Peak-End Rule

011 min

The peak-end rule at a glance

• What it is: Later ratings follow the best or worst moment and the last moment.
• Origin: Kahneman, Fredrickson, Schreiber and Redelmeier tested it in 1993 with cold water.
• Where it bites: Satisfaction scores, support closings and onboarding flows are read from memory.
• Guard against it: Collect ratings during an experience as well as after it.

021 min

Where the peak-end rule shows up

A support lead (illustrative) reads two customer surveys. Both customers had the same problem, it took the same time to fix, and in both cases the agent was polite throughout. The first customer scored the service high. The second scored it low.

The only difference is the last step. The first customer ended with a message that said what was fixed and what to expect next. The second customer ended with a refund form that rejected the input twice. The middle part of both cases was identical. If the second customer had received the closing message instead, their score would likely have been higher.

The pattern is general. When people are asked about an experience afterward, they do not replay it from start to finish. They answer from a few moments that stand out, and the final moment is almost always one of them.

032 min

Why the peak-end rule happens

Fredrickson and Kahneman proposed two ways the mind could store an emotional episode. One is like a film, where every moment is kept and the total can be added up. The other is like a collection of snapshots, where only a few selected moments are kept. Their evidence favoured the snapshots.

In the snapshot model, the overall judgment is a weighted average of those few moments, and time is not represented in it. Kahneman and colleagues proposed that people average the peak and the end of an experience instead of adding up all its moments.

This conflicts with a rule called temporal monotonicity. The rule says that adding pain to the end of an episode can only make the episode worse. People who judge by peak and end break this rule, because a mild extra period at the end lowers the average.

Why do the worst and the last moments get picked? Kahneman's account is that they are prototypical examples of the episode, so they are the most salient, which means they stand out most. A practical reason is also plausible: reconstructing a whole episode in memory is hard, and two moments are easy to recall. This explanation is a proposal, and the next sections show that it does not hold in every setting.

Memory seems to matter. In the second of two film studies, people who ranked the clips from memory showed an even smaller effect of duration than people who judged them as they watched.

042 min

Where the peak-end rule comes from

The first result came from Varey and Kahneman in 1992. They showed people lists of discomfort ratings, given on a scale from 0 to 10 at five-minute intervals, as if another person had rated an unpleasant event. An unweighted combination of the peak rating and the final rating explained 94% of the variance in the global evaluations. The series 2-5-8-4 was judged much less aversive than the series 2-5-8, even though the only difference was five extra minutes of discomfort.

In 1993, Fredrickson and Kahneman moved from lists to real experience. They had 32 people watch aversive film clips and pleasant film clips that varied in duration and intensity. The viewers rated their feelings during each clip and again at the end. The later rating followed a weighted average of the peak and final feelings, and the length of the clip added little.

The study most often cited is the cold-water experiment by Kahneman, Fredrickson, Schreiber and Redelmeier, also published in 1993. Thirty-two male university students put one hand in water at 14 °C for 60 seconds in a short trial. In a long trial they put the other hand in the same water for 60 seconds and then kept it there for 30 more seconds, while the water warmed slightly to about 15 °C. That was still uncomfortable, but less than before. The long trial contained all of the discomfort of the short trial plus an extra, easier period.

The students were then asked which trial they would repeat. Most of them, 22 of 32, chose the long one. The authors described the result as people preferring more pain to less. A replication that removed the real-time discomfort ratings found the same preference.

062 min

How the peak-end rule shows up in product, design and support

Product and design. Laws of UX, a design reference site, gives two real cases. It says Mailchimp treats the moment a first-time user finishes sending a campaign as important and does more than show a plain confirmation window. It adds illustration, animation and humor. The same page says Uber reduced its post-request cancellation rate by working on how people perceive time and waiting, so that waiting did not become a negative peak. These are the site's descriptions, not experiments, so treat them as design practice and not as proof.

