What is the right way to compute true gross retention vs net retention when half your customers are on multi-year contracts with annual escalators?

The Core Problem: Subscription Retention Math Assumes A Baseline That Usage Pricing Does Not Have

Every standard retention formula a SaaS finance team has ever used carries a hidden assumption so deeply embedded that nobody states it out loud: there is a stable contractual baseline to measure against. Gross revenue retention is "what fraction of the revenue I had a year ago do I still have," and net revenue retention is "what is that same cohort worth now, including expansion." Both formulas require a clean, unambiguous *starting number* — the ARR a customer was contractually obligated to pay you at the start of the measurement window.

In a pure-subscription business that number exists, it is written into a contract, it does not move between renewal dates, and it is sitting in your billing system as a single field. The math is mechanical because the denominator is a fact.

Usage-based pricing breaks that assumption at the root. A customer on consumption pricing — paying per API call, per gigabyte ingested, per compute-second, per seat-day, per message sent — does not have a contracted ARR. They have a *revenue stream that the customer themselves controls month to month*.

In January they spend \$8,000. In February, \$11,000. In March, \$6,500 because their own business had a slow month.

In April, \$14,000 because they launched a feature that hammered your API. There is no contract line that says "this customer is worth \$X/year." There is only a trailing series of invoices, each one a different number, none of them authoritative.

So when you sit down a year later and ask the retention question — "did this customer expand, stay flat, contract, or churn?" — you discover the question is genuinely ambiguous, not because your data is bad but because the concept itself is underdefined. Compared to *what* baseline?

Last January's single month? The average of last year? Their highest month?

Their committed minimum? Each of those choices produces a *materially different* retention number for the exact same customer with the exact same behavior. A customer who spent \$8K last January and \$7K this January looks like 88% retention against a point-in-time baseline — but if their trailing-twelve-month average went from \$9K to \$10.5K, they actually *grew* 117%.

Same customer. Same invoices. Opposite story.

The number you report depends entirely on a methodological choice you may not even realize you are making.

This is the core problem, and everything else in this entry follows from it. The mixed-model company is not doing harder arithmetic than a subscription company — it is doing arithmetic on a foundation that has to be *constructed* before the arithmetic is meaningful. The baseline is not given; it is a decision.

And because it is a decision, it must be made explicitly, documented, defended, and held constant — or your retention metrics become a Rorschach test that tells every reader whatever they want to see.

Why You Can't Just Blend Them Into One Company Number

The instinct of most finance teams confronting a mixed pricing model is to compute one company-wide NRR and one company-wide GRR — take all the revenue, all the customers, run the standard formula, report a single number to the board. This is the most common and most damaging mistake in mixed-model retention measurement, and it is worth being precise about *why* it fails.

A subscription cohort and a usage cohort have fundamentally different *retention physics*. The subscription book tends to cluster: most customers sit at exactly 100% retention (they renewed at the same price), a minority expand in discrete jumps (they bought more seats), a minority contract or churn at renewal events.

Its distribution is lumpy and renewal-gated. The usage book is the opposite: almost *no* customer is at exactly 100%, because usage never lands on precisely last year's number. Every usage customer is at 94% or 112% or 88% or 134% — the distribution is continuous, wide, and unconstrained by renewal dates.

When you average a tight, renewal-gated distribution with a wide, continuous one, the blended mean describes neither population. It is a number that exists nowhere in your customer base — a statistical artifact.

Worse, the blend hides the signal that actually matters. Suppose your subscription segment posts 96% GRR and 104% NRR — solid, healthy, predictable. Your usage segment posts 82% GRR and 128% NRR — high expansion masking real contraction underneath.

Blend them 50/50 and you might report 89% GRR and 116% NRR, which looks like a perfectly respectable mid-market SaaS. But that blended number has *erased the most important fact about your business*: that one half of your customer base is bleeding 18 points of gross revenue a year and only looks healthy because expansion is papering over it.

A board reading 89%/116% has no way to know there is a structural retention problem in half the company. The blend is not just imprecise — it is actively concealing.

The correct mental model is this: the mixed-model company runs two retention businesses under one logo. They have different baselines, different volatility, different leading indicators, different forecasting behavior, and different stories to tell investors. You compute them *separately*, with the right methodology for each, and only *then* do you combine them — deliberately, with a documented weighting — into a blended figure.

And even then, the blend is the *least* informative of the three numbers. You report it because investors ask for a single comparable, but the two segment numbers are where the truth lives. Blend last, never first.

The Baseline Problem In Usage Pricing: What Is The Denominator?

Strip the retention question down to its arithmetic core and it is a fraction: revenue-now divided by revenue-then, for a fixed cohort. The numerator is rarely controversial — it is this period's revenue from the cohort, and you can pull it from billing. The entire difficulty lives in the *denominator*: revenue-then.

For a subscription customer, revenue-then is their contracted ARR at cohort start — a single, stable, contractual fact. For a usage customer, "revenue-then" does not exist as a fact. It has to be *defined*, and the definition you choose is the single most consequential decision in the whole exercise.

There are four candidate denominators, and they are not close to equivalent. Option one: last month's usage, annualized. Take whatever the customer spent in the cohort-start month, multiply by twelve. Simple, but catastrophically noisy — if the start month happened to be a seasonal peak, every customer looks like they contracted; if it was a trough, everyone looks like they expanded.

You are baselining off a single coin flip. Option two: a trailing average — typically the trailing three or trailing twelve months as of cohort start. This smooths the coin-flip problem and is the most defensible choice for most usage businesses.

The cost is that it lags reality and dampens both real expansion and real contraction. Option three: the contracted minimum or commitment. If your usage customers sign a committed-spend deal — "you'll spend at least \$120K this year, drawn down as you consume" — then the commitment *is* a contractual baseline, and you are suddenly back on subscription-like footing.

This is the cleanest option *when it is available*. Option four: the customer's own prior-period actuals matched period-over-period — compare this March to last March specifically, to neutralize seasonality. Useful as a cross-check, fragile as a primary method because it assumes seasonality is stable year over year.

The reason this matters so much is that the *same customer* produces wildly different retention numbers under each denominator. Consider a customer whose monthly spend over a year ran: 8, 9, 7, 6, 8, 10, 12, 9, 7, 8, 11, 14 (in \$000s). Their cohort-start month was \$8K.

Their trailing-3-month average at start was ~\$8K. Their full-year average was ~\$9.1K. Their committed minimum was \$84K/year (\$7K/month).

This period they are running \$14K/month. Against last-month-annualized (\$96K), they are at 175% NRR. Against full-year-average-annualized (\$109K), they are at 154%.

Against the commitment (\$84K), they are at 200%. Three legitimate methodologies, three numbers spanning a 46-point range, *for one customer who did exactly one thing*. Multiply that ambiguity across an entire cohort and you understand why two analysts at the same company, both "computing NRR correctly," can hand the CFO numbers that differ by 15-20 points.

The denominator is not a detail. It is the whole game.

The Three Baseline Methodologies: Point-In-Time, Trailing Average, Annualized Run-Rate

In practice, mature mixed-model finance teams converge on three named methodologies for the usage baseline, and the discipline is in knowing when each is appropriate and what each one hides.

Methodology A — Point-in-time snapshot. The baseline is the customer's revenue in a single anchor month (cohort start), annualized. The numerator is their revenue in the comparison month, annualized. This is the easiest to compute and the easiest to explain, and it is *wrong for almost every usage business* as a primary metric.

It treats one month's coin flip as a structural fact. Its only legitimate uses are: (1) a directional, fast monthly read where you accept the noise, and (2) businesses whose usage genuinely has near-zero month-to-month variance (rare — think a metered utility-like product with extremely stable consumption).

What it hides: everything seasonal, everything lumpy, every customer whose anchor month was unrepresentative. If your usage NRR swings 25 points between consecutive monthly cohorts, you are almost certainly on a point-in-time methodology and should stop.

Methodology B — Trailing-3-month (or trailing-twelve) average. The baseline is the average monthly revenue over the three (or twelve) months ending at cohort start, annualized; the numerator is the average over the trailing window ending at the comparison date. This is the workhorse and the right default for the large majority of usage businesses.

The trailing-3 version is responsive enough to catch real trend changes within a quarter while smoothing single-month noise; the trailing-12 version is the most stable but lags hardest. What it hides: a genuine cliff gets averaged into a gentle slope for the length of the window, so a customer who fell off in month one of your window still looks half-alive for two more months.

