Streak mechanics for B2B retention: where the reward goes

Streak mechanics port to B2B, but you have to change the unit. In a consumer app the streak is a person and a day. In a seat-based product the streak is an account and a billing period, and the counter that matters is usage depth — seats active, workflows run, records touched — instead of consecutive logins. The three mechanics underneath are the same: loss aversion makes an accumulated position expensive to abandon, endowed progress makes a partly-filled counter more motivating than an empty one, and the hazard curve tells you when the account is actually at risk. The hazard curve is the one that changes a budget. Cost per saved account falls with every month of tenure, so the cheapest save is always the latest one — but the month that saves the most accounts under a fixed budget sits in the middle, and that crossover is what the rest of this post locates.

Three mechanics, one curve

Loss aversion

Tversky and Kahneman's reference-dependent model (Quarterly Journal of Economics, 1991) states the assumption plainly: losses and disadvantages weigh more heavily on preference than equivalent gains. A streak counter works because it converts continued use into an owned position. Skipping a period stops being a neutral non-event and becomes a loss of something already held.

In B2B this does not attach to a person. It attaches to whatever the buyer would have to explain away at renewal: fourteen months of pipeline history, four integrations, the report the CFO now expects on the first Monday. Those are the account's streak. Nobody logs in to protect them, and that is exactly why login streaks measure the wrong thing.

Endowed progress

Nunes and Drèze, The Endowed Progress Effect (Journal of Consumer Research, 2006), found that people given artificial advancement toward a goal persist harder and finish faster. The abstract describes converting an eight-step task into a ten-step task with two steps pre-completed. Loyalty & Reward Co's write-up of the car wash field experiment reports 34% redemption for the pre-stamped group against 19% for the eight-stamp group — I am citing that secondary source because the JCR abstract itself does not carry the figures. The B2B version is the onboarding checklist that starts at 3 of 12 because you counted the contract, the SSO setup and the first invite. The arithmetic of that is in the endowed progress post.

The hazard curve

The hazard rate is the probability an account dies in the next period given that it survived to now. It is rarely flat. The shape I assume throughout this post — high in the first period, falling toward a floor — is an assumption about your cohort, not a fact I can cite about subscription products in general, and it is the single input that decides everything downstream. Fit it before you trust it. Under that shape, every period an account survives makes it a better bet, and makes each remaining at-risk account both scarcer and more valuable to save. I derive the shape in streak mechanics and the hazard rate.

Why "daily login" is the wrong unit in B2B

A seat-based product is not used daily by any one person on purpose. A payroll tool is used twice a month and is not in trouble. Duolingo's Q2 FY26 shareholder letter reports 58.7 million DAUs against 140.6 million MAUs and a Current User Retention Rate of 84%, an all-time high. Read that metric carefully: Duolingo defines CURR as the share of current-user DAUs who come back the next day, and a current-user DAU is one who was also active another time in the previous seven days. It is next-day return among the already-habituated, not next-day retention across everyone. That subset is the environment a daily streak is built for. If your product's honest cadence is fortnightly, a daily counter manufactures failures the user cannot control, which is the fastest way to teach someone the counter is meaningless.

So the account streak is a run of periods containing a qualifying event, and you choose the period and the event from data you already have. The qualifying event has to correlate with renewal, and you have to check that correlation before you incentivise it — otherwise you have built a machine for buying activity you did not want.

That same letter has the other half of the lesson. Duolingo ran a one-time Streak Revival event in June 2026: 15.4 million learners restored a lost streak, and nearly 8 million of them had no active streak when the event began. Restoring a destroyed position was worth building a campaign around, which tells you what the position was worth in the first place.

The hazard model, in accounts

The continuation curve

Model the probability that an account alive in period t survives into t+1 as a curve climbing from a rough first period toward an asymptote:

p(t) = p_inf - (p_inf - p1) * exp(-k * (t - 1))

p1 is first-period continuation, p_inf is where renewal settles once the account is embedded, and k is how fast it gets there. Three numbers, fittable from one real cohort.

