manager moves this campaign to an individual daily budget of $232, leaving the remaining two campaigns to share $168/day. At the 50/50 split, each arm receives $116/day ($3,480/month). At the account’s historical $4.10 CPC and 3.8% conversion rate, each arm produces approximately $3,480 ÷ $4.10 × 0.038 = 32.3 conversions per 30-day measurement window—below the 100-conversion floor[2] but sufficient for a directional read at 80% confidence.[6]
  • Decision: Move the campaign to an individual $232/day budget before creating the experiment. Accept that 32 conversions per arm over 30 days supports only an 80% confidence directional result, not a 95% conclusive result. Set the experiment end date to 90 days (June 1–August 29, 2027) to accumulate approximately 97 conversions per arm—approaching the 100-conversion working floor—then evaluate at 95% confidence.[2]
  • Why: Custom experiments require an individual daily budget, not a shared budget[11]; failing to make this change before launch prevents experiment creation entirely.

Worked example (illustrative)

Overlapping Experiments Corrupting Bidding-Strategy Results

  • Setup: A Dallas, TX SaaS account with two Search campaigns—Brand (60 conversions/month, $5,000/month spend) and Non-Brand (130 conversions/month, $18,000/month spend)—that compete in overlapping auctions for generic software keywords. The account manager launches a tCPA experiment on the Non-Brand campaign (treatment tCPA: $185; historical CPA: $192) on March 1, 2027, then launches a second experiment on the Brand campaign on March 10, 2027 to test a new RSA headline.
  • Numbers: The Brand campaign’s RSA experiment changes messaging that also appears in non-brand auctions via broad match. From March 10 onward, users in the Non-Brand tCPA experiment treatment arm are exposed to the new brand messaging, introducing a confound. If the Non-Brand treatment arm shows a CPA improvement of $192 → $174 (a 9.4% reduction) in the March 10–April 30 window, the result cannot be attributed solely to the tCPA change—the RSA change is a simultaneous uncontrolled variable. Google warns that simultaneous overlapping experiments make results less reliable.[3] The correct schedule: run the Non-Brand tCPA experiment from March 1–April 28, 2027 (58 days, ~251 conversions per arm at 130/month × 58/30). Start the Brand RSA experiment no earlier than May 1, 2027.
  • Decision: End the Brand RSA experiment immediately on March 10, 2027. Re-run the Non-Brand tCPA experiment from March 1–April 28, 2027 without the overlapping Brand experiment. Schedule the Brand RSA experiment to begin May 5, 2027, after the Non-Brand result is finalized and applied or discarded.
  • Why: Google explicitly advises against running overlapping experiments on campaigns that share auctions because they interfere with each other and make results less reliable.[3]

12. What Changed Recently (Last 30 Days)

The following changes were announced or confirmed in Google Ads documentation and official communications as of November 2026. Where a feature’s availability is still account- or campaign-dependent, that caveat is noted explicitly.

Performance Max asset experiments

Google announced that advertisers will be able to test creative assets directly within Performance Max campaigns, comparing new assets against existing assets or testing additional creative types within the same campaign.[8] This is the most significant new experiment type introduced in this period. Prior to this announcement, Performance Max offered limited controlled testing at the asset level; creative decisions were largely delegated to Google’s machine learning without an explicit A/B framework. Asset experiments for Performance Max are rolling out progressively and may not yet be available in all accounts. When available, the recommended approach mirrors other asset tests: define one primary success metric (conversion value or conversions), isolate one creative variable (for example, a new image asset set versus the existing set), and run for a minimum of four weeks to allow the asset rotation model to stabilize.[2]

Experiment Center: unified hub for experiments and Lift Studies

Google has formalized the Experiment Center as the single destination for campaign experiments, Lift Studies, asset tests, and campaign-type impact tests.[10] Previously, campaign experiments and Brand Lift or Search Lift studies were managed in separate parts of the interface. The Experiment Center consolidates creation, reporting, and incrementality measurement. Advertisers can now compare the effect of adding a campaign type—for example, adding Demand Gen alongside a Search campaign—using the same controlled framework as a standard A/B experiment. This is particularly useful for accounts evaluating whether to expand into new campaign types without committing budget account-wide before validation.

