Module 6 · GTM Analytics & Instrumentation · Chapter 26 of 32
Cohort and Trendline Analysis
A live tile shows the business as it is right now, so it cannot see a thing that only happens over time: compounding. Cohorts and self-benchmarking trendlines add the time axis, so you watch the compound and denominator effects happen in your own data instead of arguing about them.
10 min read
Someone forwarded me a churn dashboard with a red arrow on it. The tile counted paying accounts inside a single customer cohort, and over eight quarters it had slid from 561 to 347. The note attached asked what customer success had broken. I pulled the same cohort’s ARR on the second monitor. Across those exact eight quarters it had climbed from $18.3M to $22.9M. The accounts leaving were small, the accounts staying were expanding, and revenue per surviving account had doubled from about $33K to about $66K. The cohort was not dying. It was compounding. The dashboard could only count heads, so the only thing it could see was the leaving.
That gap is what this chapter closes. In Building the Bowtie Dashboard we wired volume, conversion, and time metrics into live tiles, and a live tile is honest about one thing: the business as it is right now. Compounding is not a right-now phenomenon. It is something that happens to a fixed group of customers as they age, so a snapshot structurally cannot show it. To see it you hold a group still and watch it move through time. That is a cohort. And the most reliable yardstick to judge a cohort against is not an industry benchmark, it is the same company’s earlier cohorts, which Revenue Architecture calls self-benchmarking: your own trendline controls for your motion, your ACV band, and your segment mix in a way no external panel ever will.
A cohort is the time axis a tile is missing
A cohort is a group defined by when it started and then tracked forward on a shared clock. The dollars-versus-logos split above came from a real one: 617 accounts whose first-ever purchase landed between September 2021 and August 2023 at Anaconda. It is the same cohort the growth model used to name the compound effect and the denominator effect. This is where you watch both of them happen.
Line the cohort up by time since it started paying (the clock runs from the contract start date, not the close date) and two curves come apart. Count the logos and you get a decay curve: 100% at the start, then 72%, 56%, and 43% by year three. Measure the dollars those survivors carry and you get net revenue retention holding at or above 100% the whole way: 100, then 109, 102, 104. The logo curve is the denominator effect wearing its everyday clothes, smaller accounts churning out of the base as it ages. The dollar curve is the compound effect, the survivors renewing and expanding fast enough to more than cover what left.
View as table
| Point | Value |
|---|---|
| Start | 100% |
| Yr 1 | 72% |
| Yr 2 | 56% |
| Yr 3 | 43% |
Toggle the two curves and the whole argument from the factory chapter resolves into a picture: this business grows off its installed base, not off head count. The landing deal is a foot in the door, a median of about $10.5K against a median lifetime of about $53K, so roughly three quarters of the money shows up after the first signature. That single fact is why the cohort multiplied 3.97x overall while shedding more than half its logos, and it is invisible to any metric that treats a customer as a unit instead of a stream. It also self-benchmarks cleanly: split the same cohort by the year each account landed and the 2022 landings returned 4.67x, 2021 returned 3.34x, and 2023 returned 3.07x, so you can see which vintages produced the durable customers without reaching for an outside report at all.
Build a cohort you can trust
Cohort analysis fails quietly, because a broken cohort still produces a confident number. Every headline above survived a specific mistake I made or nearly made, and the guardrails are worth stating as a sequence rather than a warning.
- 1
Anchor the clock to the event, not the calendar
Age every account from the moment it actually started (contract start, or first delivered impact), never from the close date or a fiscal boundary. A backdated multi-year term signed late will otherwise show up as if it were years old on day one, and your retention curve inherits the error.
- 2
Define membership once, then freeze it
Base the cohort on an event you can recompute, such as an account's first-ever closed-won deal, not on a mutable label field. A "New Logo" flag on this cohort disagreed with the first-ever-win test on 95 accounts: 61 real first wins were unlabeled and 34 labels sat on accounts that had already bought.
