An anatomy of where recurring revenue actually escapes — twenty-five leak frameworks in five families — and an audit of what the public benchmark literature can and cannot tell you about it. Every claim carries a label. Every number carries its sample size. And the things we cannot yet claim are printed as prominently as the things we can.
Most benchmark reports open by telling you how large their dataset is. This one opens by telling you what it does not have.
Intake Review has delivered zero client diagnostics to date. Edition One therefore contains no proprietary client data at all. The free Revenue Leak Scorecard that feeds our intelligence layer went live in the same week this was published, so its corpus stands at zero rows on publication day. Nothing in this document is drawn from our own book of work, because there is not yet a book of work to draw from.
What Edition One is, then, is two things we can stand behind today:
One — the taxonomy. Twenty-five diagnostic frameworks, arranged in five families, that together describe the anatomy of revenue leakage in a B2B SaaS business. This is not a survey result. It is a structure, and a structure can be judged on its own merits: is anything missing, is anything double-counted, does it map to how the money actually moves. We publish it in full, including the exact question each framework asks.
Two — the audit. We took the public numbers a SaaS founder is most likely to be managing against, went to the primary sources rather than the summaries, and recorded for each one: the sample size, whether the metric's definition is published, and when the underlying survey was actually fielded. That audit produced a finding we did not expect, and it is the centrepiece of this edition.
Every factual claim in this document carries one of four labels, and the label is not decoration — it tells you exactly how much weight the claim can bear.
| Label | Means | You may |
|---|---|---|
| Verified | Read by us from the primary source, with the source, date and sample size recorded | Quote it, with the n attached |
| Calculated | Arithmetic we performed on stated figures; the arithmetic is shown | Check our working, then quote it |
| Estimate | A modelled coefficient chosen by us, not measured | Use it directionally. Never as a fact about your business |
| Unverified | Circulating publicly; we could not check it at the primary source | Treat with suspicion. We include it to show why |
Four findings. Two of them are about SaaS. Two of them are about the numbers people use to talk about SaaS.
Three reputable sources, three answers, twenty points of spread on the single most-quoted metric in software. None of them is wrong. They are measuring different populations, and only some of them tell you which.
Four of the five sources we audited state their sample size, and three publish how they calculate the metric. The failure is not concealment at the source — it is that n and definition fall off the moment a number is repeated in a deck.
28% every few years, 17% ad hoc, 4% never — 49% in total, against 29% who raise annually. On a survey of 310 companies. This is the most actionable public number we found.
Disclosed honestly inside the document, invisible by the time the number reaches a board pack. Benchmark vintage is a category of error nobody tracks.
A revenue leak is not usually a mystery. It is a number nobody owns, measured a way nobody wrote down, compared against a benchmark nobody checked. The frameworks in Section 03 exist to give each leak an owner and a definition. The audit in Sections 09–11 exists to show why the benchmark half of that sentence is harder than it looks.
Twenty-five frameworks in five families. Each one asks a single question, and each question has a dollar answer. Click a family to isolate it.
Revenue can only leak in five places: you charged too little for it (Pricing & Monetization), you won it and lost it (Retention & Expansion), you paid too much to find it (Acquisition & Funnel), you attracted it and never converted it (Activation & Conversion), or you burned the capital that bought it (Capital Efficiency & Moat). Every one of the twenty-five frameworks belongs to exactly one family, and the families are exhaustive by construction — five families, five frameworks each. If you can describe a revenue leak that does not fit one of these five, we want to hear about it, and Edition Two will carry it.
Net revenue retention is the metric a SaaS board asks about first and a valuation is priced on. We asked three reputable sources what the median is. We got three different answers.
Green line marks 100% NRR — the point at which a company grows without adding a single new customer. Bars are medians as published by each source. Verified each figure read from the primary source, 11 Aug 2026.
The temptation is to pick the number you like and move on. The honest reading is that the gap is mostly population, not contradiction:
ChartMogul measures roughly 2,700 B2B companies drawn from businesses that connect their billing system to ChartMogul, with an inclusion floor of $250k ARR. That population contains a great many small, self-serve, low-ACV businesses. SaaS Capital surveys private B2B SaaS companies above $1 million ARR — a materially larger, more sales-led population.
And here is the evidence that population is the explanation rather than an excuse: inside SaaS Capital's own data the same effect appears. Median NRR climbs monotonically with deal size — 95% at $0–10K ACV, 100% at $10–25K, 102% at $25–50K, 108% at $50–100K, and 115% above $100K. Verified
| Median NRR by ACV band | $0–10K | $10–25K | $25–50K | $50–100K | $100K+ |
|---|---|---|---|---|---|
| SaaS Capital 2025 · private B2B >$1M ARR | 95% | 100% | 102% | 108% | 115% |
A twenty-point spread between reports is therefore roughly the same size as the twenty-point spread within one report across deal sizes. Which means the practical conclusion is not "someone is lying".