AI Max experiments as a one-click Search test

AI Max experiments are now a distinct, one-click test type for Search campaigns.[4] Unlike a custom experiment that requires duplicating the base campaign, an AI Max experiment splits the existing campaign 50/50, with one arm running standard Search behavior and the other arm activating AI Max features such as expanded search-term matching and asset optimization.[4] This eliminates the setup overhead of creating a parallel campaign and ensures the split is applied to the same campaign budget. The fixed 50/50 split is set by Google and cannot be altered for this experiment type, which means each arm receives equal budget and traffic from launch. Accounts evaluating AI Max adoption should use this path rather than a manual custom experiment because it preserves a direct control comparison within the same campaign structure.[14]

Custom experiment setup simplified with up to two primary metrics

The updated custom experiment workflow now allows advertisers to designate up to two primary success metrics and presents experiment types—Ad variations, Custom, and AI Max—within a single Experiments workflow rather than separate sections of the interface.[14] Custom experiments remain available for Search, Display, Demand Gen, and Video campaigns; they are not supported for App or Shopping campaigns under the current workflow.[14] The consolidated reporting view shows the original and experiment campaigns side by side with confidence indicators, making it easier to assess whether a result has reached the 95% conclusive threshold or only the 70–80% directional threshold without exporting data.[5]

Recommendation-to-experiment workflow

Google now enables advertisers to convert a system recommendation directly into an experiment, testing the proposed change against the base campaign before applying it account-wide.[9] This is particularly valuable for bid-strategy recommendations, targeting-expansion suggestions, and campaign-setting changes where account-wide application can be difficult to reverse. Rather than accepting or dismissing a recommendation outright, advertisers can route it through the standard experiment framework with a predefined primary metric and runtime, then apply it only if the result clears the 95% confidence threshold.[2] This workflow does not change the statistical requirements: the same minimum duration, conversion volume, and significance rules apply.

Video experiments expanded globally

Google announced that video experiments—which allow structured A/B testing of creative and measurement of results before scaling a video change—are being expanded globally.[9] Availability may still vary by account and campaign type; confirm in the Experiments page before planning a video creative test. Where available, the recommended approach is to test one creative variable (opening frame, audio-on versus audio-off, call-to-action overlay, or duration), use a 50/50 split, and run for at least four weeks to capture sufficient conversion signal.[2]

Explicit guidance against overlapping experiments

Google’s updated documentation now explicitly advises against running multiple experiments simultaneously when they share campaigns, auctions, audiences, budgets, or conversion signals.[3] This guidance formalizes what was previously implicit best practice. The recommended approach is to run experiments sequentially on any given campaign and to stagger experiments across campaigns with significant auction overlap. If two campaigns must be tested simultaneously, they should target non-overlapping geographies, audiences, or query sets to prevent cross-arm contamination.

Summary of recent changes

Change What it means for practitioners Availability Source
Performance Max asset experiments First controlled creative A/B framework for PMax; test new asset sets against existing assets Rolling out; not universal [8]
Experiment Center (unified hub) Consolidates campaign experiments, Lift Studies, asset tests, and campaign-type impact tests in one location Generally available; feature set varies by account [10]
AI Max one-click experiment 50/50 split of existing Search campaign; no campaign duplication required; tests AI Max matching and asset optimization Available for eligible Search campaigns [4][14]
Up to 2 primary success metrics in custom experiments Reduces the need to export data to evaluate a second business metric; still requires one predeclared decision metric Available in updated custom experiment workflow [14]
Recommendation-to-experiment workflow Route system recommendations through controlled test before applying account-wide; reduces irreversible changes Available for select recommendation types [9]
Video experiments expanded globally Structured creative A/B testing for Video campaigns before scaling; subject to account eligibility Expanding globally; confirm in Experiments page [9]
Explicit no-overlap guidance Google now formally documents that simultaneous overlapping experiments produce unreliable results; run sequentially Applies to all experiment types [3]