You accept that lag in exchange for not chasing noise.

Methodology C — Annualized run-rate against commitment. The baseline is the customer's *contracted* annual commitment or credit purchase, not their actual usage; the numerator is their actual annualized consumption (or the renewed commitment). This is the cleanest methodology *and* the most data-dependent — it only works if your usage customers actually sign commitments or buy credit pools.

When they do, you have effectively re-created a contractual baseline, and your usage GRR/NRR behave almost like subscription metrics, with the bonus that overage above the commitment shows as expansion. What it hides: customers who consume *below* their commitment still count as 100% retained against the floor even though they are economically unhealthy — so commitment-based retention can look artificially stable while a burndown problem builds underneath.

The fix is to track *commitment utilization* as a companion metric. Most sophisticated consumption businesses run Methodology C for retention and utilization as the early-warning overlay.

The governing rule across all three: pick one per segment, document it precisely, and never change it without restating prior periods. A methodology change that is not restated is indistinguishable from a performance change, and that is how finance teams lose credibility with boards and auditors.

Distinguishing Usage Variance From Real Contraction

This is the single hardest judgment in mixed-model retention, and getting it wrong in either direction destroys the metric. A usage customer's revenue dropped 30% this month. Did they *contract* — a permanent, structural reduction in how much they will ever consume — or did they merely *vary* — a temporary dip that will mean-revert?

The retention math treats those two cases completely differently: variance should be smoothed away and ignored; contraction should be measured and counted. But at the moment the drop happens, they look identical. A single data point cannot tell you which one you are looking at.

The only reliable way to distinguish them is *time plus a structural test*. First, the smoothing window: you do not classify a revenue drop until it has persisted across your chosen window — typically three to six months. A customer who drops in March and recovers by May was variance.

A customer who drops in March and is still down in August is contraction. This is precisely why point-in-time methodology fails: it forces a classification decision on month one, before you have the information to make it. Second, the step-change test: real contraction usually looks like a *step* — usage was running at one level, then dropped to a distinctly lower level and *stabilized there*.

Variance looks like *oscillation* around a stable mean. If you plot the trailing-3-month average and it has shifted to a new plateau, that is contraction; if it is wobbling around the same line, that is variance. Third, the cause overlay: variance has benign explanations that correlate with the customer's own seasonality (retail customer slow in summer, tax-software customer slow after April), while contraction correlates with adverse events — they migrated a workload to a competitor, they downsized, they re-architected to use you less, a champion left.

The practical implementation: every usage customer carries a *rolling classification* — "stable," "in variance band," "trending down (unconfirmed)," "confirmed contraction," "confirmed cliff." The classification only hardens after the smoothing window confirms it. Your retention computation uses the *confirmed* state, not the raw monthly delta.

This means your most recent two-to-three months of usage retention are always provisional and get revised as the window matures — and that is correct behavior, not a bug. Mixed-model finance teams that refuse to accept provisional-then-revised retention numbers end up either chasing noise (classifying every dip as churn) or going blind (smoothing so hard they never see a real cliff until it is a quarter old).

The seasonal-business case makes this vivid: a customer with a legitimate 40% summer trough is *not* contracting, and a methodology that flags them every July is broken — but the same 40% drop in a non-seasonal customer that persists into the fall *is* contraction and must be counted.

The Commitment/Minimum Floor As An Anchor

The cleanest escape from the baseline problem is structural rather than analytical: if your usage customers carry a committed minimum, an annual spend commitment, or a prepaid credit pool, you have a contractual anchor and most of the ambiguity disappears. This is why so many consumption-pricing companies have moved aggressively toward committed-spend contracts — it is not only a cash-flow and forecasting benefit, it makes the retention math *tractable*.

Here is the mechanism. When a customer commits to spend, say, \$240K over the next twelve months — drawn down as they consume, with overage billed at a defined rate — that \$240K is a real contractual baseline, exactly like a subscription ARR. At renewal, you ask the same questions a subscription company asks: did they renew the commitment flat (100%), increase it (expansion), decrease it (contraction), or walk away (churn)?

Usage *above* the commitment shows cleanly as expansion. The credit-burndown variant works the same way: a customer prepays for a pool of credits, and the retention question is whether they re-up the pool and at what size. In both cases you have re-anchored a fluctuating revenue stream to a fixed contractual number, and your usage GRR/NRR start behaving like well-defined subscription metrics.

But the commitment floor introduces its own distortion that you must manage, and it is the mirror image of the point-in-time problem. A customer can be "100% retained" against their commitment while being economically sick. Imagine a customer who committed to \$240K, consumed only \$160K of it last year (a 67% burndown), and renews the commitment at \$240K again.

Against the commitment baseline, that is 100% GRR — perfect. But the customer is *not* healthy: they are paying for capacity they do not use, which is a leading indicator of either a downsell at the next renewal or outright churn. Commitment-based retention can therefore look reassuringly stable right up until a wave of under-consuming customers all decide not to re-up.

The required companion metric is commitment utilization — actual consumption divided by commitment — tracked at the customer level and watched as an early-warning system. Healthy: utilization 85-110% (consuming roughly what they committed, with some overage). Watch: 60-85%.

Alarm: under 60%. The combination — commitment-based retention as the headline number, utilization as the leading indicator — is the most robust setup available to a usage business, and it is what the better consumption-model public companies actually run internally even when they disclose something simpler.

The Consumption-Model NRR Reality: Why The High Numbers Are Real But Volatile

Pure consumption companies — Snowflake, Datadog, MongoDB Atlas, Confluent, and their peers — have historically posted net revenue retention numbers that look extraordinary next to traditional SaaS: figures in the 115-160%+ range, sometimes higher in growth phases. A subscription-only CFO sees those numbers and assumes either the consumption company is exceptional or the metric is being gamed.

Neither is quite right. The high consumption NRR is structurally real — and structurally more volatile — and understanding both halves of that is essential to reading a mixed-model book correctly.

It is real because of the fundamental mechanic of usage pricing: revenue grows automatically with the customer's own success, with zero sales motion required. A subscription company has to *sell* expansion — more seats, a tier upgrade, a new module — and each of those is a discrete deal with a sales cost and a close rate.

A consumption company expands every time the customer ingests more data, runs more queries, sends more messages, or onboards more of their own users onto the product. If the customer grows, you grow, frictionlessly. A healthy consumption customer who started at \$100K and whose own business doubled is now at \$200K without anyone at the vendor lifting a finger.

Aggregate that across a cohort of growing customers and 130%+ NRR falls out of the math naturally. It is not an accounting trick; it is the pricing model doing what it is designed to do.

But the same mechanic that makes the number high makes it volatile and macro-sensitive in a way subscription NRR never is. Because revenue tracks the customer's own activity, it also tracks the customer's own *contraction* — and the broader economy. When customers tighten budgets, they do not have to call you to downsell; they simply optimize their queries, archive cold data, turn off a pipeline, and your revenue falls *the same month*, with no renewal event required.

This is exactly what happened across the consumption-software sector in the 2022-2023 "optimization" wave: NRR figures that had been 160%+ compressed toward 120% and below within a few quarters, not because customers churned but because they consumed less. A subscription book cannot move that fast — it is renewal-gated.

A consumption book re-prices itself continuously, in both directions. So when you read a consumption NRR, you must read it as a *high-beta* number: the upside is bigger and free, the downside is faster and uncontracted. For a mixed-model company, this means the usage half of your retention will always be the more exciting and the more nerve-wracking half, and your forecasting has to treat it accordingly.

The Consumption-Model GRR Reality: The Cliff Versus The Slow Bleed

If consumption NRR is the exciting number, consumption GRR is the hard one — and it is the number a mixed-model CFO should worry about most, because it is both harder to measure and easier to flatter yourself about. Gross revenue retention asks: of the revenue I had, how much did I *keep*, ignoring all expansion.

In a subscription business, "losing" revenue means a contraction at renewal or an outright cancellation — discrete, visible, dated events. In a consumption business, a customer can erode 40% of their revenue *without ever generating a single event you would naturally call churn.* They never cancelled.

They never downgraded a plan. They are still "active," still logging in, still technically a customer. They are just... using you 40% less.

Your GRR is bleeding and nothing in your CRM says so.