Compare the two populations at their edges rather than their middles. ChartMogul's upper quartile for B2B SaaS is 97%. SaaS Capital's lower quartile at $25–50K ACV is also 97%, against a top quartile of 111%. Verified
Read that again slowly: a company in the top 25% of one dataset would sit in the bottom 25% of the other. Same metric, same year, same three letters on the chart. If you have ever been told your retention is strong, or weak, without being told which population you were being ranked inside, you now know how little that verdict was worth.
Two further figures from the same sources, for anyone sizing the AI question: ChartMogul puts AI-native NRR at 48% — near B2C levels, far below B2B. Verified And Benchmarkit reports gross revenue retention drifting down over three years, 90% to 88%, while median growth for usage-based-pricing companies ran 44% against 25% for traditional subscription. Verified — though the publisher itself cautions that usage-based pricing correlates with AI-native companies, which can bias that comparison. We repeat the caution because they were right to make it.
So the practical conclusion is this:
“Median SaaS NRR” is not a number you can manage against. Median NRR for companies of your ACV, measured your way is. Before you take any retention benchmark into a board meeting, you need three things attached to it: the population, the ARR or ACV band, and the formula. If a benchmark does not travel with all three, it is decoration.
We expected to find that benchmark publishers hide their samples. Mostly, they do not. We were wrong about where the failure happens.
| Source | Sample size stated | Definition published | Survey window stated | Headline we took |
|---|---|---|---|---|
| SaaS Capital 2025 Growth Rate Benchmarks | ✅ >1,000 | ✅ Yes | ✅ Q1 annually | 25% median growth |
| SaaS Capital 2025 retention page | ⚠️ Not on the page stated in the sibling PDF | ✅ Full formula given | ✅ Dec-23 → Dec-24 | 102% NRR at $25–50K ACV |
| ChartMogul The AI churn wave | ✅ 3,500 (~2,700 B2B) | ❌ Not published | ✅ Through Sept 2025 | 82% median B2B NRR |
| Benchmarkit / Pavilion 2025 Performance Metrics | ✅ N = 563 | ✅ Cohort method | ✅ Feb–Mar 2024 | 101% NRR |
| SBI / Price Intelligently 2025 State of SaaS Pricing | ✅ N = 310 | ✅ Yes | ⚠️ Not clearly stated | 29% raise prices annually |
| 1Capture Free Trial Conversion Benchmarks 2025 | ⚠️ “10,000+” claimed | ❌ None | ❌ None | 18.5% trial-to-paid Unverified |
Four of five primary reports state their sample size. Three publish their formula. Calculated — counted directly from the six rows above; the arithmetic is: n stated for SaaS Capital growth, ChartMogul, Benchmarkit and SBI; formula published by SaaS Capital retention, Benchmarkit and SBI.
So the folklore — that benchmark vendors bury their methods — is largely false, and we should say so plainly. The real failure mode is downstream. A number leaves a well-documented report with its n and its formula attached, gets quoted in a blog post without them, gets quoted again from the blog post, and arrives in a board pack as a bare figure with no ancestry at all.
The 1Capture row is the illustration, and we include it deliberately rather than to score a point. The page claims “10,000+ SaaS companies analyzed” and “2.5 million trial users tracked”, and names four data sources. It publishes no sampling method, no time period, and no citable dataset. Verified — we fetched the page on 11 Aug 2026 and searched it specifically for a methodology statement. That does not make its 18.5% figure false. It makes it uncheckable, which for a number you intend to run a company on is the same problem.
Can you name the population, the sample size, the formula and the date the data was collected? If any of the four is missing, you are not looking at a benchmark. You are looking at a rumour with a chart on it.
Of everything we audited, this is the number with the best sourcing and the most direct operational consequence.
On a survey of 310 SaaS companies, price-increase frequency breaks down like this Verified:
| How often prices rise | Overall | SMB | Midmarket | Enterprise |
|---|---|---|---|---|
| Annually | 29% | 23% | 32% | 27% |
| At contract renewal | 21% | 16% | 19% | 25% |
| Every few years | 28% | 28% | 29% | 30% |
| Ad hoc | 17% | 28% | 17% | 11% |
| Never | 4% | 6% | 2% | 5% |
49% of companies raise prices less often than once a year — 28% every few years, plus 17% ad hoc, plus 4% never. Calculated — 28 + 17 + 4 = 49, from the Overall column above.