Worked example (illustrative)

Using the Recommendation-to-Experiment Workflow for a tROAS Change

  • Setup: A Minneapolis, MN home furnishings e-commerce account spending $22,000 per month on a Search campaign. Current strategy: Maximize Conversion Value (no target). Google’s recommendation engine suggests applying a target ROAS of 420%. Historical 90-day ROAS: 388%. The account manager wants to validate the recommendation before applying it account-wide.
  • Numbers: Monthly conversions (purchases): 148 at an average order value of $310. At 50/50 split, each arm receives $11,000/month and approximately 74 conversions/month. Google advises keeping tROAS targets within roughly 15% of actual historical performance[3]; 420% is 8.2% above the 388% historical ROAS, within the 15% guideline. Planned runtime: September 1–October 28, 2027 (58 days). Ramp-up exclusion: days 1–7 (September 1–7).[3] Measurement window: September 8–October 28 (50 days). Conversions per arm in measurement window: 74 × (50/30) = 123 conversions—above the 100-conversion working floor.[2] Confidence threshold: 95% (p ≤ 0.05).[2] Primary metric: ROAS. Guardrail: conversion volume must not fall more than 10% below the control arm (i.e., treatment arm must maintain ≥ 111 conversions in the measurement window).
  • Decision: Route the system tROAS recommendation through the custom experiment workflow instead of applying it directly. Set treatment arm to tROAS = 420%, control arm to Maximize Conversion Value (no target), 50/50 split, September 1–October 28, 2027. Primary metric: ROAS. Guardrail: conversion volume ≥ 111 in treatment measurement window. Apply the tROAS target only if ROAS in the treatment arm exceeds 388% at 95% confidence and the conversion-volume guardrail is met.
  • Why: The recommendation-to-experiment workflow allows the proposed tROAS change to be validated against a live control before account-wide application[9]; the 123 conversions per arm in the measurement window exceed the 100-conversion working floor required for a 95% confidence result.[2]

Worked example (illustrative)

Performance Max Asset Experiment for a New Image Asset Set

  • Setup: A San Diego, CA direct-to-consumer skincare brand running a Performance Max campaign at $15,000 per month with 210 purchases per month at an average order value of $68 and a current ROAS of 955%. The creative team has produced a new lifestyle image asset set (12 images, 5 logos) to replace the current product-only image set. The account wants to test the new asset set using the Performance Max asset experiment feature once it becomes available in the account.
  • Numbers: At 50/50 split, each arm receives $7,500/month and approximately 105 purchases/month. Planned runtime: October 1–November 28, 2027 (59 days). Ramp-up exclusion: days 1–7 (October 1–7).[3] Measurement window: October 8–November 28 (52 days). Conversions per arm in measurement window: 105 × (52/30) = 182 conversions—above the 100-conversion working floor[2] and approaching the 300-conversion preferred threshold within a second 60-day run if needed. Primary metric: conversion value (revenue). Guardrail: ROAS must not fall below 860% (a 10% decline from the 955% baseline) in the treatment arm during the measurement window. Confidence threshold: 95% (p ≤ 0.05).[2]
  • Decision: Launch the Performance Max asset experiment on October 1, 2027 once confirmed available in the account. Treatment: new lifestyle image asset set. Control: existing product-only image set. All other asset types (text, video, audience signals) held constant. Primary metric: conversion value. Guardrail: treatment ROAS ≥ 860%. If treatment ROAS ≥ 955% at 95% confidence and guardrail is met, apply the new asset set. If inconclusive after 59 days, run a second 60-day window before deciding.
  • Why: Performance Max asset experiments are the first controlled creative A/B framework for PMax[8]; the 182 conversions per arm in the measurement window exceed the 100-conversion working floor[2], making a 95% confidence result achievable within the planned runtime.