An onboarding flow (illustrative). A designer has time to polish one part of a ten-step setup. The rule says to look for the step that causes the most frustration and for the last step. A smooth middle matters less for memory than a bad peak or a rough ending.

Survey timing (illustrative). A PM sends a satisfaction survey right after a checkout. The score then describes the checkout step more than the whole purchase journey. An analyst reading the scores should know which moment the survey followed.

Assistants and AI evaluation (illustrative). A team asks users to rate an AI assistant with one thumbs-up button at the end of each conversation. The rating mostly reflects how the conversation ended and its worst moment. If the team wants to find the cause of a low score, it needs the rating for each turn, not only the final one.

In all four cases, the effect is usually suffered by accident. It becomes a design choice only when a team plans the ending on purpose.

071 min

How to apply the peak-end rule

1. 1

   Map the experience as a sequence of steps.

   Mark the step with the strongest emotion, good or bad, and the final step.

2. 2

   Fix the worst peak first.

   No polish in the middle outweighs one step that causes strong frustration.

3. 3

   Design the ending on purpose.

   End on a clear confirmation, an easy exit or a small positive result.

4. 4

   Measure twice.

   Ask for a rating during the experience and another afterward. The gap between them shows what memory changed.

5. 5

   Check behaviour as well as ratings.

   The colonoscopy trial measured who returned, because memory and later choice are what the rule predicts.

6. 6

   Improve the ending by making it easier, not longer.

   Do not add real discomfort to improve a memory.

The sixth point is an ethical limit. The colonoscopy authors wrote that physicians might use special gentleness and care at the end of an aversive procedure on an awake patient. They also named the tension between the fallibility of human reasoning and the patient's autonomy. Laws of UX gives the design version of step 3: pay close attention to the most intense points and the final moments of the user journey.

082 min

Where the evidence on the peak-end rule is contested

The rule is well supported for remembered pain, but it is not a law, and three groups of findings limit it.

The first lab demonstration may owe part of its result to format. Liersch and McKenzie pointed out that memory was not involved in the original demonstration of duration neglect. People there judged another person's experience, shown as a list of numbers. In three experiments, graphs rather than number lists made duration neglect disappear. Their conclusion was that when people do not have to remember the experience, they use all the moments. They advised caution about generalising from duration neglect. This objection is aimed at the 1992 list task. It does not address the cold-water or colonoscopy results, where memory was involved.

Duration neglect depends on the design. Kahneman and Frederick wrote that duration neglect almost always appears when people cannot compare durations directly, and almost never when they can. The colonoscopy authors added that duration has a small influence unless it is highly salient, such as the wait for surgery, or correlated with intensity, such as the length of labor.

The effect on behaviour was small. The colonoscopy authors called their study a proof of concept. Return rates were only slightly higher in the extended group, and they said that more effective interventions are needed.

For a practitioner, this means the following. Treat the peak and the end as strong candidates for what customers will remember. Do not assume that duration never matters, and do not assume a small change at the end will reverse behaviour.

091 min

Common misunderstandings about the peak-end rule

"Only the ending matters." The finding is about a weighted average of two moments. In the film study, retrospective evaluations were well predicted by a weighted average of the peak affect rating and the final rating. Both count.

"The peak means the worst moment." The films in that study ran from pleasant, such as penguins at play, to highly aversive, such as an amputation. The peak can be a high point as well as a low point.

"Duration does not count at all." The 1992 study found the effect of duration small, not zero. The authors wrote that it was statistically significant but remarkably small.

"The rule describes the experience." It describes the memory. A person who lives through a bad middle part still lived through it, and may leave when a better option exists, whatever their final rating says.

101 min

The peak-end rule vs. nearby concepts

The shortest test is to ask what is being judged. If it is recall of list items, the serial position effects apply. If it is the overall judgment of one lived episode, the peak-end rule applies.

A related effect works in the other direction. The goal gradient effect describes effort that rises near a finish line, which is behaviour during the experience and not memory afterward.

?6 questions

Questions people ask

§5 sources

Sources