What to check before you touch anything
- Is there enough total traffic/conversion volume to split it into two halves (test and control) and still get a statistically meaningful read from each? An experiment needs roughly twice the data to reach the same confidence level as a direct change.
- How much does it matter to the business not to lose volume/efficiency during the test period? A Campaign Experiment limits the risk to a slice of traffic, while a direct target change affects the whole campaign right away.
- Is this a small, routine tweak (a step of 10–15%) or a fundamentally different strategy/target? Experiments are often overkill for small steps and worthwhile for bigger, riskier calls.
- Is there enough time budgeted for the test itself? An experiment needs to run at least as long as it takes to exit "Learning" plus a stable-performance period afterward, or the comparison won't be valid.
- Is it technically feasible to split traffic cleanly (Campaign Drafts/Experiments don't support every campaign type the same way — Performance Max in particular has quirks)? Worth confirming the experiment format you want is actually available for this campaign.
Possible approaches
- For bigger or riskier changes (a significant tightening/loosening of the target, or switching bid strategies entirely), many agencies prefer a Campaign Experiment — it lets you compare the new and old setup side by side under the same external conditions (seasonality, competition) before rolling the change out to the whole campaign.
- For small, incremental adjustments (the standard 10–15% target change), a direct change followed by monitoring is more common — setting up a full experiment for a small step often isn't worth the extra time and effort.
- Google recommends using Experiments specifically when you need confidence comparing two fundamentally different approaches (a different target CPA, or an entirely different bid strategy) and when your volume is large enough to split traffic without making either half statistically unreliable.
- If the campaign is small and splitting it into test/control would make both halves unreliable, it can make more sense to skip a formal experiment and instead test the change sequentially over time (before/after), understanding the limits of that comparison (seasonality and other outside factors can distort it).
- In larger accounts with several similar campaigns, an alternative is to test the new target on one of several similar campaigns while leaving the others as an informal control group, rather than splitting traffic within a single campaign.