The code

Python standard library only. Every input below is an assumption I have labelled as one, not a benchmark.

from math import exp

def p_cont(t, p1, p_inf, k):
    """Probability an account alive at period t is still alive at t+1."""
    return p_inf - (p_inf - p1) * exp(-k * (t - 1))

def save_curve(cohort, p1, p_inf, k, horizon, lift, cost, budget):
    surv = [1.0]
    for t in range(1, horizon):
        surv.append(surv[-1] * p_cont(t, p1, p_inf, k))

    def tail(a, b):
        q = 1.0
        for t in range(a, b):
            q *= p_cont(t, p1, p_inf, k)
        return q

    rows = []
    for t in range(1, horizon):
        at_risk = cohort * surv[t - 1]
        reached = min(at_risk, budget / cost)      # budget caps who you can touch
        saved   = reached * lift * tail(t + 1, horizon)
        spend   = reached * cost
        rows.append((t, at_risk, reached, spend, saved,
                     spend / saved if saved else float("inf")))
    return rows, cohort * surv[horizon - 1]

# Assume: 500 accounts, 12 monthly periods, 55% continue after month 1 rising
# toward 95%, k = 0.35. A save play adds 10 percentage points of continuation
# and costs $180 of CSM time. Budget $27,000, i.e. 150 plays.
rows, baseline = save_curve(500, 0.55, 0.95, 0.35, 12, 0.10, 180.0, 27000.0)

cheapest = min(rows, key=lambda r: r[5])
most     = max(rows, key=lambda r: r[4])
print(f"baseline survivors at month 12: {baseline:.1f}")
print(f"cheapest per save : month {cheapest[0]} at ${cheapest[5]:,.0f}")
print(f"most saved        : month {most[0]} ({most[4]:.2f} accounts, ${most[3]:,.0f})")

assert cheapest[0] != 1              # month one is never the cheapest save
assert 1 < most[0] < len(rows)       # the budgeted optimum is interior
assert rows[0][5] > rows[-1][5]      # cost per save falls as the cohort self-selects

Two answers, and they disagree

Running it prints:

baseline survivors at month 12: 56.0
cheapest per save : month 11 at $1,800
most saved        : month 4 (6.91 accounts, $24,849)

Cost per retained account falls every month, from $8,839 in month 1 to $1,800 in month 11. Month 1 looks like the obvious place to spend because everyone is there, and it is the worst place: a save bought in month 1 has to survive eleven more months of decay before it counts at the horizon, and that tail discount is 0.204. You pay full price for a fifth of an account.

Cheapest is not the objective, though. Under a fixed budget the total saved peaks at month 4 — 6.91 accounts — and month 4 does it for $24,849, less than months 1, 2 and 3 each spend to save fewer. Before month 4 the budget binds and you cannot reach everyone at risk. After month 4 you can reach everyone, but there are fewer of them and the survivors were increasingly safe anyway. The optimum sits at the crossover. Change the budget and the crossover moves; change k and it moves further.

To push your own numbers through this without writing the loop, the streak and retention hazard calculator runs the same curve in the browser with the same budget cap. Two caveats before you click: its labels say days where this post says billing periods, so read every day as a period, and its headline output is the cheapest period, not the one that saves most. It will hand you the falling cost curve, not the month-4 answer — for that, watch the spend column for where it stops pinning to your budget.

Consumer-to-B2B mapping

The translation is not one-to-one, and the rows where it breaks matter more than the rows where it holds.

The last two rows are where the mapping earns its keep. Consumer churn is instant and observable. B2B churn is decided months before it is executed, and it surfaces in usage long before it surfaces in cancellation. The seat goes quiet in March and the renewal is lost in November. The hazard curve is the only tool here that operates on March.

What a streak freeze is for a seat-based product

The mechanic

A consumer streak freeze is an item that absorbs one missed day. It sells because losing a 300-day streak hurts more than the freeze costs — that asymmetry, and how to price against it, is the whole subject of loss aversion and streak freeze pricing.

