Google Ads experiments on low volume: 50/50 split or 80/20?

The usual advice for Google Ads experiments is a 50/50 split so you get a clean read, but with only about 30 conversions a month that test can take a quarter to tell you anything you'd trust. An 80/20 split gets you a verdict sooner, but the 20% arm gets too few clicks to show a real CPA difference, so all you really learn is that the control is fine. On low volume, I'd stop chasing statistical cleanliness and test a change big enough to move the number, like a new bid strategy or a different landing page. Run it at 50/50 for six weeks no matter what, because smaller tests mostly just produce noise. Does anyone have a rough conversion floor they use before splitting traffic at all?

It really comes down to how big a CPA change you'd actually act on. If a 20% drop would make you switch bidding, that's a large effect, and 30 conversions a month can show it in about six weeks, so 50/50 is reasonable. If you'd only act on a 10% move, no split at this volume will tell you anything useful within a quarter. In that case I'd skip the experiment, make the change, and compare before and after while watching the trend.

I'm also not sure "big enough to move the number" tells the whole story. A big change is the right instinct at low volume, but it comes with more downside too. Six weeks at 50/50 on a bid strategy swap can cost real money if it underperforms. For a landing page test I'd keep the split smaller, since the 20% arm works as a safety net.

Agreed on the 20% arm for landing pages. Bid strategy tests are different, though. Switching to a Smart bidding strategy puts the experiment arm back into learning, and at 30 conversions a month that can eat a big chunk of a six-week test. The first two weeks often look worse than the steady state, so if you call it at week three you're judging the learning period, not the strategy. Either set the read date after learning settles or drop the first two weeks from the comparison.

Fair point on the learning period, but how long you wait comes down to whether the experiment switches the bidding strategy or only changes the target. Going from manual CPC to tCPA, or from Max Conversions to a target, is a real reset. Adjusting a target on a strategy that's already running usually doesn't send the campaign back into full learning, so a week-two comparison can be fair there. At 30 conversions a month, a reset might only take a week or so to settle, and a two-week read could leave very few conversions per arm. I'd check the learning status on the experiment arm before picking a read date. If it's still learning when you planned to call it, the CPA number isn't ready to act on.

Honestly it comes down to conversion lag. If most conversions land within a day or two, a week-two read is roughly fair even after a bidding change. If your sales cycle runs a week or more, the last few days of clicks haven't finished converting yet, so both arms look worse than they will. The reset arm usually suffers more because its conversions arrive later. The conversion lag report will show you how long that window is, so hold off on reading the test until the newest clicks you're counting have had time to convert.

Honestly it comes down to how long your conversion lag is compared with the learning reset. If most conversions land within two days, a week-two read is fair even after a Smart bidding switch. If the lag runs closer to a week, the reset and the lag stack up, so a week-two read is judging a partial cohort. Either way, filter by click date and only count clicks older than the lag window. The default date view mixes in the newest days, which haven't had time to convert yet, and that tilts the comparison toward whichever arm had more recent clicks.