Start with hypotheses tied to revenue

The experiment tool is only as good as the question you feed it, and the questions worth asking are specific and denominated in money. “Switching to broad match with target ROAS will increase conversion value 12 percent at flat efficiency” is a hypothesis. “Let’s try broad match” is a mood. Every test gets one primary metric, CPA, ROAS, or conversion value over cost, plus guardrails, a maximum acceptable CPA drift, a ROAS floor, and a stop-loss so a losing test can’t quietly burn budget while everyone waits for it to turn around.

The changes that deserve this treatment are the ones that rewire auction behavior or budget allocation, which is exactly why they’re too risky to flip live. Bidding strategy shifts, target CPA to maximize conversions, or onto target ROAS. Match type consolidation paired with Smart Bidding. Performance Max against Standard Shopping. RSA asset and pinning strategies. Audience layering, and landing page changes where the page is the variable. Small tweaks don’t need the machinery; structural bets do.

Run clean splits or don’t bother

Concurrency is the entire point. Test and control run simultaneously against the same market conditions, which neutralizes the seasonality, news cycles, and competitor behavior that make before-and-after comparisons unreadable. Use a 50/50 split for the fastest decisive read, or 80/20 when the change is risky enough that you want most traffic protected while the experiment earns trust. And hold the discipline that makes every test in this queue valid: one meaningful variable per experiment, because if two things changed and performance improved, you don’t know which one paid the bills, and you’ll reverse the wrong one someday.

Standardize everything that isn’t the variable. Conversion tracking and values identical across both arms, ad schedules, locations, devices, and budgets locked unless one of them is the thing under test, and for creative evaluations, ad rotation set so delivery optimization doesn’t quietly re-weight your variants mid-test. RSA tests specifically need matched asset counts and consistent pinning across arms, since an experiment comparing a full asset load against a thin one is measuring inventory, not messaging.

Then iterate on a system rather than inspiration. Success criteria and decision windows predefined, a couple of weeks minimum and a conversion floor per arm before anyone reads the results, winners promoted through the tool, learnings documented, and a rolling backlog of the next hypotheses so the account is always testing its next lever. The backlog discipline matters more than any single test, because one experiment teaches you a fact and a pipeline of them teaches you your market.

Point experiments at the waste first

The highest-ROI early tests are usually waste-cutters, because they pay for themselves out of budget you’re already losing. Trial broad match with Smart Bidding against your current match mix and let the split quantify whether the recaptured queries arrive at acceptable efficiency. Test a strict negative keyword strategy against the status quo, or brand separated from non-brand, or expensive geographies and hours excluded, and get the CPC and CPA delta as a measured fact before rolling anything out account-wide. Every one of these is a change people argue about in meetings, and the experiment ends the meeting.

Creative and landing alignment belong in the same program, since paying for the wrong clicks is waste with better camouflage. Ad variations handle RSA messaging tests at scale, value propositions, qualifiers that deter poor-fit traffic before the click, pinning against free rotation, and the mechanics of doing that without muddying delivery are ones we covered in the right way to split test ads without confusing Google. Pair the message test with a landing page experiment, intent filters, pricing transparency, so the page finishes the qualification the ad started. Fewer junk clicks in, more qualified conversions out, and a blended CPC that falls because the traffic got better, not cheaper.

Decide faster, with just enough statistics

A test is readable when three things align: enough volume, stable patterns, and a delta big enough to matter. Our working floors are 50 conversions per arm for CPA and ROAS decisions, or a few hundred clicks per arm if the question is genuinely about CTR, plus consistency across the segments you care about, device, weekday versus weekend, top geographies, since a “winner” that only wins on desktop Tuesdays is a segmentation insight, not a rollout.

Define the minimum lift worth detecting before launch, because it sets your patience budget. If a 5 percent improvement wouldn’t change any decision, don’t spend three weeks proving it, declare practical equivalence and move to the next test. If the readout shows 15 to 20 percent with the arms clearly separated, treat it as provisionally winning, promote it, and keep monitoring after promotion, since the cost of a rare false positive is smaller than the cost of habitually sitting on real wins.

Trust Google’s built-in readouts, then sanity-check them. Watch conversion lag so you’re not judging recent days whose conversions haven’t landed, confirm no parallel account changes contaminated the window, and in volatile accounts glance past the averages to whether the pattern holds week over week. Where budget is tight, run experiments sequentially at 70/30 splits with stricter stop-losses, trading some speed for safety. Run this whole loop consistently, sharp hypotheses, clean splits, waste-first targeting, decision rules honored, and Experiments stop being a feature you occasionally remember and become the account’s operating rhythm, which is the only reliable way an account stops reacting to its market and starts shaping its own numbers.