The SMB column is the one to sit with. SMB-focused companies are 3× more likely than midmarket to never raise prices (6% vs 2%) and 61% more likely to raise them ad hoc (28% vs 17%). Verified — both comparisons are stated in the source, and both match the table above.
One more figure from the same survey, because it explains a great deal of unrealised pricing power: only 24% of companies localise price based on willingness to pay. Verified
Pricing is the only one of the five families where the behaviour is publicly measured with a stated n, even though the dollar cost of that behaviour is not. We know roughly half of companies let a year or more pass without a rise. We do not know, from any public source, what that costs them. That gap — measured behaviour, unmeasured consequence — is precisely the space a diagnostic occupies.
A quieter problem than the last three, and harder to spot once a number is loose in the world.
Benchmarkit's report is published and cited as the 2025 B2B SaaS Performance Metrics benchmarks. Inside the document, the survey window is disclosed clearly:
“Survey Date: Survey was open during February – March, 2024. We asked for either CY-24 data or last twelve months data.” Verified
That is honest reporting, and the disclosure is exactly where it should be. The failure is not the publisher's. It is that the vintage does not survive the journey: by the time “101% NRR (2025)” reaches a strategy deck, the fact that it describes a market fielded in early 2024 has evaporated.
We also found, inside the same document, that the participant count is stated as N = 563 in the participant profile and as “n = 583 participants” in two glossary footnotes, without reconciliation. Verified We quote 563, because that is the figure attached to the participant profile itself. We flag it not as a gotcha — a twenty-respondent difference changes nothing material — but because a report that asks to be quoted precisely should be quoted precisely, and readers deserve to know we noticed.
We publish our coefficients for the same reason we ask everyone else to: a model you cannot inspect is not a model, it is a claim.
The free scorecard scores four of the twenty-five frameworks from seven answers. Here is every coefficient it uses. All of them are Estimate — chosen by us, anchored to published behaviour, and not yet fitted to measured outcomes.
| Term | Formula | Status |
|---|---|---|
| Retention (F13) | MRR × monthly churn × 12 | Calculated — the visitor's own churn on their own MRR. No coefficient at all. |
| Pricing (F06) | ARR × uplift, where uplift = 2% (rise inside 6 months) rising to 13.5% (never), plus 4 points if pricing is a single flat price or has no clear structure | Estimate — the recency ladder is shaped by the 49%-raise-less-than-annually finding above |
| Onboarding (F07) | ARR × (25% benchmark − your signup-to-paid band) | Estimate — our weakest assumption, see below |
| Expansion (F14) | ARR × 2% (structured monthly motion) to 14% (nothing in place) | Estimate — not calibrated against any measured sample |
| Total | Sum of the four, capped at 60% of ARR | The cap is a judgement: past 60% the model stops discriminating |
The onboarding term compares every company against a single 25% signup-to-paid benchmark regardless of deal size. A $49/month self-serve product and a $5,000/month sales-led product plainly do not share one conversion bar. We use one anyway, today, because we do not have enough rows to split it by ACV — and the public conversion figures we could have borrowed are exactly the uncheckable kind described in Section 05. When the corpus can support an ACV split, this coefficient changes and we will say so in Edition Two, including what it was before.
Grade bands are equally arbitrary and equally published: leak below 10% of ARR is an A, 10–18% B, 18–28% C, 28–40% D, 40%+ F. They are judgement, not a distribution — because we do not yet have a distribution. Re-cutting them on real data is the first thing Edition Two owes you.
Pre-registering a claim is the cheapest honesty there is: you write down what would have to be true before you would say it, before you have any incentive to bend it.
Live from the scorecard intelligence layer. Aggregates are withheld by the endpoint itself until the threshold is met — this is enforced in code, not by promise.
| Claim we want to make | Blocked until | Status today |
|---|---|---|
| “The median SaaS company leaks X% of ARR” | ≥30 completed scorecard runs | ❌ 0 rows |
| “Your leak is worse than X% of companies your size” | ≥30 runs within an ACV band | ❌ 0 rows |
| “Pricing is the largest leak family” | ≥30 runs, reported with the spread, not just the mode | ❌ 0 rows |
| “Our coefficients are calibrated” | Measured before/after outcomes from delivered diagnostics | ❌ 0 delivered diagnostics |
| Case-study evidence from real engagements | The free-for-three deliveries complete | ⏳ offers scheduled 12 Aug 2026 |
Three of those five unlock at the same threshold, and the threshold is deliberately unglamorous: thirty rows is not a research programme, it is the minimum at which a median stops being an anecdote. We would rather publish “n=31, here is the spread, treat it as directional” than wait for a number large enough to sound impressive and quietly skip the disclosure in the meantime.
