How a custom experiment works
Where other experiment types come with their purpose already decided, a custom experiment leaves the choice to you. You pick a Search or Display campaign as the base, create a variant of it, and change whatever you want to test — keywords, match types, bid strategy, ad copy, audiences, targeting settings. You then choose how to split the traffic, set a start and end date, and nominate the objective you care about.
Both versions run at the same time from that point on, competing in the same auctions under the same conditions. The results screen shows the control and the variant side by side for each metric, with an indication of how confident Google is that the difference between them is more than noise. When the dates run out, you apply the variant to the original campaign, keep the control, or turn the variant into a campaign of its own.
Why custom experiments matter
Some changes cannot be judged any other way. Switching bid strategy is the clearest case: apply it outright and you will never know whether the results moved because of the strategy or because the month was different. Restructuring match types, tightening audiences, or trying a very different set of ad messages all have the same problem.
The second reason is political rather than technical. A custom experiment turns a disagreement between an owner and whoever runs the account into a question with an answer. It also caps the downside — only part of the traffic meets the untested version, and the whole thing can be stopped without unpicking a rebuild.
Where custom experiments go wrong
The flexibility is the trap. Because you can change anything, it is easy to change several things at once, and a variant with new keywords, new copy and a new bid strategy tells you which bundle won but nothing you can reuse.
Two other failures are common. Small campaigns are tested as if they were large ones, and a difference that looks convincing in the first days dissolves by the third week — this catches a lot of local and specialist advertisers who simply do not have the volume. And landing page tests get pushed into Google Ads, where they do not belong: an ads experiment can only split paid traffic, while your page is also being visited by organic, direct and referral visitors who never enter the test at all. Use a proper page testing tool for that, or accept that you are measuring one slice of the audience.
How to act on it
Decide the single change, the deciding metric and the run length before you build anything, and write them somewhere other than your head. Split evenly unless the variant carries real risk, and leave both sides alone once it starts — editing mid-flight is the fastest way to waste a month.
Give any bid strategy test time to settle before reading it, since both arms restart their learning. At the end, apply, discard or promote, then record what you found. That habit compounds: a year of recorded A/B tests is worth more than a year of untested opinions, and it is the difference between an account that improves and one that just changes.