The Traction-Claim Corroboration Standard: Corroborated, Partial, or Unsupported

A founder's deck leads with a number: £50k MRR, $10M ARR, 120% net retention. Your job, before you spend a day in the data room, is to decide whether that headline is credible enough to earn your diligence time. This guide gives early-stage investors, platform and talent partners, and angels a fixed way to grade any headline traction claim as Corroborated, Partial, or Unsupported, using only externally observable signals, so two analysts looking at the same deal land on the same verdict.

This is a definition of done, not a teardown. The existing library has worked examples of single ARR and gross-margin claims. What has been missing is the reusable pass bar a team adopts as policy: the external-evidence checklist that gates diligence before anyone opens a private data room.

What "corroborated" means, in one gradable sentence

A traction claim is Corroborated when its restated number coheres with externally observable headcount, hiring, web and app footprint, and named customers, with no contradicting signal; Partial when signals are mixed or incomplete; and Unsupported when at least one external signal clearly contradicts the headline.

The whole standard rests on one idea investors already use: you triangulate a stated ARR, MRR, or user count against headcount, job history, web traffic, app presence, named customers, and reviews before accepting it. A claim is not credible because it is impressive. It is credible because it is internally consistent and survives contact with evidence you can pull from the open web.

The three grades map to three actions:

The bar for Corroborated is coherence, not proof. You are not auditing the books from the outside; that is impossible. You are deciding whether the number is plausible enough that the deeper work is worth doing. The arithmetic test for coherence is blunt: credible numbers mean customer count times ACV equals ARR, and implied headcount sits inside the stage band.

The external evidence that actually moves a grade

Six signals do the work: implied headcount, hiring, web traffic, app presence, named customers, and registry filings. Each proves one thing, and each lies in a specific way you can anticipate.

The most powerful is ARR per employee, because payroll is observable and hard to fake. It turns a narrative into arithmetic. Divide the restated ARR by a stage-appropriate band and you get an implied FTE count you can compare to the company's LinkedIn footprint and open roles.

Here is what each signal proves and how it fails:

Two of these deserve a hard caution. Web-traffic estimators have a floor. For sites with 250,000-plus monthly users, one widely used estimator lands within plus or minus 30% of actual about 62.9% of the time. In the 5,000 to 100,000 range it is the most accurate of the common tools. Below 5,000 monthly users it is the worst, usually overstating by two to three times. At that scale, drop traffic as evidence rather than discounting it, because the error exceeds the signal.

The second caution is the registry. A UK company publishes accounts through Companies House, and full accounts show turnover, but small companies may file abridged accounts that omit it. More important: Companies House does not verify the accuracy of information filed, which means incorrect figures can still be filed and uploaded. A registry can only ever downgrade a claim when it contradicts the headline. It should never upgrade one to Corroborated on its own. App-store estimators carry the same warning: the data is model-based, with no access to the actual console numbers.

Strip the claim to a defensible base first

Before you test a number, you have to agree on what it is. Investors verify that a founder's ARR definition matches theirs, and the fastest inflation hides in the definition, not the arithmetic. A single consistent definition for each metric must apply across board decks, investor materials, and financial statements.

ARR is forward-looking, non-GAAP, and excludes everything one-time. Strip these before you test:

• Free trials and pilots
• Professional services, onboarding, and implementation fees
• One-time charges and uncommitted usage overages
• Letters of intent and verbal interest, which are not revenue
• Expansion revenue you might lose if a customer leaves
• Signed-but-not-live contracts, which are committed ARR, not live ARR

ARR that quietly includes implementation fees or services revenue reads as sloppiness or inflation once billing data gets unpacked. The classic worked case: a founder reported $50K MRR at seed, and verified bank statements showed $31K, the gap caused by counting annual contracts as monthly revenue. That is a 38% overstatement sitting entirely in the definition.

Durability metrics decide whether the base survives. Treat GRR as the honesty check, because it excludes expansion and only counts what a company kept. For the honesty check, top-quartile GRR holds 91% and the bottom quartile sits at 76%; aim for 85%-plus, with 90%-plus marking top-tier durability. CAC payback adds a second durability read: the median is 16 months, and the top quartile is 6 months or less.

Benchmark medians disagree enough that you must name the source every time. Net retention medians run across a 20-point spread depending on who published them.

Across all SaaS companies, the 2025 median net retention is 101% and the median gross retention is 91% per SaaS Capital's 1,000-plus company set. A founder who cites "we beat the 82% median" when the relevant source median is 101% is source-shopping. Pin one named source per metric and the room to do that disappears.

The arithmetic test: implied versus observed headcount

The single most falsifiable test is ARR per employee, because it converts a claim into a headcount you can count on LinkedIn. Divide restated ARR by the stage median, and compare the implied FTE to what you observe.

The bands vary sharply by stage, so use the right row. A number that looks reasonable at $100M ARR is absurd at seed.

Worked example: a $10M ARR claim at the $5M to $20M median of $130,000 implies about 77 FTE. Observed headcount under about 40 is a two-times-plus gap and a fail signal. A 12-person LinkedIn footprint behind a $10M claim is a 6x gap, which is Unsupported on its own unless the founder has an extraordinary explanation.

