Across roughly 200 AI-native software companies above $250K in annualized revenue, median gross revenue retention measured 40 percent and median net revenue retention measured 48 percent during 2025. Median net revenue retention for the roughly 2,700 B2B software companies in the same dataset measured 82 percent. Kyle Poyar published the figures on 10 December 2025, working from a ChartMogul study of about 3,500 subscription companies categorized by an automated scrape.
Five months later, TechCrunch documented how that revenue base is counted before it is collected. Spellbook chief executive Scott Stevenson called the counting of uncollected contract value as ARR a dishonest metric and accused large funds of supporting it. TechCrunch interviewed more than a dozen founders, investors, and startup finance professionals, and reported one venture investor who had seen contracted ARR run 70 percent above collected ARR, a company marketing $50M in ARR against a collected figure of $42M, and a board that knew pilot-stage contract value was being counted as ARR during the pilot.
The gap comes from counting rules, and a buyer holding a billing export and an accounts receivable ledger can re-run those rules in an afternoon. Seven tests restate a reported ARR figure into collected recurring revenue and then estimate the share of it that persists twelve months out.
ARR means revenue currently delivered and invoiced, and four substitutions inflate it
Annual recurring revenue measures what a company would collect over the next twelve months from contracts it is delivering today if it sold nothing new. The definition requires the service to be live, the contract to be in force, and an invoice to have gone out.
Contracted ARR, sometimes labeled committed ARR, includes signed contracts whose delivery begins in a future period, so a contract signed in June with a January start date enters the ARR line in June and carries no revenue for six months. Pilot conversion value counts the post-pilot annual price of a pilot before the customer has decided to convert. Verbal commitments count revenue from customers who have stated an intent to renew without signing. Single-month annualization multiplies the most recent month by twelve, which imports any one-time fee, migration charge, or usage spike into the run rate at twelve times its one-month size.
Each substitution is defensible in isolation, and all four raise the reported number. The largest gap in the TechCrunch reporting is contracted ARR running 70 percent above collected ARR, and a company using all four substitutions can exceed that without any single line being a fabrication.
A billing export, a receivables ledger, and a contract register are enough to run the restatement
Test 1. Invoiced run rate. Sum every invoice line item issued in the trailing three months that carries a recurring subscription revenue type, exclude implementation fees, professional services, setup charges, and one-time overages, then multiply the three-month sum by four. A three-month window damps a single anomalous billing period, which twelve times the latest month does not. An invoiced run rate below 85 percent of reported ARR requires an explanation from management, and below 70 percent the gap becomes the primary diligence item.
Test 2. Cash conversion of the run rate. Divide trailing-twelve-month cash collected from customers by the average invoiced run rate across those twelve months. A company billing annually in advance will run above 1.0 during growth and a company billing monthly will run near 0.90 to 0.95, and both ranges are consistent with clean collections. A ratio below 0.75 means the company is issuing invoices that go uncollected, which Test 1 cannot detect.
Test 3. Gross dollar retention on a fixed cohort. Freeze the list of logos that were live and invoiced twelve months ago and sum their recurring revenue at that date. Those identical logos' recurring revenue today, with every dollar of upsell, cross-sell, seat expansion, and price increase excluded, divided by that starting figure, is gross dollar retention. Below 85 percent, new bookings replace lost revenue before they add any, and below 70 percent the company must win back more than 30 percent of its base every year to hold revenue flat.
Test 4. Net dollar retention and expansion concentration. Run Test 3 again on the frozen cohort with expansion included. The difference between the two figures is the expansion rate. Then rank the expansion dollars by account and count how many accounts produced half of them. If three customers supply all the expansion behind a 115 percent net figure, that 115 percent depends on three renewal decisions. Require gross and net side by side, and treat a package reporting only net as an incomplete disclosure.
Test 5. Weighted average remaining committed term. For each active contract, record months of committed term remaining, assigning one month to anything cancellable at will, then weight by that contract's recurring revenue. A weighted average below six months means most of the base can cancel within two quarters, which changes both the discount rate and the covenant package a lender would write.
Test 6. Usage revenue separated from committed subscription. Split the Test 1 run rate into revenue billed against a contractual minimum and revenue billed on metered consumption with no floor, then compute the trailing-twelve-month coefficient of variation on the metered component. Metered inference and per-token billing carry no minimum, so a customer can cut that revenue next month by making fewer calls. When metered revenue exceeds 30 percent of the run rate and its coefficient of variation exceeds 25 percent month to month, treat that portion as uncommitted.
Test 7. Customer concentration on the restated figure. Rank customers by restated committed recurring revenue and compute the top-one and top-five shares. Run the ranking on the restated figure, because removing pilot and contracted-but-undelivered revenue usually strips smaller accounts first and raises the top-five share. A top-five share above 30 percent or a top-one share above 10 percent is the threshold used here.
