maxlevel
ContactLog inStart free
BlogRuleset design

Streaks work until somebody misses a day

A streak is the cheapest retention mechanic you can run, and the day it breaks is where the real design work sits. What the evidence says about who gets the blame, how long to make the window, and when to let people start again.

2 August 202611 min read
A row of eighteen rounded tiles on a dark umber background lit by an ember glow. The first eleven brighten from dim brown to hot orange, then one tile is empty with a dashed outline, and the tiles after it start dim again. A line beneath the empty tile reads: the day it breaks.

A streak is the cheapest mechanic you have

Streaks cost almost nothing to run. Count the consecutive periods in which someone did the thing, show the number, stop. There's no economy to balance and no reward store to keep stocked. The evidence behind them is also better than most of what gets sold as gamification.

In the Peruvian summer of 2022, Aulagnon, Cristia, Cueto and Malamud randomised 60,000 primary school children across four message conditions on Conecta Ideas, an online maths platform. One group got messages highlighting their streak of consecutive weeks connected. The others got personalised reminders, generic reminders, or nothing at all. The streak group connected in 9.4 percentage points more weeks than the control group and scored 0.13 to 0.17 standard deviations higher on a maths test. It was the only treatment that produced consistent learning gains.

The two margins moved by very different amounts. Highlighting streaks lifted the likelihood of connecting at least once by 2.8 points, and the fraction of weeks connected by 9.4. It kept the children who had already started. Personalised reminders were the better tool for getting non-users through the door, which stands to reason, because you can't have a streak of one.

STREAK MESSAGES VS CONTROL, PERCENTAGE POINTS Fraction of weeks connected They kept going +9.4 pp Connected at least once They turned up at all +2.8 pp Maths scores in the streak group ran 0.13 to 0.17 standard deviations above control. 60,000 STUDENTS, PERU, SIX WEEKS, 2022
Both effects are real, and one is more than three times the other. Streaks held on to children who were already using the platform.

A streak is a retention mechanic. If your problem is that half the sales floor has never logged anything, a counter of consecutive days will do almost nothing for you.

The break is the whole design problem

Every streak ends. Yours will end on a Tuesday, when somebody's train is cancelled.

Silverman and Barasch tracked what happens next in On or off track, published in the Journal of Consumer Research in 2023. Broken streaks reduce what people do afterwards compared with intact ones. How much depends on blame. When someone reads the break as their own doing, the drop is worse. When they can pin it on circumstances outside their control, it's softer. It softens again if they get a chance to repair the break.

That turns a display question into a design question. Your ruleset decides, silently, which of those stories a missed day tells.

WHO GETS THE BLAME FOR THE MISSED DAY The streak breaks They read it as their own fault Largest drop in what follows They can blame circumstances Softer drop They can repair the break Softer still
Directions reported by Silverman and Barasch. The branches are ordered by severity; no magnitudes are claimed here, because the paper reports the comparison rather than a single headline number.

A counter that drops to zero at midnight with no explanation and no route back tells the harshest version available. So decide what a break costs before you turn anything on. Lengthening the period is the blunt fix: a weekly count absorbs a bad Tuesday that a daily count would punish. A repair route is the better one, because doing the thing twice to cover a gap hands the person back a version of events where the miss wasn't fatal.

Match the window to the behaviour

Habits don't all run on one clock. Buyalskaya and colleagues applied machine learning to two large panels, over 12 million gym visits and over 40 million hospital handwashing events, and found that gym habits took months to form while handwashing habits took weeks. There's no single magic number of days.

Your streak period is therefore a claim about how often the behaviour naturally happens. Get it wrong and you manufacture failures that have nothing to do with the person.

Take a household rota where the recycling goes out twice a week. Put a daily streak on it and the counter is broken on ten days out of fourteen while nobody has done anything wrong. Put a weekly streak on the same fortnight and it reads as two clean weeks. Same behaviour, same data. The rule invented the failures.

SAME FORTNIGHT, SAME CHORE, TWO STREAK RULES Daily streak rule 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Counter resets on ten of the fourteen days. Weekly streak rule Week 1 Week 2 Both weeks intact. ILLUSTRATIVE ROTA: THE CHORE HAPPENS ON DAYS 2, 5, 9 AND 12
An illustrative rota, not measured data. The behaviour is identical in both rows. Only the rule changes, and with it ten days of manufactured failure.

The test is easy enough. Write down how often you actually expect the action, then set the period one notch looser than that. If people recieve a nudge every morning about something they do on Tuesdays, you've built an annoyance with a number attached.

Give people a date to start again

Once a streak is properly dead, the question becomes when to restart it. Timing matters more here than you'd think.

Dai, Milkman and Riis called this the fresh start effect, in Management Science in 2014. Across more than a year of university gym records they found the probability of a visit rising after temporal landmarks: by 33.4% at the start of a new week, 14.4% for a new month, 11.6% for a new year, 47.1% for a new semester, 24.3% following a school break, and 7.5% after a birthday other than the twenty-first.

INCREASE IN THE PROBABILITY OF A GYM VISIT A birthday +7.5% A new year +11.6% A new month +14.4% After a school break +24.3% A new week +33.4% A new semester +47.1% University gym records, September 2010 to December 2011. Birthdays exclude the twenty-first.
Bar length is the reported percentage increase. A new semester is worth roughly six times a birthday.

Dates that mean something make people willing to try again. A restart on a random Thursday in March is just a reminder of the streak you lost.

This is what seasons are for, and seasons aren't built in maxlevel yet. We want them shipped soon, because a season boundary answers both problems in this post at once. It resets streaks on a date everybody knew about in advance, and it's the moment to change values without anyone reading it as personal. Until then you can run one by hand. Pick the end date before you launch, tell everybody what it is, and hold to it.

What you're actually deciding

Turning streaks on takes a minute. The decisions around them are the work: what counts as a period, how many misses it tolerates, whether a break can be repaired, and when everybody starts again from zero together. Get those wrong and you've built a machine for telling committed people they've failed.

None of it shows up in the counter. It's a number going up, right until the Tuesday it isn't.

Sources

  • Aulagnon, R., Cristia, J., Cueto, S. and Malamud, O. (2024). Streaking to success: the effects of highlighting streaks on student effort and achievement. Inter-American Development Bank. PDF
  • Silverman, J. and Barasch, A. (2023). On or off track: how (broken) streaks affect consumer decisions. Journal of Consumer Research, 49(6), 1095–1117. Article
  • Buyalskaya, A., Ho, H., Milkman, K., Li, X., Duckworth, A. and Camerer, C. (2023). What can machine learning teach us about habit formation? Evidence from exercise and hygiene. PNAS, 120(17). Record
  • Dai, H., Milkman, K. and Riis, J. (2014). The fresh start effect: temporal landmarks motivate aspirational behavior. Management Science, 60(10), 2563–2582. PDF
← All posts