The cohort table the deck didn't include.

Why founders show triangle charts instead of cohort tables — and the three reconstructions that turn a flattering retention slide into a defensible one.

01The triangle chart problem

The most-shown retention chart in seed-stage decks is the triangle: an aggregate retention curve that starts at 100% and declines smoothly over twelve months to whatever number flatters the story. The chart is comforting because it implies a steady-state retention rate. It is misleading because it averages every cohort the company has ever had — including the ones that decayed quickly and the ones that have not yet existed long enough to decay at all. The shape of the triangle is determined by the mix of cohorts, not the underlying retention behaviour of any single cohort.

A company that retains 70% of its month-one users at month twelve looks identical, on a triangle chart, to a company where half of cohorts retain 90% and half retain 40%. The two are not the same business. The first is sustainable; the second is two product–market fits stapled together, one of which is leaking customers at a rate the aggregate hides. The diligence question is whether the cohort distribution is converging to the strong number or to the weak one.

Aggregate retention also drifts upward over time mechanically. A company that grew faster recently has more recent cohorts in the denominator, and the recent cohorts have not had time to churn. The chart looks like retention improved when, structurally, it merely got younger. Founders rarely intend this — but they rarely catch it either. The triangle chart that the deck ships is honestly built and structurally misleading at the same time.

In our corpus, 12% of seed-stage decks include a cohort table. The other 88% include a triangle, a logo-retention number, an NDR claim, or an aggregate revenue retention figure. All four are weaker than the table. The work of this note is to explain why, and what to do about it on the diligence side when the table is absent.

02What a cohort table is, what it isn't

A cohort table is a two-dimensional grid. Rows are start cohorts — usually the calendar month each customer signed up. Columns are tenure periods — month one, month two, month three, and so on after signup. Cells contain the retention rate for that cohort at that tenure. A complete table for a two-year-old company is twenty-four rows tall and twenty-four columns wide, with the lower-right triangle empty because future tenure periods have not occurred yet.

What the table shows that the triangle chart hides: each row is a cohort, and each cohort can be read independently. If month-one cohorts retain better than month-twelve cohorts at the same tenure, the company is acquiring better customers as it matures and the business is strengthening. If month-twelve cohorts retain worse than month-one cohorts, the company is acquiring worse customers as it scales and the GTM motion is breaking. Both patterns are common. Both are diagnostic. Neither is visible in a triangle.

A few things a cohort table is not.

It is not the same as logo retention. Logo retention is a single number that aggregates across cohorts and obscures cohort-level variation. A cohort table that shows ninety percent logo retention in the strong cohorts and sixty percent in the weak cohorts is more useful than the aggregate that splits the difference.

It is not the same as net revenue retention. NDR mixes expansion and contraction within existing accounts and tells you nothing about whether new cohorts churn. See Note 09. A company with strong NDR and weak cohort retention is a business that grows revenue by upselling a shrinking customer base, which is durable until it isn't.

It is not the same as a triangle chart with the cohorts colour-coded. A coloured triangle chart still aggregates within each cohort line and obscures the within-cohort behaviour at the cell level. The table is structurally more informative because each cell is a separate observation, not a smoothed line.

The point of asking for the table specifically is that no other format reveals the same information. The substitutes do not substitute.

03The three reconstructions when the founder can't produce one

Some founders cannot produce a cohort table within the diligence window. Sometimes the data is in an analytics system that does not expose it cleanly. Sometimes the founder has the data but has never structured it this way. Sometimes — rarely — the founder does not have the underlying customer-level data at all, which is itself a flag.

We have three reconstructions we use when the table is absent. Each works from data the founder will share, even if the founder will not produce the table directly.

Reconstruction one — from MRR plus customer count. If the founder shares monthly MRR and total customer count for the prior eighteen to twenty-four months, we can back out a cohort approximation by solving for new-cohort size each month and applying a churn assumption iteratively. The approximation is rough at the edges but captures the directional pattern. The fit improves materially if the founder also shares net new MRR each month — gross adds minus churn — which most founders track and many will share without protest.

Reconstruction two — from billing system exports. If the founder routes through Stripe, Recurly, Chargebee, or any standard SaaS billing system, the export contains every invoice the company has issued. From the invoice stream we can construct cohort tables directly without any input from the founder beyond access to the export. We ask for the export under NDA, run the reconstruction in our own pipeline, and present the table back to the founder for comment. Founders who decline the export request are flagging a problem upstream of the diligence; founders who provide it within forty-eight hours have passed a separate test about how the company handles its own data.

Reconstruction three — from a sample of customer references. If the data is genuinely unavailable, we sample. Pull a list of all customers from the prior eighteen months, randomly sample fifteen, ask the founder to introduce us to each for a five-minute call, ask each whether they are still a customer and at what spend level. The reconstruction is small-sample and imperfect but produces a non-zero observation. A founder who cannot introduce us to a fifteen-customer random sample within the diligence window is flagging something else — usually a customer relationship managed by sales rather than the founder.