The seven tests restate a $10.0M reported figure to $5.70M of committed recurring revenue
The figures below are synthetic. They describe a company reporting $10.0M of ARR.
EXHIBIT 1 · FROM REPORTED ARR TO EXPECTED PERSISTENCE
Test 1. Recurring subscription invoices for the trailing three months totaled $0.640M, $0.650M, and $0.660M. The sum of $1.950M times four gives an invoiced run rate of $7.80M. The $2.20M gap breaks into $1.35M of pilot contracts valued at their post-pilot annual price, $0.60M of signed contracts with start dates in the following two quarters, and $0.25M of verbal renewal commitments from customers whose prior terms had lapsed.
Test 2. Trailing-twelve-month cash collected from customers was $5.95M against an average invoiced run rate of $6.40M, a ratio of 0.93, so the restated figure reconciles to cash. That same cash measured against an average reported ARR of $8.30M over the period produces 0.72.
Test 6. Test 3 needs a committed base, so Test 6 runs first. Of the $7.80M invoiced run rate, $2.10M was metered inference billing with no contractual minimum, and its monthly coefficient of variation was 34 percent. Committed subscription revenue is therefore $5.70M.
Test 3. The 141 logos live twelve months ago paid $5.40M on the committed-subscription basis. Those 141 logos pay $3.89M today with all expansion stripped out. Gross dollar retention is 72.0 percent.
Test 4. With expansion included, the frozen cohort pays $4.31M, a net dollar retention of 79.8 percent. Expansion contributed $0.42M, and two accounts supplied $0.33M of it, so net retention excluding those two accounts is 73.7 percent.
Test 5. Within the $5.70M committed base, $2.35M is cancellable monthly and $3.35M runs under annual terms with a weighted 7.4 months remaining, giving a weighted average remaining committed term of 4.8 months.
Test 7. The top five customers hold $2.19M of the $5.70M committed base, a 38 percent share.
At 72.0 percent gross retention on the committed-subscription basis, $4.10M of the $5.70M committed base is expected to persist twelve months from today, before any new logo or expansion dollar, against a reported headline of $10.0M.
The 40 percent AI-native median is measured on whatever those companies report as recurring revenue, so it is a directional reference here. The two figures rest on different bases.
The seven tests read trailing billing data, so they cannot value a pilot that has not converted
A company that rebuilt onboarding in month ten still shows the prior cohort's retention for another year, so an analyst running Test 3 reads the rebuild as history. Splitting retention by signup vintage is the partial correction, because a rising series across recent vintages is evidence the rebuild changed customer behavior.
A pilot converting at 90 percent and a pilot converting at 10 percent look identical in a billing export. Only customer references and usage depth inside the pilot separate them, and neither is in the billing export.
Companies that migrated billing platforms mid-year, or that invoice partly through a reseller, produce customer identifiers that do not match across the two dates Test 3 requires, and reconciling them by hand consumes most of the analyst hours in the restatement.
The ChartMogul sample holds roughly 200 AI-native companies identified by automated categorization of scraped web data, and Poyar describes the per-bucket cuts as directional at about 50 companies each. The dataset also skews toward self-serve subscription billing, which pulls the 40 percent median down through cheap plans. AI-native plans under $50 per month showed 23 percent gross and 32 percent net retention. The $50 to $249 band showed 45 percent and 61 percent, and plans above $250 per month showed 70 percent gross and 85 percent net. An enterprise-priced AI company should be screened against the top band, and using 40 percent as the expectation for a company with a $200K average contract value would be wrong in the company's favor.
Investors apply Carta's 42 percent AI seed premium to the revenue figure in the deck
Carta reported median pre-money valuation on AI seed rounds of $17.9M in 2024, 42 percent above the median for non-AI companies at the same stage. Peter Walker, Carta's head of insights, put early-2025 AI seed medians at $19M pre-money against $13M for non-AI, a 46 percent premium.
Tests 1, 2, and 6 run off a billing export and a receivables ledger in under four hours. Customer-level revenue at two dates adds most of a day and produces Tests 3 and 4, and the contract register behind Tests 5 and 7 adds another two hours. A buyer who runs the restatement before the term sheet prices $4.10M of expected persistence against the $10.0M headline. The same restatement run after signing surfaces the $5.90M difference at the next round, when the buyer already owns it.
Sources
- Kyle Poyar, Growth Unhinged, 10 December 2025, the AI churn wave, drawing on the ChartMogul retention study of roughly 3,500 subscription companies.
- TechCrunch, 22 May 2026, on inflated ARR.
- Carta, AI fundraising trends, 2024, and Peter Walker, March 2025.
SEND A REVENUE BASE FOR A READ
EXTERNAL FIGURES TRACED TO PRIMARY PUBLICATIONS AND LINKED ABOVE. THE WORKED EXAMPLE USES SYNTHETIC FIGURES. SECOND ORDER PUBLISHES SAMPLE MODELS ON SYNTHETIC DATA ONLY.