Twenty-five students in a tutoring program raise their test scores from 64 to 76, a 12-point jump. Impressive, until you ask what would have happened without tutoring. They were in school all term, taking the same classes and practising for the same test. A control group of 25 similar students who did not get tutoring went from 65 to 73, up 8 points. The tutoring's effect is the extra improvement: 12−8=4 points.
A control group is the experiment's answer to "compared with what?" It must be like the treatment group in every way except the treatment: the same kind of people, the same time period, the same measurements. The best way to get that is random assignment, so that both groups come from one pool split by chance. Controls do not always get nothing. In medicine they often get a placebo, a dummy pill that looks real, or the current standard treatment, so the question becomes whether the new treatment beats what patients would otherwise receive.
An untreated group is not automatically a good control. If students at one school try a new toothpaste and students at another school serve as the control, the schools may differ in water fluoridation, diet or dental visits, and those differences travel into the comparison. The cavity gap then mixes the toothpaste with everything else that differs between the schools. A control that differs systematically from the treatment group reintroduces confounding.
Think of it like this. A control group works like a twin who skipped the treatment: whatever happens to that twin, from growth to practice to recovery, would also have happened to the treated twin, so the difference between them is what the treatment added. Real groups are twins only on average, which is why they are formed by chance.
Core idea. A treatment's effect is measured against a comparable group that did not get it: subtract the control group's change from the treatment group's change, and what remains is what the treatment added.