- 3
Guard the point-in-time sum
When you value the cohort as of a date, sum only the revenue whose term spans that date AND whose deal closed on or before it. Drop the close-date guard and future backdated terms leak into earlier periods, inflating the past and flattening the very trend you are trying to read.
- 4
Rank within the landing window, not across the lifetime
To find the "landing" deal, rank an account's deals inside its landing period only. Ranking across all of an account's lifetime deals picks the biggest deal it ever did, which once understated this cohort's landing value from $27.0M to $15.6M and inflated the LTV multiple on the way.
- 5
Read the curves, not the blended average
A single 3.97x multiple hides the split between the accounts that churned and the accounts that quadrupled. Report logo retention, net dollar retention, and revenue-per-surviving-account as separate lines. The story lives in how they diverge, which is the whole point of holding the group still.
None of this is exotic. It is the discipline of measuring one clearly defined group with clean edges, which is a data-model job the GTM Engineer Academy covers on the build side and that the CRM data model chapter set up when we standardized the fields. Get the edges wrong and the analysis is worse than no analysis, because it is a wrong answer delivered with a trendline behind it.
Self-benchmarking with polynomial trendlines
Once a cohort is clean, you fit a trendline so the shape is a claim you can test rather than dots you can squint at. Revenue Architecture is specific about the method: take eight to ten periods, split each into acquisition, retention, and expansion the way First Principles taught, plot them side by side rather than stacked, and fit a second- or third-degree polynomial to each stream. A straight line is the wrong tool here on purpose. Recurring revenue compounds, and compounding is nonlinear, so a linear fit will always lag a stream that is actually bending. The target fit is an R² near 0.98.
The reason to benchmark against yourself, and not against a panel, is that your own history holds the variables that a benchmark averages away. The book’s SaaS benchmark set spans 2016 to 2022 across 868 companies by ACV band, and it is genuinely useful for one thing: telling you roughly which neighborhood your conversion rates live in. It cannot tell you your trajectory, because it is not running your motion into your market. Your last eight quarters are. A small move inside your own numbers propagates in ways a benchmark will never flag: let win rate slip from 30% to 25% and the leads you need per win climb from 3.33 to 4.0, so the same target quietly demands 20% more opportunities. You catch that in your own trendline weeks before an annual benchmark refresh could.
Cohorts run on the acquisition side too
Retention cohorts get the attention, but the same discipline diagnoses the left side of the Bowtie, and there the clock is what you have to argue about. On a standing quarter-cohort analysis I built, each cohort was the set of deals created in a quarter, tracked to the end of that same quarter: a 90-day, in-quarter conversion lens. Held that way, the funnel is remarkably stable, and it points at exactly where the motion leaks.
| Stage reached | % of created cohort | Advanced from prior stage |
|---|---|---|
| Qualification | 100% | base |
| Discovery | 68% | 68% |
| Evaluation | 35% | 51% |
| Proposal | 28% | 80% |
| Negotiation | 22% | 79% |
| Closed Won | 19% | 86% |
The in-quarter win rate held at 19.2% in two consecutive quarters, which is the kind of repeatability that lets you trust a diagnosis. And the diagnosis is unambiguous: the bottleneck is early. Only 68% of qualified deals reach discovery and only about half of those reach evaluation, while everything from evaluation onward converts above 79%. Coaching dollars aimed at closing skills would be aimed at the one part of the funnel that already works.
What you can now do
Give me a group with clean edges and a shared clock, and you can now follow it through time: build a retention curve and an LTV multiple, watch the compound effect outrun the denominator effect in your own data, fit a low-degree polynomial and use the R² as a truth check rather than a trophy, and self-benchmark this quarter’s cohort against last year’s instead of against a panel that never sold your product. You can also spot the lens artifacts, the places where the window you chose, not the team you measured, produced the number.
Cohorts answer how a population converts and compounds once it exists. They do not answer where the population came from or which spend earned it. That is the next and last analytics question in this module. In Attribution and Pipeline Analytics we trace pipeline back to its sources and measure how it converts across the funnel, and we do it without the attribution theater that assigns a tidy percentage to every touch and a real answer to none.