Alternative Hypothesis
The alternative hypothesis is the claim a hypothesis test is actually designed to find evidence for — typically that a real effect, difference, or relationship exists, directly opposing the null hypothesis's default assumption of no effect. Where the null hypothesis is the conservative default a test tries to argue against, the alternative hypothesis represents the researcher's actual substantive claim of interest, and rejecting the null hypothesis with sufficiently strong evidence is what allows a test to conclude in favor of the alternative instead.
The alternative hypothesis can be formulated in two structurally different ways that change which specific statistical test and calculation apply: a two-tailed (or two-sided) alternative hypothesis simply claims a difference exists in either direction, without specifying which way (the true effect could be positive or negative), while a one-tailed (or one-sided) alternative hypothesis makes a more specific directional claim (the new treatment is better than the control, specifically, not just different from it in some unspecified direction). This choice has to be made before looking at the data, as part of the study's design, rather than selected afterward based on which direction the results happened to point — choosing a one-tailed test after already seeing a result that happens to lean one direction is a well-documented form of statistical malpractice (a variant of p-hacking) specifically because a one-tailed test uses a more lenient p-value threshold in the chosen direction, artificially inflating the odds of a statistically significant result if the direction was selected to match the data rather than fixed in advance.
A statistically significant result (rejecting the null hypothesis) supports the alternative hypothesis but does not, by itself, confirm precisely why the alternative hypothesis is true or guarantee the effect is practically meaningful — statistical significance establishes only that an observed effect is unlikely to be pure chance, a separate question from effect size (how large or practically important the effect actually is), which is why a statistically significant result with a tiny, practically negligible effect size is a common, specifically flagged pitfall in applied statistics: with a large enough sample, even a trivially small, practically unimportant effect can become statistically significant, since statistical significance is sensitive to sample size in a way practical importance simply isn't. This gap between statistical and practical significance is exactly why reporting an effect size alongside a p-value, rather than a p-value in isolation, has become increasingly standard and recommended practice across applied research.
The alternative hypothesis in applied business contexts — most commonly in A/B testing — is typically the specific, practically motivated claim actually driving the experiment: "the new checkout flow increases conversion rate compared to the current one," with the null hypothesis serving as the conservative, skeptical default that has to be convincingly overcome by the collected data before the business actually changes anything based on the result, a framing that makes explicit why hypothesis testing's inherent conservatism (requiring strong evidence before rejecting the null) is a deliberate safeguard against making costly changes based on effects that turn out, under more rigorous scrutiny, to have been nothing more than random noise.
Read the alternative hypothesis as the substantive claim a hypothesis test exists to evaluate evidence for, always defined in direct opposition to the null hypothesis's conservative default: whether it's framed as two-tailed or one-tailed is a consequential design decision that must be locked in before data collection, not adjusted afterward to match whatever direction the results happen to show, and a rejected null hypothesis supporting the alternative establishes only that an effect is statistically unlikely to be chance, a genuinely different and narrower claim than establishing that the effect is large enough to actually matter in practice.