Worked example (illustrative)

Experiment Center: Testing the Incremental Impact of Adding Demand Gen

  • Setup: A Boston, MA B2B software company running Search campaigns at $30,000 per month with 95 qualified leads per month (MQL rate: 38% of raw form fills, or 250 form fills × 0.38). The growth team wants to know whether adding a Demand Gen campaign at $6,000 per month produces incremental MQLs or simply re-contacts users already captured by Search.
  • Numbers: Total addressable monthly budget: $36,000. Control arm: Search only at $30,000/month. Treatment arm: Search at $30,000/month + Demand Gen at $6,000/month. The Experiment Center’s campaign-type impact test divides eligible users into control and treatment groups.[10] Planned runtime: August 1–September 28, 2027 (59 days). Ramp-up exclusion: days 1–14 (August 1–14) to allow Demand Gen bidding model to stabilize.[3] Measurement window: August 15–September 28 (45 days). Expected Search conversions per arm in measurement window at 95 MQLs/month × (45/30) = 143 MQLs. Primary metric: incremental MQLs (treatment arm MQLs minus control arm MQLs, normalized for budget). Secondary guardrail: Search CPA in the treatment arm must not increase more than 10% above the control arm’s Search CPA during the measurement window (baseline Search CPA: $30,000 ÷ 95 = $315.79 per MQL; guardrail ceiling: $347.37 per MQL). Confidence threshold: 95%.[2]
  • Decision: Run the campaign-type impact test via the Experiment Center, August 1–September 28, 2027. If the treatment arm produces ≥ 15 incremental MQLs (a 10.5% lift above the 143 expected) at 95% confidence and Search CPA in the treatment arm remains below $347.37, approve the $6,000 Demand Gen budget as a permanent addition. If incremental MQLs are fewer than 8 (a 5.6% lift) or Search CPA breaches the guardrail, discontinue Demand Gen.
  • Why: The Experiment Center’s campaign-type impact test is the appropriate tool for measuring the incremental value of adding a campaign type[10]; without a controlled comparison, overlap between Search and Demand Gen audiences makes it impossible to distinguish incremental lift from audience duplication.

References

  1. [1] https://support.google.com/google-ads/answer/10682377 support.google.com
  2. [2] https://developers.google.com/google-ads/api/docs/experiments/reporting developers.google.com
  3. [3] https://support.google.com/google-ads/answer/13826584?hl=en support.google.com
  4. [4] https://support.google.com/google-ads/answer/9124739?hl=en-AF support.google.com
  5. [5] https://support.google.com/google-ads/answer/10436762?hl=en support.google.com
  6. [6] https://support.google.com/google-ads/answer/14074599?hl=en-IR support.google.com
  7. [7] https://developers.google.com/google-ads/api/docs/experiments/overview developers.google.com
  8. [8] https://business.google.com/us/accelerate/announcements/ business.google.com
  9. [9] https://support.google.com/google-ads/announcements/9048695?authuser=0 support.google.com
  10. [10] https://support.google.com/google-ads/answer/16856494?hl=en support.google.com
  11. [11] https://support.google.com/google-ads/answer/10683687?hl=en-2 support.google.com
  12. [12] https://support.google.com/google-ads/answer/6318742?hl=en-EN support.google.com
  13. [13] https://developers.google.com/google-ads/scripts/docs/campaigns/drafts-experiments developers.google.com
  14. [14] https://hustlemarketers.com/google-ads-experiments/ hustlemarketers.com
  15. [15] https://support.google.com/google-ads/answer/6261395 support.google.com
Share the Post:

Related Posts