A/B testing in advertising is running two ad variants that differ in one deliberate way — hook, creative, offer, or audience — splitting impressions between them, and letting a pre-chosen metric (CTR, CPA, ROAS) decide the winner. It replaces opinions about what works with evidence.
The discipline is in the setup, not the dashboard. A valid ad test changes one variable (two variants differing in headline and image and audience teaches you nothing attributable), declares its success metric before launch (a variant can win CTR and lose CPA — decide which verdict counts first), and runs to adequate sample. That last one kills most "tests": deciding at 50 clicks each is reading noise. As rough guidance, conversion-judged tests want 100+ conversions per variant for confident calls; CTR-judged tests need thousands of impressions per variant. Small accounts should therefore test bigger swings — dramatic differences reach significance at realistic volumes, while "Get Started" vs "Start Now" needs traffic most accounts don't have.
What to test follows a hierarchy of leverage: angle first (which problem or desire the ad leads with — fear of waste vs hope of growth; this is where 2–10× differences live), then format/creative (UGC vs studio, video vs static), then hook (first line or first second), then elements (CTA, proof points), and offer framing throughout (free trial vs demo, % off vs € off). Testing button copy while the angle is unexamined is optimizing the doormat of an unvisited house.
Use platform experiment tools (Google Ads Experiments, Meta A/B Test) when you need clean splits — normal Meta ad rotation is not a fair test, because the delivery system assigns budget unevenly within hours based on early signals. Accept the platform's bias for quick creative reads; insist on proper splits for strategic decisions.
Cadence: one meaningful test at a time per campaign, document results, retire losers, and promote winners to control. The compounding from twenty documented tests a year is the closest thing performance marketing has to interest.
SaaS Meta test, single variable = angle. A: "Stop losing deals in spreadsheet chaos" (pain). B: "Close 30% more deals with pipeline clarity" (outcome). Same image, audience, budget, 3-week run to ~120 conversions each: A wins CPA €38 vs €52. Next test keeps A's angle and tests UGC vs static format.
Until adequate sample, with a minimum of 1–2 full weeks to absorb day-of-week effects — whichever is longer. For conversion metrics, ~100 conversions per variant is a practical confidence floor; stopping at the first encouraging gap is how accounts institutionalize noise.
Default delivery optimizes, it doesn't experiment: early engagement signals steer most budget to one variant within hours. For fair splits use Meta's A/B Test tool (or Google Ads Experiments), which hold the split fixed for the test's duration.