Peak-end rule

The peak-end rule describes how memory summarizes experiences. When people look back on an episode, their judgment is dominated by how it felt at its most intense point and how it felt at the end. How long it lasted counts for surprisingly little, a pattern called duration neglect. The rule is most closely associated with Daniel Kahneman and colleagues.

Why it matters

The best-known evidence is a 1993 study by Kahneman, Fredrickson, Schreiber and Redelmeier, When More Pain Is Preferred to Less. Participants held one hand in 14°C water for 60 seconds. In another trial they held a hand in 14°C water for 60 seconds and then kept it there for 30 more seconds while the temperature was raised slightly, to about 15°C. When asked which trial to repeat, 22 of 32 participants (69%) chose the longer one, even though it contained all the pain of the short trial plus more. The milder ending made it remembered as less bad.

For product design this means people's memory of using your product, and their likelihood of coming back or recommending it, is shaped by a few moments. A smooth flow with one painful step can be remembered for that step. A frustrating ending can spoil an otherwise fine session.

How to apply it

• Do find the negative peaks with research (support tickets, session recordings, usability tests) and fix those first.
• Do design endings deliberately: a clear confirmation after a purchase, a useful summary after a task, a calm sign-off after cancellation.
• Do create a few positive peaks where they are earned, such as a helpful success state when someone finishes a hard task.
• Do make error states explain what happened and what to do next, since errors are often the peak people remember.
• Don't end experiences with friction: aggressive exit pop-ups, surprise fees at the last step, or a cancellation flow full of obstacles.
• Don't assume shortening a flow fixes it. Duration matters less to memory than the worst moment and the finish.

Common mistakes

• Using the rule to excuse a bad middle. Peaks include negative ones, and a bad step can become the peak.
• Faking positive moments. Confetti on a routine action is not a peak and can feel patronizing.
• Quoting the cold-water result loosely. The paper reports 69% (22 of 32) preferring the longer trial. Some summaries say 80%, which matches only the subgroup who felt a clear drop in discomfort at the end (17 of 21).

In AI products

AI features fail in visible ways: a confidently wrong answer, an agent that books the wrong date, a summary that misses the key decision. These moments tend to become the negative peak of the whole experience. How the product handles them shapes what people remember.

Google's PAIR Guidebook advises teams to "Address mistakes with humanity and humility, and explain the system's limits while inviting people to continue forward." In practice: admit the error plainly, make undo or correction easy, offer a way for the user to take over, and end the interaction with the task back on track. A clear recovery path can turn a failure into a tolerable memory rather than the reason someone stops using the feature.