The Story Behind the Story

The Number That Would Not Stay Still

The first sign was a number that would not stay still.

The 2024 edition of the KBCM/Sapphire survey put the median private SaaS company’s 2022 ARR at $22.5 million. The 2025 edition put the same year at $17.0 million. Same survey, same statistic, same data year: 24% smaller. EBITDA margin for 2022 did the same thing, printed at negative 26% in one edition and negative 47% in the next. Neither figure is a typo, and neither edition is wrong. The survey redraws its respondent panel every year, and each new panel re-answers history. Both values are correct for the edition that printed them.

But a benchmark whose history moves is a different kind of instrument than the one most readers believe they are consulting, and once you have seen the past move, you start checking whether it moves in a direction.

It does.

Restated pair: 2022 median ARR $22.5M (KBCM-2024) vs $17.0M (KBCM-2025); 2022 EBITDA (26%) vs (47%). Full register: Data Integrity & Corrections LogVerified

Where We Started

We did not come to this data hunting for restatements. We came with a theory. After years inside SaaS revenue organizations, our working model was that churn, retention, and expansion are not separate problems parceled out to separate teams; they are connected parts of one revenue system, and they move together. The model made predictions before we opened a single PDF: downsell should be hiding inside the netted retention metrics, expansion should stall below a certain company size, and department-level fixes should show little effect on system-level outcomes. What we needed was public data rich enough to test the predictions, and the KBCM series, sixteen years of it, seven editions of it in our archive, was the only instrument deep enough to try.

The theory got its test, and most of the predictions held. That story is the report. This chapter is about what we had to get through to run the test at all, because the record turned out to have properties nobody had documented, and reading it honestly required building tools we did not expect to need.

Pulling the Thread

Comparing the 2024 and 2025 editions line by line against the source PDFs, we verified 39 figures that the newer edition restated. Thirty-one of them, 79%, moved the past in a worse direction: less profitable, slower-growing, smaller. This is not optimistic forecasts being trued up as actuals arrive. On the figures where both editions report settled history, no estimate involved on either side, 78% still move down. The only systematic upward revisions are net retention and sales efficiency, and they concentrate in the oldest vintage.

Read that pattern again, because it is the finding in miniature: with each passing edition, the past gets poorer and slower while getting better-retaining and more efficient. The record drifts, and it drifts with a lean.

The drift is also easy to miss by design of the reading, not of the survey. Editions sometimes rename a metric while revising its history. The 2022 median contract value appears under one label in the 2024 edition and a slightly different label in the 2025 edition, $10K apart. Match on names alone, as any reasonable reader would, and the two prints never register as the same series. We stopped matching on names. Every figure in this report is pinned to one named edition and one page, and no chart splices values across editions, because a spliced line manufactures movement that is partly a measurement change.

39 restatements, 31 downward (79%); settled-vs-settled 78%; renamed contract-value series. Register and per-figure pins: Reading KBCM Across EditionsVerified

The Optimism Ledger

Restatements look backward. The same two-edition comparison lets you look forward, and check the survey’s own estimates against what landed. Every edition prints its two most recent years as forecasts, marked E, because the survey closes mid-year before the current year’s financials exist. Where a 2024 estimate could be measured against the 2024 actual the next edition reported, the actual came in worse in 10 of 12 cases, concentrated in profitability, ARR dollars, and growth. The median ARR estimate missed by 28%.

The estimates lean optimistic, and consistently so. And the lean has a history. KBCM titled its 2022 edition Resilience Amidst Market Turmoil, naming the storm on the cover, while the same release told readers that private SaaS companies had “not been materially affected by the current market landscape.” The turn was already underway. The contemporaneous read minimized it. That is not an accusation; it is a calibration. An instrument that has historically read the present kindly should have its forward columns read as a ceiling, and that is how this report reads them.

Estimate-vs-actual: 10 of 12 worse, median ARR miss −28% (estimate accuracy table); 2022 framing documented in The 2022 TurnVerified

The Record Goes Dark

The third pattern took the longest to see, because it is a pattern of absences.

Across seven editions, the survey published a full conversion funnel exactly once, in the 2023 edition, on 2022 data. It never appeared again. ARR added per fully-ramped rep was measured in exactly two editions, and discontinued after its best-ever print of $673K; by 2024, the figure you can derive from the quota and attainment the survey still publishes was back at its 2020 level. The churn-by-contract-size cut, one of the cleanest gradients in the dataset, stopped after the 2020 data year. In each case the series ends near the moment the news turned.