This produces two distinct failure modes that consumption GRR has to capture, and they require different detection. The cliff is the consumption equivalent of churn: a customer whose usage drops to near zero and stays there — they migrated off, the project ended, they went out of business.

Cliffs are relatively easy to detect *after* the smoothing window confirms them, and they should be counted as gross churn even though no cancellation was ever processed. The harder failure mode is the slow bleed: a customer who does not fall off a cliff but ratchets down — \$10K/month, then \$9K, then \$8K, then \$7K — over many months, each step small enough to look like variance, the cumulative effect a 30-40% structural loss.

The slow bleed is where consumption GRR goes to die quietly, because every individual month is deniable and the trend only becomes obvious in aggregate.

Measuring consumption GRR correctly therefore means doing something subscription GRR never has to do: affirmatively measure the floor of each customer's usage over the period, not just check whether they cancelled. The methodology (detailed in the next two sections) is to take the cohort, fix the baseline, and then measure each customer's *sustained lower bound* over the comparison period — capped at 100%, because GRR by definition ignores expansion.

A customer who oscillated between \$8K and \$14K against an \$8K baseline contributes 100% to GRR (their floor held). A customer who stepped down to a sustained \$5K against that \$8K baseline contributes ~62% — that 38-point loss is real gross churn, even though nobody ever cancelled anything.

A mixed-model company that computes consumption GRR by only counting logo cancellations and explicit downgrades is systematically *overstating* its gross retention, sometimes by 15-25 points, and will be blindsided when the slow bleed finally shows up in a renewal cohort or a cash forecast.

Computing GRR For The Usage Segment: The Step-By-Step Methodology

Here is the concrete computation for usage-segment gross revenue retention, assembled from the principles above. The goal of GRR is to measure *retained* revenue while ignoring expansion entirely — so the math is deliberately conservative and floor-oriented.

Step 1 — Define and freeze the cohort. Pick every customer who was an active usage customer as of the cohort-start date (e.g., all usage customers active on the first day of the trailing measurement year). This cohort is *fixed*. New customers acquired after the start date are never added to it — they belong to later cohorts.

The cohort can only stay the same or shrink; that is what makes GRR a clean retention measure.

Step 2 — Compute the baseline for each cohort member. Apply your chosen methodology consistently. If trailing-3-month average: for each customer, average their monthly usage revenue over the three months ending at cohort start, then annualize. If commitment-based: use each customer's contracted commitment as of cohort start.

Sum the per-customer baselines to get the cohort baseline revenue. This is your denominator and it never changes for the life of the cohort.

Step 3 — Measure each customer's sustained floor over the comparison period. This is the step unique to usage GRR. For each customer, look at their revenue across the comparison period (smoothed by the same window) and identify their *sustained lower bound* — the level they held, not their peak and not a single-month trough.

In practice: take the trailing-3-month average at the comparison date, or the minimum sustained 3-month average over the trailing year, depending on whether you want a point-in-time or worst-sustained read. Cap each customer's retained revenue at their baseline — GRR never gives credit for expansion.

Step 4 — Handle the edge states explicitly. Confirmed cliffs (usage at or near zero, persistent past the smoothing window): retained revenue = 0. Confirmed contractions: retained revenue = sustained lower level, capped at baseline. Customers in unconfirmed variance: use their smoothed current level, flagged as provisional.

Customers who moved to a different segment (migrated to subscription): see the migration section — do not silently drop them.

Step 5 — Aggregate. Usage GRR = (sum of per-customer retained revenue, each capped at its baseline) ÷ (cohort baseline revenue). Because every numerator term is capped at its denominator term, usage GRR is mathematically bounded at 100% — as it should be. A well-run consumption business often posts usage GRR in the 78-92% range; subscription GRR in a healthy SaaS is usually 88-95%.

The gap between them is the structural cost of the consumption model's lack of a renewal gate, and it is *normal* — the point of measuring it correctly is to know its true size, not to make it disappear.

Computing NRR For The Usage Segment: Capturing The Full Expansion Picture

Net revenue retention for the usage segment uses the *same fixed cohort and the same baseline* as GRR — the only thing that changes is that NRR lets the numerator run past 100%, capturing the full expansion that GRR deliberately ignores. Keeping the cohort and baseline identical between the two metrics is non-negotiable: if GRR and NRR are computed off different cohorts or different baselines, the pair becomes incoherent and you cannot reason about the gap between them.

Step 1 and Step 2 are identical to the GRR computation — same frozen cohort, same per-customer baseline, same cohort baseline revenue as the denominator.

Step 3 — Measure each customer's ending revenue, expansion included, no cap. For each cohort member, take their smoothed revenue at the comparison date (trailing-3-month average annualized is the standard). Unlike GRR, you do *not* cap at baseline — a customer who grew from \$8K/month to \$14K/month contributes their full \$168K annualized.

This is where the consumption model's automatic, sales-free expansion shows up.

Step 4 — Net the contractions and churn against the expansions. The cohort's ending revenue is the sum of every member's current smoothed revenue: expanders contribute more than their baseline, flat customers contribute roughly their baseline, contractors contribute less, cliffs contribute zero.

NRR nets all of this together. Usage NRR = (sum of per-customer current smoothed revenue) ÷ (cohort baseline revenue).

Step 5 — Decompose the NRR into its drivers, because the headline number alone is misleading. A usage NRR of 124% could be a *healthy* book — broad-based expansion across most customers, modest churn — or a *fragile* one — a few whales expanding massively while the long tail bleeds.

These are completely different businesses with the same headline number. So always report usage NRR alongside its bridge: starting cohort revenue → (+) expansion from growing customers → (−) contraction from shrinking customers → (−) churn from cliffs → ending cohort revenue. The bridge is what tells you whether 124% is durable or whether it is one big customer away from being 95%.

For a mixed-model company specifically, this decomposition is also what lets the board see that the usage segment's high NRR is *real expansion*, not a baseline artifact — which is the credibility question every investor will eventually ask.

Computing Both For The Subscription Segment: The Standard Contractual Method, Stated Cleanly

The subscription half of a mixed-model book is the easy half — but "easy" does not mean "skip the rigor." It means the baseline is *given* rather than constructed, so the computation is mechanical. The mixed-model company still has to do it carefully and, critically, has to do it on the *same cohort timing and the same measurement period* as the usage segment, or the two halves will not be combinable.

(This entry's companion piece, q97, covers general GRR-vs-NRR computation in depth; what follows is the subscription method stated cleanly enough to slot alongside the usage method above.)

Cohort and baseline. Take every customer who had an active subscription as of the cohort-start date. Each customer's baseline is their contracted ARR as of that date — a single, stable number sitting in the billing or contract system. Sum them for the cohort baseline.

There is no smoothing window, no methodology choice, no provisional classification — the contract says what it says.

Subscription GRR. Over the measurement period, for each cohort member, count only *losses*: full churn (cancelled, baseline ARR lost entirely), and downgrades/contraction at renewal (partial ARR lost). Retained revenue = baseline ARR minus losses, with expansion ignored entirely.

Subscription GRR = (cohort baseline − total churn − total contraction) ÷ cohort baseline. It is bounded at 100% by construction.

Subscription NRR. Same cohort, same baseline, but now add expansion: upsells, cross-sells, seat additions, tier upgrades within the cohort. Subscription NRR = (cohort baseline − churn − contraction + expansion) ÷ cohort baseline. It can and should exceed 100% for a healthy book.

The key discipline points for the mixed-model context. First, *timing alignment*: run the subscription cohort on exactly the same start date and measurement length as the usage cohort, so the two segment numbers describe the same slice of time. Second, *expansion attribution*: when a subscription customer also starts buying usage, that incremental usage revenue does *not* belong in subscription NRR — it belongs in the usage segment (or the hybrid treatment below).

Keep the two revenue types in their own buckets even for a single customer. Third, *renewal-event hygiene*: subscription contraction is recognized at the renewal event, not when the customer hints at it — which means subscription metrics lag reality in the opposite way usage metrics do (usage moves too fast, subscription moves too slow).

Knowing both lag directions is what lets you read the blended number sensibly.

Combining The Two Into A Blended Company Number

Once you have four clean numbers — subscription GRR, subscription NRR, usage GRR, usage NRR — the blend is straightforward arithmetic, but the *presentation discipline* around it is where mixed-model companies either build credibility or lose it.