The B2B analogue is written into the contract instead of sold as an item: unused credits that roll one period, seats that go dormant instead of forfeit, a tier threshold assessed on a trailing average rather than a single month. All of them do the same job. They stop one bad month from destroying an accumulated position, because an account that has already lost the position has nothing left to protect and churns more freely.

Where the reward should land

Two rules I would hold to. The reward lands on the account rather than the individual, because the individual leaves. And it lands as commercial value — credits, a better tier, terms — because a badge inside a B2B tool has nobody to display it to. Loyalty tier threshold calculus covers what a tier has to be worth before it moves behaviour, and leaderboards without churn covers the comparison mechanic that backfires most often.

The ethics line

A streak is legitimate when breaking it costs the user only what they genuinely lose. It becomes manipulation the moment you engineer an artificial loss to force behaviour that benefits you alone: burning an accumulated position to punish one lapse, timing a nudge to land at the moment of maximum anxiety, or making a paid freeze the only affordable escape from a penalty you invented. Endowed progress is fine when the pre-filled steps represent real work already done. It is a lie when they do not.

The test I use is disclosure. If I showed the user the model — the curve, the reward placement, the budget cap — would they call it reasonable? Month 4 in the example above survives that test. A nudge engineered to hit at 11pm on day 29 does not. There is more on where that line sits in incentive design and dark patterns. It is also why every create tool in the agent system on this site writes a proposal into an approval queue instead of touching a business table — and why that restriction lives in the tool surface rather than the credentials, since the key the agent holds could write anywhere. The human gate exists because automated incentive systems drift.

What I can and cannot show you

I searched for peer-reviewed work applying streak mechanics to seat-based B2B products and found none. What comes back is vendor blog content quoting retention lifts with no primary source attached — I chased the widely repeated "14% day-14 retention lift from streak wagers" figure through four vendor posts and never reached a primary study, so I have not used it. Treat this post as a model port. The mechanics are well evidenced; the B2B application is my reasoning.

What I can show is first-hand. I built the calculator this post links to, along with four others at /tools — five in total. They are plain client-side React components with no server call and no database behind them; the model runs entirely in your browser, which is why you can read the source of the answer it gives you. And the one large dataset I own says something relevant about intervening early on everything at once: Search Console for this site over the 85 days to 3 September 2026 shows 4,553 impressions and 22 clicks, with 111 of 182 ranking pages earning five or fewer impressions in the quarter. Effort spread across everything produced almost nothing. Cutting the published count from 267 to 103 and concentrating on the survivors was the same move as spending at month 4 instead of month 1 — the full teardown has the numbers. That is an analogy from search data, not a retention result, and I am labelling it as one.

FAQ

Do streaks work for B2B SaaS?

The mechanics do. Loss aversion and endowed progress are not consumer-specific. The daily-login implementation usually does not, because most B2B products have a weekly or monthly natural cadence, and a daily counter manufactures failures the user cannot avoid. Rebuild the counter on periods containing a qualifying event, at the account level, and it holds up.

What is the B2B equivalent of a streak freeze?

Contractual grace on an accumulated position: rolling unused credits one period, letting a seat go dormant instead of being forfeited, or assessing a usage tier on a trailing average instead of a single month. The purpose is identical — stop one bad month from destroying the position the account is trying to protect.

When is the cheapest time to save a churning account?

Cost per retained account falls with tenure, so the literal cheapest intervention is always the latest one. That is the wrong question. Under a fixed budget, the number of accounts you actually save peaks at an interior period — month 4 in the worked example above — where the budget stops binding but the at-risk pool is still large.

Are streak mechanics manipulative?

They are when the loss is manufactured rather than real. Rewarding a position the user genuinely built is legitimate; engineering an artificial forfeit to extract behaviour that serves only you is manipulation. The disclosure test settles most cases: if showing the user your model would embarrass you, do not ship it.

---

Running a retention programme and unsure where the spend should land? Fit the curve to one real cohort first, because its shape decides the answer far more than the reward cost does. Run your numbers through the streak and retention calculator or get in touch.