We want to be careful with this one, because a survey trims questions for many reasons, and response burden is real. But the direction of the trimming is not random, and its effect is measurable in what this report could and could not decompose: new-logo CAC payback stretched from 31 to 37 months across exactly the window in which the funnel went unpublished, so the deterioration cannot be split into falling conversion versus rising cost per lead. The metrics that soured are the metrics that went dark, and the blackout sits at the center of the industry’s efficiency story. Inference

Funnel printed once; per-ramped-rep discontinued after $673K; payback 31 to 37 months. Full record: The Dark Panel and Payback DeteriorationCalculated

Who Holds the Pen

The 2025 edition, to its credit, published something no edition in our archive had: a breakdown of who answers the survey. Sixty of 71 respondents sit in the CFO, finance, or accounting function.

That single disclosure reorganized how we read everything else. This is a finance-office survey, answered from the systems of record a finance office runs, and those systems were built for the acquisition-led model. It is why the instrument carries dozens of spend and sales-productivity benchmarks and has never once measured Customer Success spend as a share of revenue. It is why Sales has an output metric and CS has a capacity metric. It is why the renewal, the single event where retained revenue is won or lost, has no owner anywhere in sixteen years of questions. The gaps are not carelessness. They are a portrait of a field of vision, and the honest conclusion is the one we drew in The Benchmark’s Blind Spots: the survey reports what its respondents’ systems can see, and those systems cannot see the half of the business that now produces most new ARR at scale.

None of this is deception, and we want the record to show the opposite of contempt: KBCM has produced the only longitudinal public instrument of its kind and given it away as a public good for sixteen years. This report exists because that survey exists. But an instrument’s blind spots become the industry’s blind spots, because analysts, boards, and operators train on what the benchmark shows. Over sixteen years, a benchmark does not just measure an industry’s model. It quietly teaches it.

Respondents by role: KBCM/Sapphire Survey 2025 p8 (CFO 29, Finance 28, Accounting 3); instrument coverage compared across editions in The Benchmark’s Blind SpotsVerified

Building the Instrument to Read the Instrument

By this point the requirements had written themselves. A record that restates its own history cannot be read by copying numbers out of the latest PDF. So every figure in this report was read twice from its source, once as text and once from the rendered page image. Every figure is pinned to a single named edition. The certified set is held in three independently maintained formats that must agree before a figure ships, and the reconciliation between them has caught real errors before publication rather than after. Every restatement we verified carries a classified reason, and every multi-year series carries a continuity flag, so a chart cannot silently join two respondent pools that are not comparable. The apparatus is documented in full in Methodology, and none of it was built for ceremony. It was built because the record demanded it.

What the apparatus buys is the report’s one methodological claim: representational fidelity. We can show you what the publishers printed, in which edition, and how it moved. We cannot certify that the underlying survey data is true, and we do not claim to.

What This Buys You

Three consequences follow, and they frame everything after this page.

First, this is why the report’s conclusions differ from the headline of any single edition. Read in isolation, each edition reports retention as stable. The decline is visible only when each figure is held to its own edition and the full series is read at once. The finding lives in the reconciliation, which is precisely why nobody printed it before: the reconciliation is the work.

Second, the biases we found run in a knowable direction, and it is the conservative one. Companies that decline, sell distressed, or shut down stop answering surveys, so survivorship tilts every median toward health, and the restatement lean means the past was consistently rosier at first print than in hindsight. The structural decline this report documents is likelier understated than overstated.

Third, the reader rule the whole apparatus supports: trust the gradients and the directions, hold the levels loosely, and treat any single-year improvement as unproven until a second edition confirms it. Panel turnover can move a median. It cannot manufacture a gradient measured inside a single pool, and the gradients carry this report’s strongest findings.

The registers that follow are the working. Every restatement, every estimate miss, every figure’s source page, sortable and checkable. We built them so this report would not ask to be trusted. It asks to be checked, which is a thing the record it describes was never built to invite.

Reader rule and panel-limitation analysis: Data Integrity & Corrections Log; companion field guide: How to Misread a BenchmarkCalculated