What split percentage controls
When you run a campaign experiment in Google Ads, the account holds two versions of the same campaign: the original and a variant carrying your change. Split percentage is the dial that decides how much of the eligible traffic each version receives. Set it evenly and both sides collect data at a similar pace. Tilt it towards the original and the experiment runs more slowly, but less spend is exposed to an untested idea.
Google also asks how the split should be applied. Cookie-based splitting keeps a returning visitor in whichever version they first saw, so their experience stays consistent. Search-based splitting assigns every individual search independently, which fills the sample faster. The first suits a change that affects the landing page; the second suits a change that only affects the auction.
Why split percentage matters
The split governs two things at once: how quickly you can trust the result, and how much money is exposed to the change. An uneven split protects the budget but stretches the test out, and a test that runs too long starts collecting seasonal noise it was never meant to measure. An even split reaches a readable answer sooner, which is usually the safer choice on a small account where conversions arrive slowly.
It also decides whether the comparison is fair at all. Both versions must be drawn from the same auctions over the same days, and the split is what guarantees that. This is exactly what a before-and-after comparison cannot give you, because there the two periods are never truly alike.
Where split percentage goes wrong
The most common mistake is moving the dial mid-flight. Changing the split after the experiment has been running mixes two different sampling arrangements into a single data set, and the reported result no longer describes either one cleanly. If the split is wrong, end the experiment and start again rather than adjusting it in place.
The second is treating a small share for the variant as the cautious choice. It is cautious with money and reckless with the conclusion: the variant gathers so few conversions that you end up guessing, then roll out a change nobody actually proved. Smaller advertisers hit this quickly, because the auctions are thinner to begin with and the variant’s share of a thin auction is thinner still.
How to set it
Start from an even split unless the change is genuinely risky — a new bid strategy on the campaign that carries the revenue, for instance. Decide the end date before you launch, based on how long each side needs to accumulate a workable number of conversions, and then leave both the split and the experiment alone until that date arrives.
Read the result on the metric you named at the outset, not on whichever metric happens to look best when you open the report. And if the account simply cannot feed two sides, do not run a lopsided experiment as a compromise. Test something with a larger expected effect instead, or accept a weaker read and say so plainly when you report it.