The public statistics endpoint behind the counter above returns the row count at any n, but withholds every median and percentile until n ≥ 30. Nobody has to remember the rule, and no future version of this page can quietly break it, because the data to break it with is not served.
A short version of the full scorecard, running entirely in this page. Nothing is sent anywhere and nothing is stored from this form.
Five questions cover four leak families with two coefficients defaulted. The full seven-question scorecard asks for your ARPU and pricing model as well, sizes the leak in customers rather than only dollars, and is the version that feeds the corpus above.
Run the full 7-question scorecard →Everything needed to check us, disagree with us, or reproduce this.
Each source was located by public search, then retrieved at the primary document — the publisher's own page or PDF — on 11 August 2026. Where a publisher offered a PDF, the PDF was downloaded and its text extracted locally rather than read through a summariser, specifically so that sample sizes and survey windows could be quoted verbatim. Two figures in our early drafting came from search-result summaries and were corrected after primary retrieval contradicted them; the corrected values are the ones printed here.
| Source | n | Population | Definition | Retrieved |
|---|---|---|---|---|
| SaaS Capital — 2025 Private B2B SaaS Growth Rate Benchmarks | >1,000 | Private B2B SaaS | Published | 11 Aug 2026 · PDF, local text extraction |
| SaaS Capital — retention benchmarks page | see above | Private B2B SaaS >$1M ARR | Full formula published | 11 Aug 2026 · page fetch |
| ChartMogul — The SaaS Retention Report: The AI churn wave | 3,500 | Software cos ≥$250k ARR, through Sept 2025 | Not published | 11 Aug 2026 · page fetch |
| Benchmarkit / Pavilion — 2025 B2B SaaS Performance Metrics | 563 | B2B SaaS; survey Feb–Mar 2024 | Cohort method, published | 11 Aug 2026 · PDF, local text extraction |
| SBI / Price Intelligently — 2025 State of SaaS Pricing | 310 | SaaS, SMB / midmarket / enterprise | Published | 11 Aug 2026 · PDF, local text extraction |
| 1Capture — Free Trial Conversion Benchmarks 2025 | “10,000+” | Not specified | Not published | 11 Aug 2026 · page fetch, methodology searched for |
SaaS Capital: “(Monthly Recurring Revenue in December of 2024 only from customers who were customers in December 2023) ÷ (Total MRR in December 2023)”. Benchmarkit: cohort method using ARR from the cohort of customers at the beginning of the period. ChartMogul: no definition published on the report page. That third entry is not an oversight on our part — we looked for it, and its absence is the finding.
| Finding | What would prove it wrong |
|---|---|
| The NRR spread is population, not error | Showing the three sources produce different medians on the same population and ACV band |
| 4 of 5 sources state n | Finding a stated sample size we missed — or that one we credited is not really stated. Our count is in the table in Section 05; check it |
| 49% raise less often than annually | A larger, better-sampled pricing survey with a different distribution |
| Our leak coefficients | Measured outcomes from real engagements. We have none. This is the weakest part of the document and we would rather you knew |
Intake Review sells revenue diagnostics. A document arguing that revenue leaks are large, poorly measured and worth diagnosing is directly commercially convenient for us. You should apply the same test to Section 08 that we applied to everyone else in Section 05: the coefficients are published, so check whether they are flattering to our own product. We think they are conservative. You are better placed than we are to judge whether we managed it.
Every substantive claim in this edition, with its label. Filter it.
| Claim | Label | Basis |
|---|
The State of SaaS Revenue Leaks — Edition One. Published 11 August 2026 by Intake Review, part of Alvi Ventures. Compiled from public sources retrieved at primary on the publication date, and from a leak taxonomy of twenty-five frameworks maintained by Intake Review.
How to cite: Intake Review (2026). The State of SaaS Revenue Leaks, Edition One. intakereview4500.com/benchmark
Corrections. If a number here is wrong, we want to know more than we want to be right — a report that audits other people's disclosure has no business being precious about its own. Write to hello@intakereview4500.com and corrections will be published in Edition Two with the original value shown alongside.
Edition Two will carry: the first scorecard corpus statistics with n stated, re-cut grade bands if the distribution warrants it, findings from the first delivered diagnostics, and any coefficient we were forced to change — including what it used to be.
Offline copy. The PDF twin is printed from this same page, so the two cannot disagree: THE-STATE-OF-SAAS-REVENUE-LEAKS-EDITION-ONE.pdf (18 pages).
Download the PDF ↓ Run the free scorecard → See the full diagnostic →