In this article

Forty new paying accounts sounds like a good month. If 30 accounts leave during the same period, the business grows by ten. If the remaining customers also reduce their subscriptions, recurring revenue might fall.

A growth model makes those relationships visible. It connects the way customers arrive and use the product to what the business gains and loses. The useful output is a set of conditional statements: if more of these customers reach this outcome, and their later behavior stays similar, this is what could change.

That is more demanding than drawing a funnel. It means keeping units consistent, accounting for time, and recognizing where one assumption depends on another. Start small enough that someone can explain every number.

First decide what you are counting

A person, a workspace, and a paying company can all be called a customer in the same meeting. They cannot occupy the same cell in a model.

For a collaboration product, ten active people might belong to one paying workspace. Multiplying those people by the workspace price overstates revenue. Yet counting only paying workspaces could hide a decline in use by the people inside them. Keep both where they matter, and write down the relationship between them.

Choose one segment for the first version. A small team buying with a card and a large organization negotiating a contract may reach the same product outcome through different paths. Averaging them immediately gives you a conversion rate that describes neither journey well.

Define the customer outcome in ordinary language. In a scheduling tool, “fills a real shift” says more than “completes onboarding.” It gives the model a behavioral event to connect to later payment and recurring use. Exclude sample work if it does not demonstrate the outcome you mean.

Reconcile the account count before projecting it

For a simple subscription business, start with an accounting identity:

Closing paid accounts = opening paid accounts + new paid accounts + reactivated accounts − lost paid accounts.

Each category needs a consistent rule, including what happens if an account leaves and returns within the same period. An unexplained remainder is a sign to investigate the records or definitions. Do not bury it in a conversion assumption.

Revenue needs its own reconciliation. Existing customers can upgrade or downgrade without changing the account count. Stripe’s explanation of monthly recurring revenue separates new revenue, expansion, contraction, and churn. That decomposition is useful because the same net increase can come from very different customer behavior.

Keep recurring run rate separate from cash collected. A customer paying annually brings in cash on a different schedule from the monthly equivalent of their subscription. For usage billing, consumption and realized rates add another relationship; multiplying accounts by a single subscription price may be inadequate.

The model does not need every accounting detail to answer a product question. It does need to avoid silently changing the meaning of revenue halfway through the calculation.

Build one path through the numbers

Here is an illustrative model for the scheduling product. One acquisition cohort contains 5,000 qualified visits. Eight percent create an account, 40% of those accounts fill a real shift, and 25% of the activated accounts become paying customers within a fixed observation window.

The calculation is:

5,000 × 8% × 40% × 25% = 40 new paid accounts.

The rates are conditional: activation is measured among signups, and payment among activated accounts. If your payment rate is instead measured among all signups, multiplying by activation again would count the same loss twice. If some accounts pay before filling a shift, represent that route explicitly rather than forcing them into the sequence.

Timing matters just as much. A 30-day payment window can carry an October signup into November. Cohort conversion estimates therefore cannot be dropped directly into an October cash or closing-account forecast. Use mature cohorts to estimate transitions, then place expected conversions in the calendar periods when they occur.

For a simplified period in which all 40 paid starts occur, suppose the business opens with 400 paid accounts, reactivates ten, and loses 30. It closes with 420. At a uniform $80 monthly subscription, that is a $33,600 month-end recurring run rate. The simplifications are visible, so another person can challenge them.

Ask how sensitive the result is

Change one input while holding the rest constant. This is sensitivity analysis: it shows what the arithmetic implies, before asking whether the change is achievable.

Change to the illustrative modelAdditional paid accountsAssumption carrying the result
Qualified visits rise from 5,000 to 6,0008 new accountsAdditional visitors convert at the same rates
Activation rises from 40% to 50%10 new accountsNewly activated accounts pay at the same rate
Period losses fall from 30 to 2010 accounts preservedThose accounts remain paid at period end

These are not equally easy interventions. They are not even estimates of intervention effects. The table identifies relationships worth investigating.

Try a less comfortable assumption. If acquisition grows by 20% but signup conversion falls from 8% to 6%, the new-account path produces 36 paid accounts at the unchanged downstream rates. More traffic produces fewer customers in this scenario. The model is useful precisely because it can expose that possibility before it becomes a target to celebrate.

Treat a new channel’s inherited conversion rates as assumptions to verify. Its customers may have different intent and economics, which is why Brian Balfour’s Four Fits framework considers the fit among market, product, channel, and business model together.

Add detail where the answer changes

A single activation rate may be adequate until two segments behave differently enough to change a decision. A single price may be adequate until expansion drives much of the revenue. Add those distinctions when they explain the business, rather than adding every available breakdown.

Keep observed inputs separate from assumptions. A measured conversion rate from mature cohorts is different from an estimated five-point improvement. A modeled referral route needs evidence that invitations lead to new independent accounts, not just a count of invitations sent. Feedback effects also take time; a customer acquired today rarely produces all their future referrals today.

When actual results arrive, replace the input estimates and inspect the differences. Did fewer eligible accounts arrive? Did the conversion window mature later? Did the new segment require more assistance? The discrepancy should lead to a more accurate relationship or a better question.

The model can tell you that ten extra activation points would be consequential. It cannot tell you whether a particular redesign will produce them, or whether that work deserves the next engineering slot. Those are further decisions. Its contribution is to make the expected path to growth explicit enough that the team can test it.

Sources

Something we should clarify?

Suggest a correction