The mechanic: revenue-weight, do not customer-weight. The blended company GRR is the revenue-weighted average of the two segment GRRs, where the weights are each segment's share of the *cohort baseline revenue* — not the count of customers in each segment. If the subscription segment is \$30M of baseline and the usage segment is \$20M, the weights are 60% and 40%.

Blended GRR = (0.60 × subscription GRR) + (0.40 × usage GRR). Blended NRR is computed the same way. Equivalently and more robustly, you can compute it bottom-up: total retained (or ending) revenue across *both* segments' cohorts, divided by total baseline across both — which automatically revenue-weights and is the formulation least prone to error.

Customer-weighting is wrong because a \$2K/month customer and a \$200K/month customer would count equally, which describes no economic reality.

The presentation rule, which is the actual point of this section: always report all three, never just the blend. Every board deck, every investor update, every internal retention review shows *subscription segment GRR/NRR*, *usage segment GRR/NRR*, and *blended company GRR/NRR* — three rows, not one.

The blend alone is the least informative number you have; it is a summary statistic that, as established earlier, can describe a population that does not exist and can conceal a structural problem in half the company. Reporting only the blend is not a simplification, it is an omission.

The two segment numbers are where a reader can actually see the business: where the volatility is, where the gross bleed is, where the durable expansion is. The blend exists only because investors want a single comparable figure for the company as a whole — give it to them, but never *instead* of the segments.

The weighting must also be disclosed and held constant. Note the segment mix (e.g., "60% subscription / 40% usage of measured baseline") alongside the blended number, because if the mix shifts, the blend will move *even if both segment numbers are flat* — and a reader needs to be able to tell a mix-shift from a performance change.

A mixed-model company that is rapidly shifting toward usage will see its blended retention metrics drift simply from re-weighting; disclosing the mix is what keeps that honest.

The Cohort Definition Across Mixed Models: Handling The Hybrid Customer

The hardest cohorting problem in a mixed-model business is not the pure-subscription customer or the pure-usage customer — it is the hybrid customer who carries *both* a subscription component and a usage component on a single contract. A customer might pay \$60K/year for a platform subscription *and* \$8-15K/month in metered API consumption on top of it.

Which segment does this customer belong to? The wrong answers are "pick the bigger one and assign the whole customer there" (you lose half the revenue's true behavior) and "put the customer in both cohorts whole" (you double-count them). The right answer is decompose, do not assign.

The principle: the cohort unit for a mixed-model business is not the *customer*, it is the revenue stream. A hybrid customer is decomposed into two cohort entries — their subscription component goes into the subscription cohort with a contracted-ARR baseline, and their usage component goes into the usage cohort with a smoothed or commitment baseline.

The same customer, split into the two revenue types, each measured with the methodology appropriate to that revenue type. At the blended level they recombine automatically because the bottom-up blend sums across both cohorts. This is the only treatment that produces coherent numbers: it lets a customer who grew their subscription but shrank their usage show up correctly as expansion in one segment and contraction in the other, instead of being flattened into a single misleading customer-level number.

There is one important nuance: the decomposition has to be stable over the life of the cohort. If a customer's contract bundles subscription and usage in a way that makes the split ambiguous — a single committed dollar figure that covers both a platform fee and a consumption allowance — you need a documented allocation rule (e.g., "the platform-fee line of the order form is the subscription component; everything metered is the usage component") and you apply it identically at baseline and at the comparison date.

The danger is letting the allocation drift, because then a customer can appear to expand or contract purely from a re-classification. Document the allocation rule, version it, and treat a change to it the same way you treat any methodology change: restate history. The hybrid customer is also the customer most likely to be on a *committed-spend* usage deal, which is convenient — it means the usage component of many hybrids can use the cleaner Methodology C baseline even when your standalone usage customers are on trailing-average baselines.

The "Customer Migrated From Subscription To Usage" Problem

A specific and increasingly common edge case deserves its own treatment: the customer who *changes pricing models mid-life* — they were on a flat subscription, and partway through the cohort year they re-papered onto usage-based pricing (or the reverse). This is not a hypothetical; it is the dominant pattern in the industry-wide drift toward consumption pricing, and handled carelessly it can manufacture phantom churn or phantom expansion that pollutes both segments.

The naive treatments both fail. Treatment one — "the subscription ended, count it as subscription churn, and the usage contract is a brand-new customer": this is wrong because it records a 100% gross loss in the subscription segment and a new logo in the usage segment for a customer who never went anywhere.

Your subscription GRR craters and your usage new-logo count inflates, both fictitiously. Treatment two — "ignore the model change, just follow the dollars": this is wrong because it puts contractual-ARR-baselined revenue and smoothed-usage revenue into the same cohort line, mixing two incompatible baseline methodologies and making the customer's retention uninterpretable.

The correct treatment is continuity-of-revenue with a segment hand-off, documented as a re-pricing not a churn. The customer's revenue is followed continuously — there is no gross loss event — but the cohort entry is *re-classified* from the subscription segment to the usage segment at the migration date, with the baseline carried forward.

Concretely: the customer's pre-migration contracted ARR becomes the *baseline* for their post-migration usage measurement (it is the cleanest available anchor — it is literally what they were worth before), and their post-migration consumption is measured against it. They exit the subscription cohort not as churn but as a flagged "migrated out" line, and they enter the usage cohort as a flagged "migrated in" line carrying their prior baseline.

Both flags are disclosed. This way the company-level blended retention is unaffected by the re-pricing (correct — no economic event occurred), the subscription segment's GRR is not falsely punished, and the usage segment's NRR is not falsely inflated. The footnote in board materials reads: "X customers representing \$Y of baseline migrated from subscription to usage pricing during the period; treated as continuity-of-revenue re-pricing, not churn or new logo." Auditors and sophisticated investors expect exactly this treatment, and a company that instead lets migrations show as churn-plus-new-logo will be asked, pointedly, to restate.

Leading Indicators For Usage-Customer Retention

One genuine *advantage* the usage segment has over the subscription segment is that its retention is far more *predictable in advance* — because consumption data is a continuous, high-frequency telemetry stream, whereas a subscription customer is a black box between renewal dates.

A subscription company often does not know a customer is unhappy until the renewal conversation; a usage company can see a customer cooling off in close to real time. Exploiting that requires building the right leading indicators, and they are different from the indicators a subscription business watches.

The consumption trend line is the master indicator. For every usage customer, maintain a rolling 30-day and rolling 90-day consumption curve and watch the *slope*, not the level. A customer whose 90-day slope has turned negative — even while their absolute spend still looks fine — is the earliest possible warning of contraction.

This is the "usage decay" signal: it shows up months before the retention metric itself moves, because the retention metric is a smoothed lagging measure and the slope is a leading one. Breadth-of-usage indicators matter alongside volume: a customer using three of your product's capabilities is structurally stickier than one using one capability at the same dollar level — single-feature usage is fragile because a single competitive displacement or internal re-architecture kills it.

Active-entity counts — how many of the customer's own seats, projects, environments, or pipelines are live on you — predict consumption durability better than the dollar figure does. Commitment utilization (for committed-spend customers) is, as covered earlier, the canonical early-warning overlay: a customer drifting below 60% utilization is telegraphing a downsell or churn one to two renewal cycles out.

And the "first 90 days" consumption ramp is sharply predictive of long-term retention — a usage customer who never reaches healthy consumption velocity in their first quarter is at elevated churn risk for their entire life, far more so than a subscription customer with a slow start.

The practical payoff for a mixed-model company: the usage segment, despite being the more *volatile* half, can actually be the more *forecastable* half if you instrument it properly — because you are not waiting for a renewal date to learn the truth. The discipline is to build a usage-health score from these indicators and to treat a deteriorating score as actionable *now*, not at renewal.

Subscription retention is mostly managed at renewal events; usage retention must be managed continuously, and the leading indicators are what make continuous management possible.

The Revenue Recognition Wrinkle: Reconciling Retention Math With How Finance Books It

There is a structural mismatch between how usage revenue is *recognized* under accounting rules and how the retention math above treats it, and a mixed-model CFO has to reconcile the two or the retention metrics will not tie to the financial statements — which is the fastest way to lose an audit committee's trust.

Subscription revenue is generally recognized ratably — a \$120K annual contract is booked at \$10K/month, smoothly, regardless of when the customer actually uses the product. Usage revenue under ASC 606 is generally recognized as consumed — you recognize it in the period the customer actually burns the API calls, the compute, the storage.

