Streak effect

A streak is a count of consecutive periods in which the user did the thing, shown back to them. The only randomised field test of it measured notification copy rather than a counter. Aulagnon, Cristia, Cueto and Malamud (2025) split 60,000 Peruvian fourth- to sixth-graders on a maths platform into four arms for six summer weeks: messages highlighting their streak of completed weekly assignment sets, personalised reminders with positive reinforcement, generic reminders, and nothing. The bases were low: 5.3% of controls opened the platform at all, and those who did connected in 31.7% of weeks. Streak messages raised connecting at least once by 2.8 percentage points, behind personalised reminders at 3.8 and ahead of generic ones at 1.4, and raised the share of weeks connected by 9.4 points, ahead of 6.9 and 2.5. Learning barely moved: among the 2.3% who sat the endline test the streak arm scored 0.13 to 0.17 standard deviations above control, with no arm differing from another. Effects landed on the days the messages went out. The streak framing won depth of return and lost reach, and it counted weeks, not days.

The causal core is one set of seven lab studies, and its finding is that the display does the work. In Silverman and Barasch's (2023) Study 2, participants all completed the same four strength exercises in an app; those whose fourth log succeeded saw an intact run of four, those told the log had failed to load saw an “X”. Asked to choose their next activity, 66.23% of the intact group stayed with strength work against 57.86% of the broken group — identical behaviour, different counter, different next action. Study 3 crossed intact or broken with whether a log was shown at all, the condition design turns on. Highlighting an intact streak raised continuation to 92.47% from 64.94% unlogged; highlighting a broken one cut it to 45.21% from 60.90% unlogged. The counter does not fail safe. It points wherever the last period went.

The no-log comparisons are the weak half of the record, and the authors say so. Study 4 ran intact 65.84%, no log 54.73%, broken 47.98%, with intact beating no log but broken only directionally below it; Study 5 ran 82.55%, 77.78% and 67.92%, with the broken-versus-no-log contrast marginal and the intact one not significant. Neither contrast was powered, and which half reached significance flipped between the two. The mechanism tests are more consistent: sense of accomplishment mediated the effect (indirect effect .49, 95% CI [.25, .82]) and negative emotion did not (.11, 95% CI [−.07, .32]), even though the broken group reported far more of it. Felt progress moved behaviour, not fear of loss, which makes loss-framed streak copy the least supported thing a team usually builds first.

Two findings about breaks are the practical payload. In Study 6, blame changed the size of the drop: 53.17% continued with an intact streak, 42.00% when the break was attributable to the app's quota message, and 28.87% when participants had chosen the behaviour that broke it. In Study 7, offering repair recovered most of the damage — intact 93.14%, repairable break 85.20%, plain break 68.66%. The field messages behaved accordingly: the cell for a student who had not connected said they could reach three consecutive weeks today, never that they had lost anything, and runs of consecutive non-use were no longer in the streak arm than in the reminder arm. Willingness to protect a run is also its commercial risk — 45% of Study 6 participants would watch an advertisement to keep a streak and 43% to repair one, and Duolingo already sells a freeze.

Two popular beliefs have no support. Length is not a dial: three logged behaviours against six, and four against twenty, both reproduced the intact-versus-broken effect with no interaction with length, so the goal looks roughly categorical over the range tested and nothing is known above twenty. Making the log public added nothing measurable — 55.78% against 43.85% overall, with no interaction with whether the log would be shared. And the figure product write-ups quote most means least. Silverman and Barasch's own field study of 980 university employees in a 30-day 7,000-step challenge found that an intact run predicted meeting the goal that day even controlling for each person's overall completion rate, but it is correlational, and they say plainly it cannot separate the streak from whatever made certain people both likely to hold one and likely to keep using the app. A chart showing long-streak users retaining better is that confound.

The habit story does not survive contact with the habit literature. Lally, van Jaarsveld, Potts and Wardle (2010) had 96 volunteers repeat a self-chosen daily behaviour in a fixed context for 84 days; an asymptotic curve fitted 62 of them and fitted well for 39, with a median 66 days to 95% of the automaticity plateau and a range of 18 to 254. Across 140 missed days automaticity fell by 0.29 of a point on the self-report habit index, and the gain across an omission (.55) was indistinguishable from the gain across three consecutive days (.79), with no longer-term cost. One missed day costs almost nothing behaviourally while an interface that resets to zero charges the user everything; Khan Academy withdrew its streak in January 2021 on that reasoning, saying that as designed streaks can be demotivating when something outside the learner's control breaks them. Attached to money the mechanism is better evidenced: Mehr, Silverman, Sharif, Barasch and Milkman (2025) report across six preregistered studies (N = 4,493) that people completed more work for streak incentives — rising payouts for consecutive tasks — than for larger flat ones, and that the consecutiveness requirement rather than the escalating amount carried it, which buys the lift and the overjustification problem together.

The honest state of it: no published randomised test of an in-app streak counter exists, and none of a streak freeze. One field experiment tested streak-framed notifications, the causal studies are single-session choices on Mechanical Turk, and the one field test of streak feedback inside a consumer app — Hillegass, Silverman, Gershon, Riedel, Grüning and Wolf's working paper on a self-control app — reports in its seminar abstract that the streak became a competing goal and social-media use went up, which we could not verify in the paper itself. Duolingo's widely repeated figures are company-reported with no method published and should not be cited as evidence. Evidence of harm is correlational: among 2,483 Belgian early adolescents, van Essen and Van Ouytsel (2023) found problematic smartphone use associated with maintaining Snapchat streaks, with weak correlations between fear of missing out, self-control and the number of streak partners and days; Hristova and colleagues (2022) interviewed 25 pupils aged 14 to 18 and found maintenance becoming an obligation negotiated between friends. Regulators have begun to classify the family: the Spanish data protection authority's 2024 guide lists streaks and “perfect week” rewards among addictive patterns, and the European Parliament's resolution of 12 December 2023 asks the Commission to turn attention-seeking features off by design. Build the counter, build the repair with it, and treat anything you read about long streaks as a claim about the kind of person who has one.