The reconstructions are presented to the IC alongside the founder's headline retention claim. Where the reconstructions match the claim, the dossier notes the agreement and the diligence moves on. Where they diverge by more than ten points, the dossier asks the founder to explain the gap. The gap is usually explainable. The diligence is not the gap; the diligence is whether the founder can explain it.

04Reading a healthy retention curve

A healthy cohort table has three properties we look for in the dossier.

Property one: the lines should converge to a non-zero floor. Each cohort line, read horizontally across the table, should approach a steady-state retention rate as it ages. The floor is the long-run retention rate of the cohort. A floor of 60% at month twenty-four is healthy for vertical SaaS. A floor of 80% is healthy for horizontal SaaS. A floor of 95% is healthy for enterprise infrastructure. A floor below 40% at any twenty-four-month horizon is a flag regardless of category — the company is leaking customers faster than the GTM motion can replace them.

Property two: the lines should be parallel. Each successive cohort, read vertically, should track close to the cohorts above and below at the same tenure. The lines do not have to be perfectly parallel — small variations are normal and reflect the noise of small monthly cohorts. Large divergences are the signal. A cohort that retains thirty points below the surrounding cohorts at month six tells you something specific happened in that acquisition window — a pricing change, a product change, a marketing-channel shift — that changed the customer profile. The diligence is to ask what changed.

Property three: the most recent cohorts should not be worse than the historical cohorts. This is the property most decks hide. A company scaling acquisition without holding cohort quality constant produces newer cohorts that retain worse than older ones, even though the aggregate retention number improves (because the newer cohorts are larger and younger). Reading the table by recency rather than by aggregate is the single largest correction available against the deck's claimed retention.

Two pathological patterns appear often enough to be named. The first is the inverted curve — newer cohorts retain better than older ones — which usually indicates the older cohorts were grandfathered onto plans the company no longer sells. The second is the bimodal curve — half the cohorts retain at 80%+ and half retain at 30% — which indicates the company is selling into two ICPs and only one of them has product–market fit. The deck's aggregate retention number conceals both.

05What 88% of decks are hiding

The 88% of decks that do not include a cohort table are not, in most cases, hiding bad retention. The retention is often fine. What the deck is hiding is the founder's willingness to be examined at this level of detail.

The cohort table is the single most data-dense exhibit a founder can include in a seed-stage deck. It commits the founder to a number that cannot be slightly massaged. Every other retention metric on the deck has a range of defensible definitions. Logo retention can be calculated by month, by quarter, by year, by cohort, by anniversary. NDR has six common definitions and the founder can pick the most flattering. Aggregate revenue retention has even more degrees of freedom. The cohort table has none. The table is the table. Once it is on the deck, the founder is committed.

Founders who include a cohort table at seed are signalling something specific. They are willing to be wrong in a way that is publicly auditable. They have done the work of pulling the data, structuring it, and committing to whatever it shows. The act of including the table is itself a diligence-positive signal — not because the cohort numbers are uniformly good in our corpus (they are not), but because the willingness to expose them is correlated with founders who do the work of measuring everything else honestly.

Founders who exclude the cohort table are not, by default, hiding something. They are usually following the pitch convention they have been coached on. The convention does not penalise the omission, so founders comply. The cohort table is an above-the-line ask the IC has to make for itself. It is rarely volunteered, even when the underlying data would have flattered the deck.

06How to ask without breaking the call

The cohort-table ask is delicate in a partner meeting because it implies a specific kind of doubt — that the deck's retention number is incomplete. Most founders read the ask as confrontational even when it is not intended that way. A few framings keep the call productive.

Ask after the founder has finished the deck, not in the middle of the retention slide. Interrupting the retention slide invites the founder to defend the slide. Asking after the deck lets the founder treat the question as a follow-up rather than a challenge.

Ask for the table by its specific name. “Can you share a cohort table — rows by signup month, columns by tenure month, cells with retention rates?” Specificity removes the ambiguity that lets a founder send back a different artefact and claim to have answered the request. Founders who do not know what a cohort table is — and there are a few — will say so, and the call can move to a different question about how the company measures retention.

Ask in writing immediately after the call. The founder forgets the verbal request more often than the written one. The written request also produces a paper trail that disambiguates between “the founder did not produce the table” and “the founder forgot the ask.” The distinction matters at IC.

Offer to reconstruct. “If the table isn't in a clean format yet, we can run it from a Stripe export under NDA.” The offer removes the founder's plausible objection that the data is not in a presentable state. Most founders accept the offer. The ones who decline are flagging something specific about the data, which is itself diagnostic.

The diligence question is not whether the table exists. The diligence question is what happens when you ask for it.