This means usage revenue on the income statement is *inherently lumpy*, matching the customer's consumption pattern, while subscription revenue is smooth. The retention math in this entry deliberately *smooths* usage revenue (the trailing-average methodologies) to extract a stable retention signal — which means your smoothed retention denominator/numerator will not equal the as-recognized revenue in any given month. That is expected and correct, but it has to be explained.

There are three specific reconciliation points to manage. First, the commitment/credit accounting: when a customer prepays for a credit pool or signs a committed-spend deal, the cash arrives and a contract-liability (deferred revenue) is created, then drawn down as consumed — so the *retention baseline* (the commitment) and the *recognized revenue* (the burndown) diverge within the year and only reconcile at the period boundary.

Your retention bridge needs a documented tie-out to the deferred-revenue rollforward. Second, overage and true-ups: usage above a commitment is often billed and recognized in arrears, sometimes a month or a quarter late, so the period in which expansion *shows up in retention* (when consumed) may differ from when it *shows up in billed revenue* (when invoiced).

Third, the measurement-period mismatch: if retention is measured on trailing-3-month smoothed figures but the financials are monthly as-recognized, you need a standing reconciliation schedule that walks from "as-recognized usage revenue" to "smoothed usage revenue used in retention," with the smoothing adjustment as an explicit, signed line.

The governing principle: the retention metrics and the financial statements are computed for *different purposes* — retention wants a stable behavioral signal, the financials want faithful period-by-period representation — and they are *allowed* to differ, but every difference must be a documented, repeatable reconciling item, never an unexplained gap.

What Investors Expect For Mixed-Model SaaS

The investor lens on mixed-model retention has matured considerably, and a CFO who understands what sophisticated investors are actually looking for can turn the reporting into a credibility asset rather than a source of suspicion. The headline expectation is simple and non-negotiable: segment transparency. An investor seeing a single blended retention number from a company they know is half-usage will assume the blend is hiding something — because, as established, it usually is.

The companies that earn analyst trust are the ones that *volunteer* the subscription-segment and usage-segment numbers, the baseline methodology, and the segment mix, before being asked.

What investors specifically want to see and why: (1) The two segment numbers, so they can value the two revenue streams differently. Usage revenue and subscription revenue do not deserve the same multiple — usage revenue is higher-growth but higher-beta and lower-visibility, subscription revenue is lower-growth but more predictable.

An investor cannot do that differential valuation if you only give them a blend. (2) The baseline methodology, stated explicitly. A sophisticated investor knows that "142% NRR" means nothing without knowing whether it is point-in-time, trailing-average, or commitment-based — and they will discount a number whose methodology is undisclosed, because an undisclosed methodology is assumed to be the most flattering one.

(3) Consistency over time. The single fastest way to lose investor trust is to change the retention methodology between reporting periods without restating — it reads as either incompetence or manipulation. (4) The NRR bridge / decomposition, so they can see whether high usage NRR is broad-based or whale-concentrated.

(5) GRR, not just NRR — increasingly investors lead with GRR for usage businesses precisely because NRR can be flattered by expansion while GRR exposes the true gross bleed; a usage company that only discloses NRR is now read as evasive.

The comparables context matters too: investors benchmark mixed-model companies against both pure-subscription peers and pure-consumption peers (Snowflake, Datadog, MongoDB, Confluent, Twilio), and they expect a mixed-model company to *bridge* the two — to show that its subscription segment behaves like the subscription comps and its usage segment behaves like the consumption comps.

A company that presents its retention this way is implicitly telling investors "we understand our own business well enough to measure each half on its own terms," and that is exactly the signal that supports a premium multiple. The reporting discipline is not overhead — it is part of the equity story.

The Measurement-Period Question: Why Usage Businesses Need Longer Or Smoothed Windows

A subtle but consequential issue: the *measurement period and cadence* that works for a subscription book is often wrong for a usage book, and a mixed-model company has to resolve the tension. Subscription businesses frequently measure retention on monthly cohorts and feel comfortable doing so, because subscription revenue is stable month to month — a monthly cohort's NRR for a subscription book is a reasonably steady series.

Run monthly cohorts on a *usage* book and you get a wildly oscillating series, because each monthly cohort's baseline is anchored to whatever the usage happened to be in that month, and usage's natural variance whips the metric around. A usage business that reports a monthly-cohort NRR series will appear to have a retention problem that is actually just a measurement-cadence problem.

The resolutions, in order of preference: (1) Use quarterly cohorts for the usage segment even if you use monthly for subscription — a quarter is long enough that the cohort-start baseline is itself an average of three months, which damps the anchoring noise. (2) Use trailing-twelve-month measurement windows rather than point-in-time comparisons — comparing trailing-12 to trailing-12 is the most stable read available for a noisy book and is the methodology most consumption public companies effectively report on.

**(3) If you must report monthly for the usage segment, report it as a trailing-3-month or trailing-12-month *smoothed* series, never raw monthly-cohort points — and label it clearly as smoothed. (4) Match the measurement period to the customer's natural usage cycle** — if your customers have a strong annual seasonal pattern, anything shorter than a 12-month window will confuse seasonality for performance.

The deeper point for a mixed-model company is that you may legitimately run *two different measurement cadences* — monthly cohorts for the subscription segment, quarterly or trailing-twelve for the usage segment — and that is acceptable *as long as the blended number is computed on a common, clearly stated period* and the difference in segment cadences is disclosed.

What is not acceptable is forcing both segments onto the cadence that suits the subscription half and then being surprised that the usage half's retention looks chaotic. The volatility of usage revenue is a property of the revenue, not a defect in the company — the measurement period has to be chosen to see through that volatility to the underlying retention signal, and that almost always means *longer or smoothed* windows for the usage segment than the subscription segment requires.

Public-Company Examples: How Mixed And Usage-Model Companies Disclose Retention

It is worth grounding all of this in how actual public companies have handled usage-model retention disclosure, because their definitional choices illustrate the trade-offs and because investors will benchmark a private mixed-model company against them.

Snowflake is the canonical consumption-NRR company. It reports a net revenue retention rate that has historically been among the highest in software (peaking well above 160% in its hyper-growth phase, compressing toward the 120s-130s as the business matured and the 2022-2023 optimization wave hit).

Its definition is explicitly trailing-and-cohort-based: it measures the revenue of a cohort of customers in a trailing period against the same cohort's revenue in the prior comparable trailing period — a trailing-twelve-style methodology, precisely because point-in-time would be unreadable for a pure-consumption book.

Snowflake notably does *not* lean on a logo-churn GRR narrative; the consumption model makes logo churn a weak signal, and the company steers attention to NRR and the cohort revenue curves.

Datadog is a hybrid in the truest sense: customers carry both subscription-like product commitments and usage-driven expansion across many products. Datadog reports NRR (consistently above 120% through most of its public life) and emphasizes *multi-product adoption* as the durability story — the breadth-of-usage indicator from the leading-indicators section, made into an investor narrative.

MongoDB discloses Atlas (consumption) separately from its legacy subscription/license revenue, exactly the segment-transparency practice this entry argues for — letting investors see the consumption engine on its own terms. Twilio is the cautionary example: a heavily usage-based communications business whose NRR (its "dollar-based net expansion rate") swung significantly with customer consumption patterns and a few large customers, illustrating both the high-beta nature of consumption NRR and the whale-concentration risk that the NRR decomposition is meant to expose.

Confluent straddles a subscription (Confluent Platform) and consumption (Confluent Cloud) split and discloses NRR while walking investors through the cloud-consumption dynamics separately.

The common thread across all of them: the credible disclosers *name their methodology*, *separate consumption revenue from subscription revenue* where both exist, *lead with cohort/trailing definitions rather than point-in-time*, and *contextualize volatility* rather than hiding it.

A private mixed-model company designing its own retention disclosure should treat these as the template — and should expect its investors to ask why, if it deviates from them.

Building The Mixed-Model Retention Dashboard

Translating all of this into an operating system means building a retention dashboard with a specific architecture — and the architecture is materially different from a single-pricing-model dashboard. The governing principle: everything is segmented, and the segmentation is the default view, not a drill-down.

The dashboard has three parallel column groups — subscription segment, usage segment, blended company — each showing GRR, NRR, the cohort baseline revenue, and the period-over-period trend. The blended group is visually *last and least emphasized*, a deliberate design choice that reinforces the reporting discipline.

