TL;DR
- The Problem: Two credible 2026 datasets report median net revenue retention twenty points apart.
- The Cause: Sample, method and cohort window, not broken math.
- The Leak: Median gross retention sits at 84%, so expansion carries eighteen points of loss.
- The Fix: Decompose NRR by segment and cohort before you compare it to anything.
- The Move: A five-step diagnostic with one owner per leak.
Two benchmark reports crossed my desk in the same week this fall. Both were about net revenue retention.
One put the median at 102%. The other put it at 82%. Both had real data behind them and both were published this year. Neither was lying to me. That gap is the whole problem with how B2B teams use a retention benchmark.
Ten minutes here will change how you present the number next quarter.
Two 2026 Benchmarks, Twenty Points Apart
The first is the 2026 Aleph x Benchmarkit SaaS and AI Performance Benchmarks. It covers 342 SaaS and AI-native companies on full-year 2025 data. Median net revenue retention lands at 102%, with a top quartile of 110%. The bottom quartile sits at 92% and best-in-class starts at 120%.
The second is ChartMogul’s SaaS Retention Report, built with analyst Kyle Poyar. It scraped roughly 3,500 software companies and measured retention from billing data. Among about 2,700 B2B SaaS companies above $250k ARR, median NRR is 82%. The upper quartile reaches 97%. No survey respondents, no opt-in list.
Same metric, same year, twenty points apart.
The gap comes from sample construction. Aleph surveys companies that opted into a benchmark program. That population skews toward businesses with a finance function and a reason to compare themselves. ChartMogul measures every company whose billing data it can reach. That pulls in a long tail that would never answer a survey. One number describes the companies that care about benchmarks, and the other describes the market.
Method moves the number too. A twelve-month retained cohort of existing accounts produces a different figure than an annualized rate computed month by month. Let new-logo revenue leak into the expansion term and the drift grows again.
So when a marketing leader tells me their NRR is 96% and asks whether that is good, the question is malformed. It has no answer without a segment and a cohort attached.
I stopped quoting a single benchmark in client work for this reason. The number I use depends on segment, on pricing model, and on whether the account was sold a use case that survived production. Two companies with identical blended retention can be running opposite businesses. That is what a benchmark hides when it gets printed as one figure.
The Blended Number Hides Three Leaks
Net revenue retention is a ratio. Starting ARR plus expansion, minus contraction, minus churn, divided by starting ARR. One number falls out of four moving parts. The parts move for entirely different reasons.
The number that belongs beside it is 84%.
That is median gross retention in the same Aleph dataset. Gross retention excludes expansion and caps at 100%. The median company therefore loses sixteen points of revenue a year to churn and downgrades. Expansion has to add eighteen points to land at 102% net. Same customers, two very different stories.
Three leaks sit inside that arithmetic. Logo churn is a fit and onboarding problem. Contraction is usually a seat or usage problem, and weak expansion is a packaging problem.
A blended number folds all three into one line item that nobody owns. It also hides which one is eating the year.
A 102% net figure built on an 84% gross figure means expansion is carrying the year. Test the load before you celebrate the ratio.
Benchmark Against Your Segment, Not the Median
Retention moves with contract size, and the movement is large. Directive’s 2026 synthesis puts a good SMB range at 85% to 95% NRR. A strong enterprise range runs 115% to 130% or better.
| Segment | Good | Strong |
|---|---|---|
| SMB, under $10k ACV | 85–95% | 95–105% |
| Mid-market, $10k–$50k | 95–105% | 105–120% |
| Enterprise, $50k+ | 105–115% | 115–130%+ |
Those ranges are sanity checks rather than targets. They come from an agency synthesis of benchmark datasets, not a primary study, so hold them loosely. Pricing model moves the number just as hard. In the Aleph data, usage-based companies post 108% median NRR against 98% for seat-based, a ten-point structural gap.
Seat-based retention depends on headcount at the customer. Headcount is exactly what AI tooling is compressing right now. The usage-based 75th percentile reaches 155%, which shows how much room the pricing model creates. If your customer just bought an agent to cover three seats, that is a pricing problem. It arrives in your retention metric, and no amount of customer success activity fixes the math.
The Signals Your Health Score Is Missing
Every renewal review I sit in runs the same three inputs. Login frequency, ticket volume, and whether the CSM thinks the account is happy.
Those inputs are real, and they arrive late. By the time usage drops, the decision has usually been made in a room you were not in. I have run this decomposition on renewal books where the blended number looked fine and the enterprise cohort was quietly shrinking underneath it. The blended figure held because a handful of mid-market accounts expanded.
Nothing in the health score flagged the churn that followed. Marketing had earlier signals the whole time. It collected them and never put them on the account.
Which accounts read the renewal guide you published in September? Which champions showed up to the customer webinar? Did the champion change jobs, and did a new name appear on the record? Which customers searched the help center for the two features they never adopted? All of it is first-party, already captured, and missing from the health score.
None of that lives in the CRM.
Net revenue retention is a customer intelligence problem wearing a finance costume. The ratio tells you what happened to the base. It says nothing about what you understood about those accounts twelve months earlier.
This connects to the segmentation work and to the ICP-to-retention argument I have made before. A churned account that was a bad fit did not become a bad fit in onboarding. The story started in targeting and showed up in retention two years later.
I run this in a plain spreadsheet across two afternoons and never touch a dashboard. The shape of the answer matters more than the precision of the model, so resist the urge to buy something.
A Five-Step NRR Diagnostic
Pick the twelve-month window and the exact starting ARR. Exclude new logos from both sides, then write the definition down.
ACV band, pricing model, tenure, and whether the original champion still holds the same role.
Net hides everything. Gross retention per cut shows which part of the book is leaking. Recompute it for all four cuts before you form a hypothesis.
Split expansion into price increases, seat adds, cross-sell and usage growth. Each has a different durability, and only two survive a budget review.
Logo churn goes to the person who owns ICP fit. Contraction goes to pricing and product. Expansion goes to whoever owns packaging. Three names, three dates, one review at the end of the quarter.
What Good Looks Like Once You Split It
Good has a shape, and the shape differs by segment. A 96% blended number can be excellent for an SMB book and mediocre for an enterprise one. The expansion math only works when you know which segment is supplying it.
SaaS Capital’s 15th annual survey of more than 1,000 private B2B SaaS companies adds the growth link. Lifting NRR from the 90% to 100% band into the 100% to 110% band adds about five percentage points of growth. Companies with the highest NRR report median growth 173% above the population median. Read that as correlation with a strong mechanism, since expansion and growth feed each other.
Median growth across that survey was 22%, down from 25% the year before. Slower growth puts more weight on the base.
When I show this to a board I put the gross line first. It ends the argument about whether the growth is real before it starts. Nobody asks for the blended figure again.
Inside the same dataset, companies growing above 50% post a median NRR of 111%. Companies growing below 10% post 92%. The constraint on those slow growers sits in the installed base, not in the pipeline. Marketing owns part of that base relationship, which is why this lands on a CMO’s desk. Product owns the other part. Neither of them can fix it with more acquisition.
“Life favors the specific ask and punishes the vague wish.” That line from Tools of Titans is the argument in one sentence.
A blended NRR target is a vague wish. A cohort cut with an owner and a date is a specific ask.
Which cut in your book would you bet on if someone asked you to name one today?
Your NRR benchmark is a sample, not a standard. Decompose the number by segment and cohort, name the leak, and the benchmark stops being the point.



