The band is wide for a reason. At the extreme top, Bessemer Cloud 100 "AI Supernovas" averaged $1.13M ARR per employee against a $283K median for public SaaS, and one company reached roughly $400M ARR with 146 employees, about $2.7M per head. Those are outliers, not the band. If a founder claims outlier efficiency, that claim itself needs corroboration, not a free pass.

Payroll is observable and hard to fake, so implied FTE exposes inflation faster than any narrative.

When you need to count the revenue-generating headcount behind a claim precisely rather than eyeballing a LinkedIn number, enumeration by role and skill is the hard part. Refolk resolves people to named current employers, so you can size a sales or engineering team against the implied FTE the arithmetic demands.

One honest limit on role-level counts: in Refolk's index, zero matches returned for the title "Account Executive" and for "Customer Success Manager" when filtered to Computer Software plus United States as a bare title-industry-country combination. Role counts resolve best against a named company or a specific skill, not a generic title filter, so frame the query around the company you are testing.

The procedure: seven steps from claim to grade

Run these in order. The whole pass fits inside the 30 to 45 minutes investors spend on desk research before a first call. If a step produces a clear fail, you can stop early and decline.

There is a real order dispute worth naming. Some practitioners run external corroboration before stripping the base, because desk signals are faster to pull. I put definition-and-strip first, because the number you test should already be clean. Corroborating an inflated number against external signals just tells you the inflated number is consistent with itself.

A mismatch between a reported figure and what appears in filings typically warrants a direct conversation with the founder before the process goes further. That conversation is not a failure of the standard; it is the standard working. Partial exists precisely so you can take that call with specific questions rather than a vague unease.

The funnel figures are illustrative of the shape, not measured counts. The point is directional: most of the thinning happens before any data room opens, which is exactly where you want to spend the least.

How this goes wrong: failure modes and false positives

The most valuable part of a standard is knowing where it misfires. Each failure mode below has a false positive to watch for and a cheap check that resolves it.

Annual-vs-monthly inflation

A headline MRR that silently counts annual contract value as monthly revenue. The deck shows strong MRR; the bank shows roughly 60% of it. Check: ask for the billing export and the as-of date.

Services dressed as ARR

Implementation and professional-services fees folded into ARR. You see high "ARR" with lumpy, non-renewing line items. Check: reconcile ARR to GAAP revenue and strip one-time items.

NRR shown without GRR

Expansion masks churn behind a flattering net number. A 110% NRR can hide a 76% GRR. Check: demand GRR, which caps at 100% and cannot be inflated by upsell.

Benchmark source-shopping

The founder cites the most flattering median. "We beat the 82% median" when the relevant source median is 101%. Check: fix one named source per metric and hold the deal to it.

Traffic estimate taken as truth

An estimator reading on a sub-5,000-user site treated as fact. "100K visits" that is a two-to-three-times overstatement. Check: discount small-site estimates and require registry or billing corroboration.

Registry treated as audited

Companies House turnover accepted as verified. A self-filed figure, or abridged accounts with no turnover at all. Check: remember filings are unverified and small firms can omit turnover; use the registry only to contradict, never to confirm.

Headcount lag

LinkedIn undercounts a fast-hiring team and triggers a false fail. Check: cross-read open roles and recent joiners before grading Unsupported on headcount alone.

The checklist you adopt as policy

Run this before calling any claim graded. The grade holds only if every line is satisfied. This is the artifact you paste into your deal template so every analyst applies the same bar.

A clean scoring rubric keeps the grade reproducible. Copy this into your notes and fill the right column:

Claim (dated, one line):        ________________
Restated base (committed only): $________________
Implied FTE (ARR / band):       ________________
Observed headcount:             ________________
Implied-vs-observed ratio:      ________ (>=2x = fail)
GRR (named source):             ______%  healthy / watch / reprice
NRR (same source):              ______%
Traffic (only if >5K/mo):       ________ or DROPPED
Named live customers:           ________
Registry check:                 confirms-nothing / contradicts
Deciding signal:                ________________
VERDICT:                        Corroborated / Partial / Unsupported

Keeping the standard current

Adopt the checklist as written, but re-anchor the numbers on a schedule, because the bands move. ARR per employee rose 29% in one year, from a $150K median to $193K, and public-company net dollar retention fell from 125% in 2022 to 112% in 2025. A band you set and forget will mis-grade claims within a cycle.

Two maintenance habits keep the standard honest. First, refresh the named benchmark source for each metric once a year and note the publication it came from in your template, since the 20-point spread between sources means a stale source quietly changes your bar. Second, re-check the estimator thresholds, because the reliability floor for traffic tools is a property of their panels, not a fixed law; if a tool changes its methodology, the 5,000-user floor moves with it.

The grade itself has a short shelf life. A company hiring hard will clear the headcount test next quarter even if it fails today, so a Partial graded on headcount lag is worth re-running rather than treating as final. When you re-grade, pull fresh headcount and open roles rather than trusting a count from last quarter. In Refolk's index, headcount-by-skill enumeration resolves to named current employers - for example, the index returns 3,641 US professionals with Rust skills at employers including NVIDIA, Cloudflare, and Temporal - which is the kind of current, countable footprint that keeps a stale grade from standing.

Where a signal is genuinely unavailable, say so in the record and grade Partial rather than guessing. A standard that forces a confident Corroborated on thin evidence is worse than one that honestly marks the gap and routes it to a founder conversation. The discipline is not in always reaching a verdict; it is in two analysts reaching the same one.