Within the usage segment, the dashboard exposes the things subscription dashboards never need: the baseline methodology in use (stated on the dashboard itself, not buried in documentation), the NRR decomposition bridge (starting cohort → expansion → contraction → churn → ending), the commitment-utilization distribution for committed-spend customers, and the customer-classification census (how many customers are stable / in-variance / trending-down-unconfirmed / confirmed-contraction / confirmed-cliff).

The leading-indicator layer sits alongside: the count of usage customers with a negative 90-day consumption slope, the breadth-of-usage distribution, the first-90-day ramp cohort health.

Two infrastructure requirements make or break the dashboard. First, billing-system reconciliation across two revenue types. A mixed-model company very often has the two revenue streams in two different systems — a subscription-management/CPQ system for contracted ARR and a metering/rating system for consumption — and the dashboard is only trustworthy if there is a hard, scheduled reconciliation that ties both back to the general ledger and to each other.

The single most common cause of a mixed-model retention number being *wrong* is not a methodology error; it is a billing-data integration error where usage events are double-counted, dropped, or mis-timed. Second, the baseline-methodology documentation has to be a living artifact attached to the dashboard — a versioned document stating exactly how each segment's baseline is computed, what the smoothing windows are, how edge cases (cliffs, migrations, hybrids) are handled, and the change log of any methodology revisions with their restatement impact.

A dashboard without that attached documentation is a dashboard that cannot survive an audit or a diligence process. Build the reconciliation and the documentation first; the visualizations are the easy part.

The Forecasting Implication: Usage Retention Is Harder To Forecast, So Forecast It Differently

Retention metrics are not just a scorecard — they feed the revenue forecast, and the usage segment's retention behavior forces a different forecasting approach than the subscription segment's. A subscription book is forecastable largely from its *contract structure*: you know the renewal dates, you know the contracted amounts, you apply historical renewal and expansion rates by cohort, and you get a defensible forward number.

The contracts themselves carry most of the forecast.

The usage segment has no equivalent contractual scaffolding (except for the committed-spend portion), so forecasting it from "contracts" is impossible — the revenue is whatever customers consume, and that depends on the customers' own businesses. This means usage-segment forecasting has to be leading-indicator-driven and probabilistic rather than contract-driven and deterministic.

The right approach: build the usage forecast bottom-up from the consumption-trend signals — the rolling 90-day slopes, the active-entity growth, the commitment-utilization trajectories, the first-90-day ramp health — and express it as a *range with explicit scenarios* (a base case, an optimization-pressure downside, an expansion upside) rather than a single point.

The committed-spend portion of the usage book *can* be forecast more deterministically — the commitments are contractual floors — so a sophisticated usage forecast often splits into "committed floor (high confidence) + consumption-above-commitment (probabilistic range)."

The consequence for a mixed-model company's overall forecast is that the two halves carry different *confidence intervals*, and the combined forecast must reflect that. The subscription half is a relatively tight distribution; the usage half is a wide one, and macro-sensitive — usage forecasts have to carry an explicit assumption about the customer base's own growth and budget environment in a way subscription forecasts do not.

Presenting a single point-estimate company revenue forecast that buries a wide-distribution usage segment inside a tight-distribution subscription segment gives the board false precision. The honest forecast says: "the subscription segment will land within this narrow band; the usage segment will land within this wider band, driven by these consumption assumptions; here is the combined range." That structure also makes the forecast *diagnosable* when it misses — you can see immediately whether the miss was subscription (a contract/renewal issue) or usage (a consumption issue), which a blended point estimate can never tell you.

Board Reporting For Mixed-Model Retention

Board reporting is where all of the methodology either pays off or unravels, and the standard for a mixed-model company is more demanding than for a single-model company — not because boards want more numbers, but because the wrong presentation actively misleads a board into the wrong strategic conclusions.

The non-negotiable structure: three numbers, every time, in the same format, every period. Subscription segment GRR/NRR, usage segment GRR/NRR, blended company GRR/NRR — with the segment revenue mix stated alongside. A board that sees only the blend cannot do its job; it will either over-worry about a healthy usage volatility or, worse, miss a structural gross-retention problem in half the company because expansion is masking it in the blend.

The three-number format is what lets the board reason correctly.

Defend the baseline methodology proactively, once, and then hold it. The first time a mixed-model company presents usage retention to its board, it should walk through the baseline methodology explicitly — why trailing-3-month average (or commitment-based, or whatever was chosen), what it smooths, what it lags, why the alternative methodologies were rejected.

After that initial defense, the methodology becomes a stated constant, and the only methodology content in subsequent board meetings is *change disclosure* — and changes should be rare and always accompanied by a restatement of prior periods so the board can see the change isolated from performance.

A board that watches a company's retention methodology drift loses confidence in every other number management presents.

Show the decomposition, contextualize the volatility, and connect retention to the forecast. The usage NRR should always come with its bridge so the board can see whether expansion is broad or whale-driven. The usage segment's volatility should be *explained as a property of the model*, not apologized for as a performance issue — a board that understands consumption revenue is high-beta will not panic at a normal quarter of optimization pressure, and a board that does not understand it will.

And the retention numbers should tie visibly to the revenue forecast: "usage NRR is trending here, the leading indicators say this, therefore the usage revenue forecast is this range." The board reporting standard, in one sentence: consistent three-number format, methodology stated and held constant, decomposition always shown, volatility contextualized as structural, and a clear line from retention through leading indicators to the forecast range.

Five Real-World Scenarios

Scenario 1 — The 50/50 subscription-usage org. A data-infrastructure company with \$40M ARR-equivalent, roughly half from platform subscriptions and half from metered consumption. The trap: a single blended NRR of 118% that the board has been happy with for two years. On segmentation, the subscription half is 99% GRR / 106% NRR (healthy, boring) and the usage half is 80% GRR / 131% NRR — meaning the usage book is bleeding 20 points of gross revenue, masked entirely by expansion.

The fix: segment the reporting, expose the usage GRR problem, and discover it is a slow-bleed concentrated in customers who never reached healthy first-90-day consumption velocity. The blend was concealing the single most important operational issue in the company.

Scenario 2 — A subscription company adding a usage tier. A mature subscription business launches a consumption-priced add-on. For the first year the usage revenue is 6% of total. The right move is *not* to build a full parallel measurement system — it is to footnote the usage piece, measure it simply, and keep the headline retention metrics as subscription metrics with a disclosed usage footnote.

Building the elaborate machinery for a 6% tail is the over-engineering counter-case. The trigger to build the real parallel system is when usage crosses ~15-20% of revenue.

Scenario 3 — A usage company adding subscription commitments. A pure-consumption company starts requiring committed-spend contracts. This is a *gift* to the retention methodology — it converts the noisy trailing-average baseline into a clean commitment baseline for the committed portion.

The right move is to migrate the committed customers to Methodology C (run-rate against commitment) plus commitment-utilization as the early warning, while keeping uncommitted customers on trailing-average. The company now runs a cleaner, more forecastable usage book — and should disclose the methodology split.

Scenario 4 — A seasonal-usage business. A consumption company whose customers are e-commerce platforms — usage spikes every Q4 and troughs every Q1. Monthly cohorts and point-in-time baselines make the retention metric look like a sine wave. The fix: trailing-twelve-month windows and quarterly (or annual) cohorts only, so seasonality is *inside* every measurement window and cannot be confused for performance.

The classification system explicitly knows the seasonal pattern and does not flag the Q1 trough as contraction.

Scenario 5 — A customer mid-migration between models. A \$300K-ARR subscription customer re-papers onto usage pricing in month seven of the cohort year. Handled naively, this is a \$300K subscription churn plus a usage new logo — fictional on both counts. Handled correctly, it is a continuity-of-revenue re-pricing: the customer exits the subscription cohort flagged "migrated out," enters the usage cohort flagged "migrated in" carrying the \$300K as their usage baseline, the blended company retention is unaffected, and the board materials footnote the migration.

No phantom churn, no phantom logo.

The Decision Framework

Pulling the entire methodology into an operating sequence, the mixed-model retention computation follows six steps, in order, every period:

1 — Segment first, always. Split every revenue dollar into the subscription book or the usage book *before* any retention math happens. Decompose hybrid customers into two cohort entries by revenue stream. Never start with a blended population.

