Every experiment needs a comparison, and the control group provides it.
Without a control, any change in the results could come from the season, the news or a competitor's sale.
The control experiences everything the treatment does, except the change itself.
That is why both groups must run over exactly the same period, not one after the other.
The effect is the treatment result minus the control result, such as 11% conversion against 10%.
It is often reported as a relative lift as well, which here is 10% higher than control.
More treatment groups can test several ideas at once, such as 3 different prices against the current one.
Each extra group needs its own share of traffic, so tests with many groups take longer.
Each extra comparison is also another chance for a false positive, so the threshold may need adjusting.
Users in the control must not be exposed to the change by accident, such as through a shared link or a stored copy of a page.
Exposure can leak the other way too, and both kinds of leak shrink the measured difference towards 0.
Groups usually get equal shares of traffic, which gives the most precise comparison for a given total.
A small control, such as 10%, is sometimes used when the change is expected to be safe and the team wants it out widely.
In practice the control group's results are checked first, because a control that behaves oddly signals a broken test.
The usual mistake is comparing the treatment with last month's numbers instead of with the control from the same days.