Power Users and the L28 Engagement Model
Power users are the slice of your monthly audience that keeps showing up, and the L28 model measures exactly how much: it counts how many of the last 28 days each user was active, then plots the whole population as a histogram from 1 day to 28. A user with L28=10 came back on ten of the past twenty-eight days. Read the shape of that histogram and you learn something a login count never tells you: whether your monthly active number is a handful of devoted regulars or a crowd of people who wandered in once and left.
So why does this beat the metric on everyone's dashboard? Because "monthly active users" treats a person who opened the app once the same as a person who opened it every single day. Both count as one MAU. The average hides them. L28 refuses to.
What L28 actually measures
The Ln family is simple. Pick a window of n recent days, then for each user count how many of those days they did the thing you care about. L7 uses the last seven days, L28 uses twenty-eight, L30 rounds to a calendar month. The credit for the "28" convention goes to Facebook's growth team, and the clearest write-up I know is Jonathan Hsu's 2017 Social Capital essay, which frames each user's L28 as a unit of value you can track month over month. If total L28 across your base is climbing, engagement is deepening. If MAU climbs while total L28 flattens, you're adding warm bodies, not habit.
Here's the part people miss. L28 is a per-user number first, a distribution second. The distribution is where the story lives. Plot active-days on the x-axis and the percentage of monthly users in each bucket on the y-axis, and you've drawn what Andrew Chen and a16z call the power user curve. Their framing is that the curve exposes both your hardcore daily-return segment and the sheer variability across users, nuance that a single ratio flattens into nothing.
Reading the curve: the smile you want to see
The shape has a name. A healthy engagement product tends to make a smile: a bump of light users on the left (1 to 3 active days), a sag in the middle, and a second spike pinned to the far right where the everyday people live. That right-hand spike is your power-user cohort. Chen's original piece is blunt about it. The goal is a curve that rises again at the high-frequency end rather than sliding to zero.
Think of it like a subway line at rush hour versus midday. A frequency chart of riders would show a wall of commuters who ride twice a day, five days a week, and then a long thin tail of tourists who rode once to see the city. Two completely different populations sharing one turnstile count. Average their trips and you'd describe a rider who exists nowhere on the platform.
Three shapes you'll actually run into:
- The smile. Left bump, dip, right spike. You have a real habitual core. This is the target for daily-use products.
- The slide (or "L"). A tall left bar that decays toward zero with no recovery on the right. Lots of one-and-done visitors, no daily habit. Common and honest for low-frequency products, worrying for one that claims to be a daily tool.
- The desert. Everything clustered in the low-to-middle range with no spike anywhere. Nobody hates you, nobody's hooked. Often the scariest, because there's no obvious segment to double down on.
Why the average lies: a worked example
Let me use a tiny dataset, because I trust twelve real rows more than a paragraph of theory.
Twelve users, one 28-day window, counting days they opened the product:
| User | Active days (L28) |
|---|---|
| A | 1 |
| B | 1 |
| C | 2 |
| D | 2 |
| E | 3 |
| F | 3 |
| G | 22 |
| H | 24 |
| I | 25 |
| J | 26 |
| K | 27 |
| L | 28 |
Mean active-days here is about 13.7. And 13.7 describes not one of these twelve people. Nobody sits near the middle. You've got six dabblers hugging 1 to 3 days and six near-daily regulars up at 22 to 28. That's a bimodal distribution, and the mean plants a flag in the empty valley between the two humps. If your weekly review reports "average user is active ~14 days a month," everyone nods, and everyone is wrong.
The histogram tells the truth instantly: two peaks, a hollow middle, a textbook smile. Now the product questions get sharp. What do the six power users do in their first week that the six dabblers never did? What's the one screen the regulars live in? You can only ask those questions once you've stopped hiding them inside an average.
This, incidentally, is why I gently side-eye "average session count" on so many exec dashboards. It's not that it's fake. It's that a bimodal population makes the average a number describing an imaginary person. Vanity isn't the same as wrong, but it's close enough to cost you a roadmap.
L7 or L28? Match the window to the rhythm
Which window you pick isn't a style choice — it changes the answer. The rule of thumb, echoed by Reforge's write-up on the curve, is to match the window to how often a satisfied user would naturally return.
| Product rhythm | Window | Why |
|---|---|---|
| Messaging, social, news feeds | L28 / L30 | Daily habit is the whole game; you want the far-right spike visible |
| B2B SaaS, productivity, dev tools | L7 | A happy weekly-workflow user looks "low" on a 28-day scale and false-alarms you |
| Weekly/occasional (payroll, delivery, banking) | L7 or custom | Expecting daily use is a category error; measure the intended cadence |
Run a Notion-style tool on L28 and even your best accounts might show 8 to 12 active days, which looks anemic until you realize they use it every workday and take weekends off. On L7 the same behavior reads as a proud 5-of-7. Same people, different verdict. So before arguing about whether your curve is healthy, argue about whether the window fits the product's natural pulse.
Where this goes wrong
A few traps I've watched teams fall into, mine included.
Defining "active" as merely "opened." If your active event is an app-open, your power-user spike might be push-notification reflexes, not value. Anchor the histogram on an action that means something — a message sent, a report generated, a track played. The curve is only as honest as the event under it.
Reading the curve once and framing it. A power user curve is a snapshot. The interesting motion is how the shape shifts month to month, which is exactly Hsu's growth-accounting point: watch total L28 and watch the right-hand spike fatten or thin. A single screenshot tells you where you are, not where you're heading.
Comparing your curve to someone else's benchmark. DAU/MAU bands help you sanity-check the neighborhood — Vmobify's 2026 breakdown puts social and messaging above 50%, gaming across a wide 20 to 50% band, fintech and B2B SaaS around 10 to 25%, and e-commerce lower still, with roughly 20% being a decent consumer baseline. Useful as a smell test. Useless as a target, because a fintech app chasing a messaging app's curve is trying to become a habit it has no business being.
Letting the ratio replace the shape. DAU/MAU is one number squeezed out of the whole distribution. Two products with an identical 25% stickiness can have wildly different curves: one a healthy smile, one a slow slide with a fat left bar and no core. The ratio is the shadow; the histogram is the object.
Turning the curve into a decision
Once you can see the two populations, the work is figuring out what separates them. Pull the days-active histogram from your raw event stream — most analytics platforms will build it, and an AI-native tool like Kixo can compute that distribution straight from your event data when you ask for it in plain language — then segment on it. Compare the onboarding paths of your right-spike users against your left-bump users. Look at which feature the regulars adopted in week one. That week-one behavioral fork is usually where your activation work should go.
And there's a money angle worth chasing. Your power-user spike and your high-value accounts are often the same people, which is why it's worth cross-referencing engagement depth against spend before you decide what to protect. If you want to connect frequency to revenue, our sibling guide on predicted-LTV models walks through how RFM and probabilistic models turn behavior into a dollar forecast. For structuring the metric that sits above all of this, our piece on the north-star metric tree shows where engagement depth hangs in the wider map.
The short version, since you've read this far: MAU tells you how many people your product touched, and the L28 curve tells you how deeply. One is a headcount. The other is a habit. Build for the second one, and the first tends to follow.