2 — Pick the baseline methodology per segment and freeze it. Subscription: contracted ARR at cohort start (given). Usage: choose point-in-time (rarely), trailing-3-or-12-month average (the default), or run-rate-against-commitment (cleanest where commitments exist). Document the choice; do not change it without restating history.

3 — Smooth usage variance before classifying. Apply the 3-6 month smoothing window and the step-change test. Do not classify any usage revenue movement as contraction or churn until the window confirms it is structural, not seasonal. Accept that the most recent months are provisional.

4 — Compute GRR and NRR per segment, on the same cohort and baseline. Usage GRR measures the sustained floor, capped at 100%. Usage NRR measures ending revenue, expansion uncapped. Subscription GRR/NRR by the standard contractual method. Keep cohort timing aligned across both segments.

5 — Revenue-weight into the blend, and report all three. Blended GRR/NRR is the revenue-weighted (or bottom-up) combination of the two segments. Always present subscription, usage, and blended — never the blend alone — with the segment mix disclosed.

6 — Document the methodology and hold it constant. Attach a versioned methodology document to the dashboard and the board materials. Reconcile both revenue streams to the GL. Treat any methodology change as a restatement event. Consistency is the metric's credibility.

Five-Year Outlook

The defining trend for this entire topic is that usage and hybrid pricing are becoming the default, not the exception — which means the "mixed-model retention problem" is on its way to being simply "the retention problem." Across infrastructure, data, developer tools, AI/LLM products, communications, and increasingly application software, the industry has been migrating from pure-seat subscription toward consumption and hybrid models, and the AI wave is accelerating it hard — AI products are almost inherently usage-priced because their cost-to-serve is consumption-driven (tokens, inference, compute).

A growing share of SaaS companies will be mixed-model within five years, and a material share will be majority-usage. The implication: the methodology in this entry stops being a specialist concern for a subset of companies and becomes baseline financial competence for SaaS finance teams generally.

That ubiquity is driving a second trend: emerging standardization of usage-retention disclosure. Today there is real definitional inconsistency — companies report NRR on point-in-time, trailing-3, trailing-12, and commitment bases and often do not say which. Over the next five years, expect investor pressure, analyst frameworks, and possibly accounting/disclosure guidance to converge on a more standardized set of definitions — most likely centered on trailing-twelve-month cohort methodologies and on mandatory segment disclosure for mixed-model companies.

The companies that adopt rigorous, transparent, segmented retention reporting *now* will be ahead of that curve; the ones reporting a single undisclosed-methodology blend will be forced to catch up under scrutiny.

A third development: tooling will mature. Today, computing rigorous mixed-model retention often requires custom data engineering because billing systems, metering systems, and BI tools were built for subscription logic. Expect the revenue-platform and analytics vendors to build native mixed-model retention computation — segmented cohorts, smoothing windows, commitment baselines, the NRR bridge — as standard functionality, which lowers the barrier for smaller companies to do this correctly.

Finally, expect GRR to overtake NRR as the headline usage metric. As investors get more sophisticated about consumption models, they increasingly recognize that NRR flatters a usage business (expansion is free and automatic) while GRR exposes the truth about durability — so the disclosure emphasis is shifting toward gross retention as the metric that actually distinguishes a healthy consumption business from a leaky one.

Final Framework

The mixed-model retention computation, reduced to its load-bearing principles:

The blueprint. Usage revenue has no contractual baseline, so the baseline must be *constructed* — and because it is a construction, it must be explicit, documented, and constant. Segment the book into subscription and usage before any math. Decompose hybrid customers by revenue stream, not by assigning the whole customer to one segment.

Compute GRR and NRR separately per segment on aligned cohorts. Revenue-weight into a blend, and report all three numbers — subscription, usage, blended — with the mix disclosed.

The baseline-methodology decision guide. Subscription baseline is contracted ARR — given, not chosen. Usage baseline is a choice between three: point-in-time snapshot (only for genuinely stable usage or fast directional reads — it overstates both churn and expansion), trailing-3-or-12-month average (the default for most usage businesses — smooths noise, lags real cliffs), and run-rate-against-commitment (the cleanest, available only when customers carry committed spend — pair it with commitment-utilization as the early warning).

Pick one per segment, document it, never change it without restating.

The dashboard spec. Three parallel segment column groups with the blend last and least emphasized. For the usage segment, expose the baseline methodology, the NRR decomposition bridge, the commitment-utilization distribution, and the customer-classification census. Add the leading-indicator layer — negative-90-day-slope counts, breadth-of-usage, first-90-day ramp health.

Build the two-revenue-type billing reconciliation and the versioned methodology documentation first; the visualizations are secondary.

The board-reporting standard. Three numbers every period in an unchanging format, segment mix stated alongside. Defend the methodology once, then hold it constant and only ever disclose *changes* — always with restatement. Always show the NRR decomposition.

Contextualize usage volatility as a structural property of the model, not a performance defect. Draw a visible line from retention through leading indicators to a ranged forecast.

The discipline in one sentence: a mixed-model company runs two retention businesses under one logo — measure each on its own terms, combine them deliberately, disclose all three numbers, and never let a blended average conceal what is happening in half the company.

The Mixed-Model Retention Computation Flow

The Usage-Variance Versus Real-Contraction Decision Tree

Sources

1. ASC 606 — Revenue from Contracts with Customers (FASB) — Governs recognition of usage/consumption revenue as-consumed versus subscription revenue ratably; the basis for the revenue-recognition reconciliation. https://www.fasb.org
2. Snowflake Inc. — SEC 10-K and Investor Materials (NYSE: SNOW) — Net revenue retention rate definition and trailing-cohort methodology for a pure-consumption model. https://investors.snowflake.com
3. Datadog Inc. — SEC 10-K and Investor Materials (NASDAQ: DDOG) — Hybrid subscription-plus-usage retention disclosure and multi-product adoption narrative. https://investors.datadoghq.com
4. MongoDB Inc. — SEC 10-K and Investor Materials (NASDAQ: MDB) — Atlas consumption revenue disclosed separately from subscription/license revenue. https://investors.mongodb.com
5. Twilio Inc. — SEC 10-K and Investor Materials (NYSE: TWLO) — Dollar-based net expansion rate for a usage-heavy communications business; volatility and whale-concentration illustration. https://investors.twilio.com
6. Confluent Inc. — SEC 10-K and Investor Materials (NASDAQ: CFLT) — Confluent Platform (subscription) versus Confluent Cloud (consumption) split disclosure. https://investors.confluent.io
7. Bessemer Venture Partners — State of the Cloud and the "10 Laws of Cloud" — Industry benchmarks and frameworks for NRR/GRR, including usage-model retention dynamics. https://www.bvp.com
8. OpenView Partners — SaaS Benchmarks and the Usage-Based Pricing Report — Multi-year research on the shift to usage-based pricing and its retention-measurement implications.
9. KeyBanc Capital Markets — Annual SaaS Survey — Cross-company benchmarking of GRR, NRR, and pricing-model mix for private SaaS.
10. a16z — "The New Business of AI" and consumption-pricing analyses — On why AI and infrastructure products trend toward usage pricing and the resulting retention-metric behavior. https://a16z.com
11. SaaS Capital — Retention and Pricing Research — Private-company GRR/NRR benchmarks and the relationship between pricing model and retention volatility.
12. Snowflake 2022-2023 "optimization" disclosures — Public commentary on consumption compression during the macro budget-tightening cycle, illustrating usage-NRR macro sensitivity.
13. AICPA — Revenue Recognition Audit and Accounting Guide — Practitioner guidance on contract-liability (deferred revenue) treatment for prepaid credits and committed-spend arrangements.
14. Klipfolio / ChartMogul / Maxio (SaaSOptics) — Subscription and Usage Analytics Documentation — Tooling-side definitions of cohort-based GRR/NRR and the gaps for usage models.
15. Battery Ventures — Software Retention Benchmarks — Comparative GRR/NRR data across subscription, hybrid, and consumption companies.
16. Meritech Capital — Public SaaS Comparables and Retention Analyses — Net revenue retention comparables across the consumption-model public cohort (Snowflake, Datadog, MongoDB, Confluent).

