GRR Dropped to 84%: Instrument the SMB Churn Bleed

The retention number most SaaS teams planned around just moved. The headline from one widely shared 2026 benchmark is blunt — "Your 88% GRR Benchmark Is Gone. The New Median Is 84%." — and it describes the largest single-year drop in gross revenue retention in four years. At the same time, net revenue retention held roughly flat at 102–106%. That gap is the whole story: the top-line number that boards watch looks fine, while the number underneath it is bleeding.

This is not a finance post. It is a measurement post. The GRR decline is concentrated in a segment, driven by a cause that shows up in product usage long before it shows up in a renewal, and invisible in any metric that blends segments or counts logos instead of revenue. If your analytics cannot separate the SMB cohort, read gross retention apart from expansion, and surface the usage signals that precede a cancellation, you will learn about this bleed from the renewals report — which is the one place it is already too late to act on it.

The number: 84% median, and where the four points went

Median GRR for B2B SaaS now sits around 84%, down from 88% a year earlier, with the top quartile near 91% and the bottom quartile near 76%. On its own, a four-point move in a median looks like noise. It is not, because the movement is not spread evenly across the customer base — it is concentrated almost entirely in small accounts. The H1 2026 retention data puts realistic gross dollar retention for sub-$25K-ACV SaaS in a 70–80% band, far below the blended median.

The segment split matters because it changes what the number means for you specifically. If you sell mostly to enterprise, your GRR may not have moved at all; if you sell mostly to SMBs, your gross retention may have fallen well past the four-point headline and the median is actively flattering you. A single company-wide GRR figure is an average over segments that are now behaving very differently, and averaging them together is exactly how the problem hides. The first instrumentation job is therefore not to compute a better company GRR but to stop computing a single one — to split it by ARR tier so the SMB line can be read on its own.

Why NRR hides the bleed

Net revenue retention adds expansion on top of what is retained, so a strong upsell motion in large accounts can hold blended NRR near 102–106% while gross retention erodes beneath it. The blend also spans segments that diverge sharply — enterprise NRR around 118%, mid-market around 108%, SMB around 97% — so the aggregate can look healthy while one tier deteriorates. A board watching NRR alone sees a flat, reassuring line and no reason to investigate.

This is a specific instance of a general measurement failure: an aggregate that averages away the signal inside it. It is the same shape as Simpson's paradox in aggregate metrics, where a trend reverses or disappears once you split by segment. The defence is to never report retention as one number. Show GRR and NRR side by side, each split by ARR tier, so expansion in the enterprise cohort cannot visually cancel contraction in the SMB cohort. The discipline of reading gross and net separately is the same one behind measuring churn as logo versus revenue, gross versus net — and in 2026 it is the difference between seeing the problem and being surprised by it.

The cause lives in usage data, not the invoice

The structural reason SMB retention is re-rating is that AI copilots have collapsed the pricing power of thin, single-purpose tools. When a general copilot inside a product a team already pays for can approximate what a standalone workflow tool did, the small-team buyer cancels and folds the job into existing software. The effect is heaviest on products that were, in the honest framing of the 2026 benchmark reports, a thin layer of logic over a database — exactly the shape a copilot substitutes most easily.

That substitution is observable in product data before it reaches the renewal. An account being hollowed out by a copilot does not usually go dark all at once; it narrows. The specific workflows a copilot can replace go quiet while the rest of the product holds, active-usage frequency drifts below the account's own baseline, and the breadth of features touched contracts toward one or two surfaces. These are leading indicators, and they are the kind of signal that moves before revenue does — the category covered in leading indicators of growth beyond MRR and NRR. The invoice is a lagging indicator of a decision the usage data already described.

Instrument account-level usage, not just user-level

Detecting this bleed requires measuring at the account level, because the unit that churns is the account, not the individual user. A product analytics setup that only tracks anonymous or user-level events can tell you that usage fell somewhere but not which accounts are contracting, and the accounts are what renew. You need to roll user events up to the account, carry the account's ARR tier and vertical as properties, and compute engagement against each account's own historical baseline rather than a global average that a few power accounts dominate.

The build is concrete. Attach an account identifier and ARR tier to every event as it enters analytics; derive a small set of account health signals — active days per period, distinct features used, active seats, and usage of the specific surfaces most exposed to copilot substitution; and compare each to the account's trailing baseline so a drop registers as a drop, not as a number that happens to be below someone else's. This is account-level product analytics, the distinction drawn out in account-level versus user-level analytics, and it is the substrate every downstream renewal-risk view depends on.

Build the renewal-risk cohort

With account-level signals in place, the deliverable is a renewal-risk cohort: the set of accounts whose usage is trending toward cancellation, surfaced weeks before the renewal date while intervention is still possible. Define it from the leading indicators — accounts below a threshold on active-usage frequency against their baseline, contracting in feature breadth, or shedding seats — and weight it by the revenue at stake so the list is ordered by what you would actually lose. Build the cohort on recurring revenue, not logo count, so a downgrade from a larger plan registers as risk rather than waiting for a full cancellation.

The cohort is only useful if it is genuinely predictive, which means validating it against accounts that actually churned rather than trusting that the signals feel right. The honest version of this work ties into predictive churn analytics and its leading indicators, where the discipline is to check whether your early-warning signals preceded real losses and at what false-positive rate. A risk cohort that flags half the base warns no one; a tight one, ordered by revenue at stake and refreshed weekly, turns the GRR line from a quarterly autopsy into a list a success team can work before the renewal.

What to run this sprint

Start by splitting one number. Take your current company-wide GRR and break it into three lines by ARR tier — sub-$25K, $25K–$100K, $100K+ — on recurring revenue, not logos. In most SaaS businesses selling across segments, the SMB line will be visibly worse than the blended median you have been reporting, and seeing that gap is the point: it tells you whether the 2026 re-rating is your problem or someone else's. If the SMB line is in the 70–80% band, the macro story is now your story.

Then instrument three account-level signals you can ship this week — active-usage frequency versus baseline, feature breadth, and active seats — and stand up a crude renewal-risk list ordered by revenue at stake. It does not need to be a model yet; a ranked list from three honest signals beats a renewals report that arrives after the decision. Pair it with the cohort discipline in acquisition versus behavioural cohort analysis so you are reading retention by how accounts behave, not just when they signed. The 88%-to-84% move is a benchmark headline; whether it becomes a number you watch bleed or one you work is a question of instrumentation.

Frequently asked questions

The questions below cover the 2026 GRR benchmark figures, why gross retention fell while NRR held, how a healthy aggregate hides a deteriorating segment, the product-usage signals that predict SMB churn, and how to instrument gross revenue retention by segment.