A confounder sits upstream of both treatment and outcome (a fork in graph terms), so the treated and untreated groups differ even before treatment does anything. Customer engagement confounds an opt-in program’s effect on revenue: engaged customers both enroll more and spend more anyway.
Confounding is the reason “correlation is not causation” has teeth, and closing the backdoor paths confounders open, by randomization, stratification, regression adjustment, matching, or weighting, is the central task of observational causal inference. The hard part is that confounders can be unmeasured, and no test on the data you have can confirm you caught them all.