Numbers

The Baseline-Methodology Spread (illustrative single customer)

• Monthly usage over the year: \$8K, 9K, 7K, 6K, 8K, 10K, 12K, 9K, 7K, 8K, 11K, 14K
• Cohort-start month: \$8K → last-month-annualized baseline \$96K
• Trailing-3-month avg at start: ~\$8K → annualized ~\$96K
• Full-year average: ~\$9.1K → annualized ~\$109K
• Contracted minimum: \$7K/month → \$84K/year
• Current run-rate: \$14K/month → \$168K annualized
• NRR vs last-month-annualized: ~175%
• NRR vs full-year-avg-annualized: ~154%
• NRR vs commitment: ~200%
• Spread across three legitimate methodologies: ~46 points, same customer, same behavior

Typical Retention Ranges

• Healthy subscription GRR: ~88-95%
• Healthy subscription NRR: ~105-120%
• Healthy usage GRR: ~78-92% (structurally lower — no renewal gate)
• Usage NRR for strong consumption businesses: ~115-160%+ in growth phases
• Consumption-sector NRR compression in 2022-2023 optimization wave: from 160%+ toward ~120% and below within a few quarters
• Common overstatement of usage GRR when only counting logo cancellations: 15-25 points too high

Smoothing And Classification Windows

• Standard usage smoothing window: 3-6 months
• Provisional (unconfirmed) retention period: most recent ~2-3 months, revised as window matures
• Trailing-average methodology options: trailing-3-month (responsive) or trailing-12-month (most stable)
• Recommended usage cohort cadence: quarterly or trailing-twelve (not monthly)

Commitment Utilization Bands

• Healthy: ~85-110% of commitment consumed
• Watch: ~60-85%
• Alarm: under ~60% (leading indicator of downsell/churn 1-2 renewal cycles out)

Blending Mechanics (illustrative)

• Subscription segment baseline: \$30M → weight 60%
• Usage segment baseline: \$20M → weight 40%
• Subscription: 96% GRR / 104% NRR; Usage: 82% GRR / 128% NRR
• Blended GRR: (0.60 × 96%) + (0.40 × 82%) ≈ 90.4%
• Blended NRR: (0.60 × 104%) + (0.40 × 128%) ≈ 113.6%
• Note: the blend conceals an 18-point gross bleed in half the company

Over-Engineering Threshold

• Usage tail below ~15-20% of revenue: footnote it, do not build a parallel measurement system
• Usage crossing ~15-20% of revenue: build the full segmented system

Leading Indicators

• Master signal: rolling 90-day consumption slope (turns negative months before retention metric moves)
• First-90-day consumption ramp: sharply predictive of full-life retention
• Breadth-of-usage: multi-capability customers structurally stickier than single-feature at same dollar level

Counter-Case: When The Elaborate Mixed-Model Methodology Is Overkill Or Actively Misleading

The entire framework above assumes the usage segment is large enough and real enough to justify a parallel measurement system. For a meaningful number of companies, that assumption is false, and applying the full methodology is not rigor — it is waste, false precision, or outright distraction.

A serious CFO should pressure-test whether this problem is actually their problem.

Counter 1 — The 95%-subscription company with a tiny usage tail should not build any of this. If usage revenue is 3-7% of the total — a subscription company that bolted on a small consumption add-on, an overage line that a handful of customers trip — building segmented cohorts, smoothing windows, commitment-utilization dashboards, and a parallel board-reporting track is straightforwardly over-engineering.

The right treatment is a footnote: "X% of revenue is usage-based; it is measured on a trailing-3-month basis and included in the blended figures; segment detail available on request." The headline metrics stay subscription metrics. The trigger to build the real system is when usage crosses roughly 15-20% of revenue and starts to materially move the blend — below that, the machinery costs more in analyst time and board attention than the insight it produces.

Many companies build the elaborate system years too early because it feels sophisticated, and it just adds reporting overhead nobody reads.

Counter 2 — Methodological purity on a genuinely noisy cohort produces false precision, not insight. There is a real failure mode where a finance team, having internalized that the usage baseline is a choice, spends enormous effort debating point-in-time versus trailing-3 versus trailing-6 versus commitment-based, builds elaborate reconciliations between them, and reports usage NRR to a tenth of a percent — for a cohort of 40 customers whose individual month-to-month variance is ±30%.

The underlying data simply does not support that precision. The honest treatment of a small, noisy usage cohort is a *range with wide error bars and a stated low-confidence flag*, not a precise-looking point estimate. Presenting "usage NRR was 123.4%" off a 40-customer cohort with ±30% individual variance is false precision that will mislead the board into treating a noise-dominated number as a signal.

Sometimes the rigorous answer is "we cannot measure this precisely yet, here is the range, here is when the cohort will be large enough to measure properly."

Counter 3 — The segmentation can become the distraction that hides the real problem. The most dangerous version of the counter-case: a company spends two quarters building the perfect segmented retention dashboard, debating baseline methodologies, refining the smoothing windows — and the entire exercise is *avoidance*.

The real issue is not that the retention is hard to *measure*; it is that the usage business model is not actually retaining customers. The customers are not sticking. The product does not create durable consumption.

No methodology refinement changes that. When a team is on its third baseline-methodology revision in a year, the question to ask is not "which methodology is most accurate" but "are we using methodology debates to avoid confronting that the usage GRR is genuinely bad and the model needs fixing, not the metric." Measurement sophistication can become a sophisticated way of not looking at the answer.

Counter 4 — Two measurement systems is real organizational cost. Running segmented cohorts, two baseline methodologies, two cadences, the reconciliation between two billing systems, the versioned methodology documentation, and the three-number board format is not free. It consumes finance-team capacity, it requires data engineering, it creates more surface area for errors, and it demands ongoing maintenance.

For a company where the usage segment genuinely warrants it, that cost is worth paying. For a company on the margin, the cost can exceed the value — and the team would be better served by a simpler, slightly-less-precise approach that they can actually maintain correctly than by an elaborate system they maintain badly.

A correctly-computed simple metric beats a sophisticated metric riddled with reconciliation errors.

Counter 5 — Sometimes the blend really is fine. The entry argues hard against blend-only reporting, and that is the right default — but there is a narrow case where it is acceptable: when the two segments genuinely behave similarly (similar GRR, similar NRR, similar volatility), the blend is not concealing anything, and the segment split is not decision-relevant.

This is uncommon — subscription and usage retention physics usually differ — but it does happen, particularly in early-stage hybrid companies where the usage book is still small and behaving like the subscription book. The test is empirical: compute the segments, and *if* they are genuinely close and stable, you can lead with the blend and keep the segments as a footnote.

The rule is "blend last, never first" — but the deeper rule is "don't hide a difference that exists," and if the difference genuinely does not exist, the elaborate separate presentation is just ceremony.

The honest verdict. The full mixed-model retention methodology is essential for any company where usage revenue is a material, structurally-distinct share of the business — roughly 15-20%+ and behaving differently from the subscription book. Below that threshold, footnote it and move on.

And at every threshold, watch for the two pathologies: false precision on noisy small cohorts, and methodology debates standing in for confronting a real retention problem. The methodology is a tool for *seeing the business clearly* — when it stops doing that and starts being a substitute for hard answers, it has become the problem it was meant to solve.

Measure as rigorously as the data supports and the materiality justifies — no more, and no less.

Related Pulse Library Entries

• q97 — How do you compute gross revenue retention versus net revenue retention? (The general GRR-vs-NRR computation; this entry is the usage-based / mixed-model complication on top of it.)
• q1899 — What replaces SDR teams if AI agents replace SDRs natively? (Sales-motion restructuring — relevant to why consumption expansion needs no sales motion.)
• q9501 — How do you start a bookkeeping business in 2027? (Finance-operations adjacency.)
• q9502 — How do you start a CPA firm in 2027? (Revenue-recognition and audit-readiness adjacency.)
• q9601 — How do you start a fractional CFO business in 2027? (The CFO-level reader of this entry.)
• q9602 — How do you start an outsourced controller business in 2027? (The controller who builds the reconciliation infrastructure described here.)
• q9603 — How do you start a tax preparation business in 2027? (Revenue-recognition adjacency.)
• q9604 — How do you start a financial advisor business in 2027? (Equity-story and valuation adjacency.)
• q9701 — What is the best practice management software for bookkeeping firms? (Tooling adjacency for finance-ops systems.)
• q9705 — How do you prep tax packages for CPAs? (Reconciliation-discipline adjacency.)
• q9801 — What is the future of bookkeeping in 2030? (Long-term finance-ops outlook context.)
• q9802 — How will AI change bookkeeping by 2030? (AI-driven shift toward consumption pricing context.)