The plateau: retention and product-market fit

Product-market fit means a real market needs the product enough to keep using it without being pushed. You can't measure it with one number, but a cohort's retention curve shows one of its clearest signals.

Every curve falls at first. Curious people try the product and leave; that's normal and says little. What matters is what happens next:

• The curve keeps sliding toward zero. Each month loses another slice of users. Nobody found lasting value; the product is only renting attention.
• The curve flattens. After the early drop, the share of active users stops changing much from month to month. That flat stretch is a retention plateau: a group of people who keep coming back on their own.

A plateau at a meaningful level is strong evidence that, for those people, the product does a job they need done. Read the slope of the later months, not the early drop: falling from 60% to 35% and then holding is far better news than falling only to 50% and then losing two points every month.

Once you see a curve flatten, put a number on it.

1. Find where the flat stretch starts. Look for the month after which the curve moves by only a point or two in total. For couples that's month 6: from 55.0% in month 6 to 52.7% in month 12, a change of about 2 points over half a year.
2. Average the flat months. Couples, months 6–12: (55.0 + 55.3 + 53.8 + 53.8 + 53.7 + 52.5 + 52.7) ÷ 7 = 376.8 ÷ 7 = 53.8%.
3. Say which measure and cohort it is: "plateau of the summer 2025 couples cohort, monthly activity by app opens, months 6–12: 53.8%".

Averaging several flat months is steadier than picking one, because any single month wobbles by a point or so. Don't start the average too early: including months 3–5, while the curve is still falling, would pull the level up and hide a slow decline.

A plateau also needs enough history. A cohort followed for four months can't show one yet; you can only say the curve hasn't stopped falling so far.

A plateau answers one question well: do the people who try the product find lasting value? About 54% of last summer's couples kept opening Halves every month for a year, without being pushed. That's real.

It doesn't answer several other questions that "product-market fit" is often taken to include:

• How big the market is. A plateau says nothing about how many more couples there are to win.
• Whether people will pay. Staying with a free tier isn't the same as buying Halves Plus; that needs conversion-to-paid data.
• Whether it will last. One summer's cohort is one data point. Newer cohorts may come from different channels and flatten lower, or not at all.

The size and origin of the cohort matter too. A plateau at 50% of 20 hand-picked beta users is mostly a story about those 20 people. A plateau at 30% of thousands who found the app in a store is much stronger evidence, because nobody chose them and the number is far less noisy.

A retention curve is only as honest as its qualifying action: the event that makes a user count as active. Change it and the same users can produce a different curve.

Some rows are written by the server, not by a person (Events, properties and users shows how to tell who caused a row). Many flatmates set up rent as a recurring bill, and from then on the server adds an expense_added event every month, whether anyone opens the app or not. Count "any event" and those people stay "active" for as long as the bill runs.

Subtract one curve from the other, month by month: in month 1 it is 68.9 − 65.9 = 3.0 points; by month 12 it is 50.4 − 39.4 = 11.0 points. It grows because people who stopped using the app keep accumulating server events; that widening gap is the signature of activity the user didn't do.

The rule: retention measures what people do. Use a user action (opening the app, adding an expense by hand, settling up), not server-side events. If automated use might be real value, check whether those flats still settle up.

With two curves for the same flats, you get two plateau levels over months 6–12:

• by app opens: (41.6 + 40.6 + 40.9 + 40.4 + 40.0 + 39.6 + 39.4) ÷ 7 = 282.5 ÷ 7 = 40.4%;
• by any event: 356.5 ÷ 7 = 50.9%.

Report the lower one as the headline, because it measures what people do, and mention the other in the same sentence: "flats plateau at about 40% by app opens; counting server-created rent bills puts them near 51%". Anyone who later sees the 51% figure on a dashboard will know where it comes from.

Maya's objection is fair: if the rent bill runs on its own, the product may be doing its job without being opened. That's a hypothesis, and it can be checked. Do those flats still record settle-ups? Do they cancel the bill when someone moves out, or does it run forever in an abandoned group? Until a user action supports it, server activity is not retention.

The same habit protects any metric: when two definitions give different numbers, publish the stricter one and say what the looser one adds.

Some uses of a product end when the job is done. A trip group exists for a holiday: people split the hotel and the dinners, settle up and stop. Trip users' curve falls to 5.6% by month 5, which on a monthly chart looks like no plateau at all.

But that fall is the product working. The group's lifetime is short by design, and it isn't the same as user churn: the person hasn't rejected Halves, they simply have nothing to split until the next trip. The end of the curve proves it: from month 9 it climbs back to 21.8% in month 12, mostly in new trip groups, as the next summer comes round.

So judge such uses by their own rhythm. For trips the question is "do people come back for the next trip?", which you read as the return a year later, or as the share of trip users who start another trip group. Monthly activity is the wrong yardstick.

The opposite mistake is just as easy: a rising tail is seasonal return, not a better plateau. One cohort shows a single summer-to-summer cycle, nothing more.

Don't blend segments whose usage rhythms differ: the blended curve follows whichever segment is in season, not anyone's retention. Put 600 couples, 800 flats and 1,600 trip users into one curve and you get 29.6% in month 3 and 32.7% in month 12, as if retention improved over the year.

It didn't. Inside the blend, couples slid from 60.3% to 52.7% and flats from 48.9% to 39.4%. The blend rose only because the biggest group, trips, came back for the next summer: 8.4% → 21.8% of 1,600 people outweighs the slow decline of the other two.

A blended curve is a weighted average, so it moves when the segments move or when their weights do, a mix shift. When segments have different rhythms (monthly for flats, yearly for trips), the blend describes none of them.

So split the curve by the segment that sets the rhythm, here the type of the first group, before you read a plateau. Show each line with its own measure: monthly activity for couples and flats, return for the next trip for trips. A single blended line is fine only when every segment inside it behaves alike.

For a product used in different ways, "do you have product-market fit?" has no honest yes or no. Answer per segment, with evidence:

• Couples: plateau at about 54% (months 6–12, app opens). Strong signal.
• Flats: plateau at about 40% by app opens; about 51% if server-created rent bills counted, which is why you quote 40%.
• Trips: judged by return for the next trip: 21.8% active again in month 12, mostly in new groups.
• Limits: one summer cohort; nothing yet on market size, willingness to pay or this year's cohorts.

One blended curve hides that couples and flats slowly slide while trips follow the seasons. A simple mean of the three plateaus mixes a monthly measure with a yearly one. Showing only the best segment answers a different question, "is there a group for whom Halves works?", so say that's the one you're answering.

How it fits together: read the curve's shape, measure the level on flat months, count only what people did, and judge each segment by its own rhythm, so the answer to "do we have fit?" is per segment, with limits. Next, Case: retention analysis by segment puts it into one memo.