One ad test saw CTR jump from about 1.1% to 5.8%. Sounds like a win. But trial conversion rate dropped from roughly 22% to 14%, and cost per trial rose from about $0.32 to $0.43, according to LiftArchive. The creative that attracted more clicks attracted worse prospects. If the team had called the test on CTR alone, they'd have scaled a campaign that made pipeline more expensive.
That pattern repeats constantly in B2B SaaS, where the gap between a click and a qualified opportunity can be weeks or months. It's the core reason most Google Ads experiments produce misleading readouts: they run too short, measure the wrong thing, or declare victory at a confidence level that doesn't match the stakes.
Pick One Metric That Maps to Pipeline
Google's own Ads Decoded guidance is clear: choose a primary metric tied to a business outcome (qualified conversions, cost per qualified conversion) before launch. CTR can be a leading indicator, but it isn't a business outcome. Neither is raw conversion volume if your conversion event is a content download that never touches sales.
Before you open the Experiments tab, write the hypothesis: "If we change [one variable], then [primary metric] will move by [directional estimate] because [mechanism]." If you can't fill in that sentence, you don't have a test. You have a hope. For B2B teams with longer sales cycles, this usually means your conversion event should be as far downstream as your data pipeline allows: a qualified meeting, an opportunity created, or at minimum an MQL with a disqualifier filter applied.
The Minimum Thresholds Most Teams Ignore
Google's documentation on Demand Gen experiments spells out the math: conversion-related results require at least 100 data points to start appearing, and conversion-based bidding experiments need at least 50 conversions per arm. That second number kills most B2B tests before they begin.
If your campaign generates 30 conversions a month, a 50/50 split gives each arm 15 over four weeks. You won't reach statistical significance. Google recommends running experiments for at least four to six weeks, and longer when conversion delays are substantial (which, in B2B, they almost always are).
When volume is low, three options exist, none painless: move the conversion event upstream to increase volume but accept you're measuring a proxy; consolidate campaigns to pool volume into one experiment with enough statistical power; or extend the test to eight or twelve weeks and manage stakeholder expectations accordingly. Low volume is a constraint you plan around, not a problem you ignore.
Match Confidence to the Cost of Being Wrong
Google's Ads Decoded guidance suggests 80% confidence may be suitable for directional adjustments, while 95% is more appropriate for major structural changes. An 80% threshold means roughly a 1-in-5 chance you're acting on a false positive. For a small bid adjustment you can reverse in a day, that's a reasonable trade-off. For a campaign restructure or a shift from Search to Performance Max, a 20% false-positive rate is expensive.
Write the confidence requirement into your experiment brief before launch. Decide what "major" means for your account (anchor it to spend thresholds or pipeline impact), and don't let quarterly urgency push you into declaring significance early.
New Tooling Makes Governance Easier (If You Use It)
Performance Max asset experiments now let teams test different asset sets (seasonal versus evergreen, for example) without conflating creative changes with bidding or targeting shifts. That's a genuine improvement: isolating one variable in PMax has historically been close to impossible. Google Ads API v24.1 added direct experiment reporting, including control-arm statistics, point estimates, margins of error, and p-values. For teams running experiments across multiple accounts, this means standardized dashboards instead of eyeballing the Experiments tab in each account.
The operational implication: if your team doesn't have API access or a BI layer pulling experiment data, you're reading results through a straw. You can't audit the margin of error, compare across experiments systematically, or build institutional memory about what's been tested.
Run It This Week
Setup: Pick one campaign with enough conversion volume (aim for 50+ conversions per arm over the test window). Write a falsifiable hypothesis. Choose one primary metric aligned to a pipeline stage, not a platform proxy.
Launch: Use a 50/50 traffic split. Set the confidence threshold before you start: 80% for a reversible change, 95% for a structural one. Set a stop-loss: if CPA rises more than 25% in either arm after two weeks, pause and diagnose.
Readout: Wait four to six weeks minimum. Report the primary metric, the confidence level achieved, and at least one secondary metric (like conversion rate by stage) to catch the CTR-up-but-quality-down trap.
Next test: Whatever you learn, document it in a shared log. The experiment itself is useful; the institutional pattern recognition over ten experiments is where the compounding happens.
Most teams don't have a testing problem. They have a rigor problem. The tooling is good enough now. The question is whether you'll commit to writing the hypothesis first, setting the threshold before results come in, and waiting long enough to trust the answer you get.