Product-led growth flywheel: 5 stages, and which one to fix first

Your flywheel spins at the speed of its weakest stage. That is why so much growth work changes nothing: improving a stage that isn’t the constraint changes nothing downstream. A great referral program feeding users into broken onboarding produces more people who churn faster.

So the useful question is never “how do we improve the flywheel.” It’s “which of the five stages is the constraint right now, and what is the one metric that proves it.”

Funnel or flywheel: what is the difference?

The PLG flywheel is a growth model where each customer stage accelerates the next, so satisfied customers produce more customers.¹ A funnel ends at purchase and leaks users at every step. A flywheel returns them: adoption drives expansion, expansion funds retention, advocacy feeds acquisition. The median private B2B SaaS company roughly breaks even on the customers it already has: expansion replaces what churn takes away. Only companies above that break-even line grow without constantly buying new ones.²

The distinction matters for one practical reason. In a funnel, a weak stage costs you the users in it. In a flywheel, a weak stage starves every stage after it, which is why finding the constraint comes before improving anything.

Where the three PLG failures land on the flywheel

If you arrived here from the three ways PLG fails, this is the translation. Those failures are diagnoses at company level. The flywheel locates them at stage level, which is where the work actually happens.

Two of the three land on Expand, which is why Expand is the easiest stage to misread. An NRR problem caused by underpricing and an NRR problem caused by a procurement wall look identical in the metric and have nothing in common as fixes.

What are the 5 stages of the PLG flywheel?

The five stages are Evaluate, Activate, Adopt, Expand, and Advocate. That is this handbook’s cut of a model the industry has been drawing since Dave McClure’s pirate metrics in 2007; the labels vary between versions, the shape does not.¹ Each turns users from strangers into advocates, and each has one metric worth watching.

None of those metrics has a target here, and the reason holds for every number in this handbook. Each one is a ratio, and you define its denominator. What counts as a signup, whether an email address, a verified account or a completed profile, is your decision, and so is what counts as activated. Require email verification and your activation rate rises, because the denominator lost everyone who never clicked. Two companies with identical products report different numbers, which makes a published median a fact about how somebody else defined their terms. What to keep instead closes the book.

The exercise: rank your five transitions, then fix only the worst one

Five stages, and five gaps, because the wheel closes: the last transition feeds the first. The gaps are where users are actually lost, and ranking them needs no outside number, because you are only ever comparing yourself to yourself. If you ran the diagnostic in the previous chapter you already hold three of these five, which is the point: this is the same measurement at a finer grain.

Get each number twice: as it is now, and as it was a year ago or in the best stretch you have had. Rank by the gap between those two, not by which percentage looks lowest. The five rates are not the same kind of number and comparing them with each other tells you nothing, but each one compared with its own past tells you where you are losing ground.

Rough numbers are fine, and a missing one is itself an answer: a transition you cannot measure is rarely the one you are managing well. If you would rather work from a single list than a ranking, the checklist runs the same material in the order the work is usually done. Then fix the rank-1 transition and deliberately leave the remaining four alone for a quarter. That instruction is the entire method. It still works when your product is unusual, which no benchmark number does.

Why is NRR the flywheel’s master metric?

The fourth transition has a name of its own, and net revenue retention is the one number that answers whether the wheel turns forward at all. It’s Expand’s stage metric and the whole flywheel’s health indicator, because expansion is where a flywheel actually pays off.

NRR answers one question: if you stopped acquiring customers today, would you grow or shrink? A flywheel spinning forward means customers stay rather than churn, spend more, and bring others. NRR captures all three in a single figure, which no stage metric does.

The obvious next question is which band counts as good, and that is the one to refuse. Published medians sit a long way below the bar most people quote, so the widely repeated “healthy is above 120%” bar describes an unusually good year while presenting itself as a floor.² What that does to a company reading its own dashboard is worked through in Expand. The comparison that means something is against your own last four quarters.

Snowflake disclosed 158% net revenue retention in its 2020 S-1, meaning it grew 58% a year from existing customers before signing anyone new.³ Two things about that number. It is a ceiling, not a target. Snowflake charges for consumption, so its revenue grows on its own as customers move more data in. A product sold per seat gets no such free push. And it did not hold. Snowflake now reports around 125%.³ One of the strongest expansion engines in software fell back toward the average, which is the strongest available argument against benchmarking yourself to an outlier’s peak.

NRR is also the one metric no single stage owns: it moves when retention improves, when expansion improves, or when advocacy brings better-fitting customers. That is what makes it a summary rather than a target. You still fix one stage at a time; NRR is how you check the fix reached the bottom line.

How do you find your flywheel’s constraint?

One stage is holding the wheel. Find it, fix it, then find the next one. This is the Theory of Constraints applied to growth: in any system, one bottleneck sets total throughput, so effort spent anywhere else is wasted until that bottleneck moves.⁴ The discipline it demands is leaving four stages alone while you work on one.

1. Run the ranking above, then take the rank-1 transition. If two are close, prefer the earlier one. A broken Activate stage makes Adopt and Expand look broken too, so the earliest constraint often moves three metrics at once. One exception, covered below: if you already know your problem is unit economics rather than activation, start at Expand. Stage order is for readers starting cold, and following it blindly with an economics problem walks you through three fixes that make it worse.
2. Confirm it’s the constraint, not a symptom. Compare drop-off at this transition against every later one. If a later transition loses a larger share, the metric here is being dragged down by cohorts that were already doomed upstream, and this stage is a symptom.
3. Interview three users who dropped at that transition. Ask what stopped them. The answer is usually more mundane than the theory.
4. Fix one thing, then re-measure before moving on. Two simultaneous fixes tell you nothing about which worked.

The instinct is to spread effort evenly, celebrate signup growth, and copy whichever stage a competitor is famous for. Each of those spends money on a stage that is not the constraint. Find your own weakest gap, fix what is behind it, and let the rest sit.

The chapters that follow run in wheel order, which is an order of operations for the product rather than an instruction about where to start. If your rank-1 transition is signup-to-activated, the aha moment and time to value are the chapters to open first; the acquisition chapters solve a problem you do not have yet. A few patterns serve more than one stage and appear on more than one card above, and each gets its chapter where it first does real work.

Read in order and the wheel starts where a stranger first meets you. The cheapest version of that is the one where they meet the product itself, doing its ordinary job, without anybody having decided to market to them: co-experience is where the Evaluate stage begins.

Updated August 2026