# Control group: definition, formula and examples

Source: https://leverprice.com/glossary/control-group
Updated: 2026-11-17

> A control group is the set of visitors who keep seeing the current version during a test, so every variant has a baseline measured at the same time. In a price test, the control is the price you charge today. A concurrent control cancels out seasonality and traffic swings that ruin before-and-after comparisons.

The control group is your "nothing changed" baseline. If you charge $29 today and want to try $35 and $39, the visitors who keep seeing $29 during the test are the control. Their results show what would have happened this week anyway, so the difference between them and the other groups can be credited to the price.

## How to calculate it

```text
lift        = (variant metric − control metric) ÷ control metric
split check = Σ (observed − expected)² ÷ expected      (a chi-square test)
```

Example: over the same two weeks, the $29 control earns $0.70 per visitor and $39 earns $0.77. The lift is ($0.77 − $0.70) ÷ $0.70 = 10%.

Also check that each group got the traffic you intended. A 50/50 split of 10,000 visitors should land near 5,000 each. If you see 5,300 and 4,700, the chi-square value is 300² ÷ 5,000 × 2 = 36, which chance alone would produce about twice in a billion tests. This is a sample ratio mismatch: something in the assignment is broken, such as bots landing in one group or a redirect that drops visitors, and the results can't be trusted until it's fixed.

## Why it matters for pricing

A common way to judge a price change is before and after: the price went up in January, so compare January with December. That comparison mixes the price with everything else that changed between the two months, such as traffic sources and seasonal demand. A control group running at the same time sees the same week as the variants, so those swings cancel out.

How you size it matters too. For two groups, the variance of the comparison scales with 1 ÷ n₁ + 1 ÷ n₂, so an even split is the most precise. Keeping only 10% of visitors on the control to limit risk gives the comparison 2.8 times the variance of a 50/50 split on the same traffic. If the goal is to limit how many visitors see a worse price, a [multi-armed bandit](https://leverprice.com/glossary/multi-armed-bandit) does that while still learning.

## Common mistakes

- Comparing with last month instead of a control running in parallel.
- Changing the control's page while the test runs, so the baseline moves with it.
- Ignoring a lopsided split. A sample ratio mismatch usually means a tracking or assignment bug, and luck is the least likely explanation.
- Mixing existing customers into the groups. A price test should enroll new visitors, while current subscribers keep the plan they bought.

## Related terms

The control anchors every [price test](https://leverprice.com/glossary/price-testing), and [sticky assignment](https://leverprice.com/glossary/sticky-assignment) keeps each visitor in one group. [Exposure](https://leverprice.com/glossary/exposure) defines who is counted in each group, and [statistical significance](https://leverprice.com/glossary/statistical-significance) says whether the gap with the control is real.

## Frequently asked questions

### Does a multi-armed bandit have a control group?

It usually keeps your current price as one of the arms, and that arm plays the role of the control, but its share of traffic changes as the bandit learns. Lift against the current price is still the number to report. The differences are covered in [bandit vs A/B test for pricing](https://leverprice.com/blog/bandit-vs-ab-test-pricing).

### How big should the control group be?

In a two-way test, the same size as the variant. With several variants, give the control somewhat more traffic than each variant, because every comparison runs through it.
