Catch unit economics drifting by segment.
Revenue grew. Blended margin moved half a point. You lost the segment.
A payment-mix shift, a shipping-zone change, a support-heavy cohort, a discount code that leaked. Each one takes a third off the contribution margin of one segment and almost nothing off the blended number. Margin watches contribution per order by segment and names the segment and the day it broke.
Fourteen unit-economics series, one segment quietly going underwater, detection computed live in your browser, including the blended margin that barely moved.
The problem
Blended margin is an average, and averages are where this hides.
Monthly finance review looks at revenue, gross margin, and a variance commentary. All three are blended across every segment, and a segment that is a fifth of orders can lose a third of its contribution margin while the blended number moves less than the mix explanation you already write every month. So it gets written off as mix, again, and the drift compounds until someone models CAC payback for that segment and finds it never paid back.
The insight
Contribution margin per order is stationary. Revenue is not.
Revenue confounds volume with profitability, which is why a growing business cannot read its own margin decline off a P&L. Divide by the order and growth cancels: contribution margin per order, shipping cost per order, and support cost per customer all sit flat under any steady growth rate, so a shift in them is a shift in economics and nothing else. That is exactly the stationarity a changepoint detector assumes, and it is why the same method that goes blind on a revenue chart is decisive on a per-order chart. Then segment it, because the blended average is the second place the signal dies, and correct for the fact that segmenting multiplies the number of tests you are running.
CUSUM on day-over-day percentage changes of every per-unit economics series, Bayesian Online Changepoint Detection (Adams & MacKay 2007) for abrupt shifts, and Benjamini-Hochberg FDR control across every segment × metric combination, because a dozen segments times six cost metrics at α=0.05 with no correction produces several false findings in every run.
How it works
Four steps, no data science team
Your order and payments data, the shipping invoices, the support cost allocation, and the refund ledger. Warehouse tables or direct connectors. Nothing new to instrument.
Revenue minus processing, fulfilment, shipping, support, and refunds, allocated per order and per customer, sliced by acquisition channel, cohort, geography, payment method, and product line.
Each series runs through CUSUM and BOCD. Benjamini-Hochberg across the whole run means a hundred segment-metric tests do not fill the channel with findings that are just the multiple-comparisons problem.
The alert names the segment, the date, and which cost line moved first, and shows the peer segments that did not move. That is the difference between knowing margin fell and knowing which vendor conversation to have.
Who it is for
The finance lead who owns contribution margin
Finance and operations leads at e-commerce and marketplace companies past $10M in annual revenue, where segments are genuinely different businesses and nobody has time to rebuild the unit-economics model every month.
Pricing
- –Full detection engine
- –12-month history
- –Monthly digest
- –Three segments
- –Unlimited segments
- –36-month history
- –Cost driver isolation
- –Slack and email alerts
- –Cohort payback recomputed on every break
- –Warehouse-native deployment
- –Custom cost allocation model
- –SSO and audit log
- –Finance team onboarding
Competition
What exists, and what it does not do
| Who | What they do | The gap |
|---|---|---|
| Looker / Tableau / Metabase | The dashboards your analytics team already built on the warehouse. | Shows the number to whoever opens it. Nothing tests whether it changed, and nobody opens the segment-level view weekly. |
| Profit analytics tools | Blended profitability by product, channel, or SKU for e-commerce. | Reporting layer, not a detection layer. Good at telling you margin by SKU for the quarter, silent on the day a segment broke. |
| NetSuite / accounting stack | The books, closed monthly, with variance commentary. | Monthly cadence and account-level granularity. A segment is not an account, and by close you are four weeks late on something that started on a Tuesday. |
| Anomaly detection in BI tools | Alerting widgets on a metric crossing a threshold or a percentage band. | Per metric, threshold-based, and usually pointed at revenue. Blended margin moving half a point crosses nothing, which is why the incident is invisible in the first place. |
The honest weakness: the numbers we test are only as good as the cost allocation feeding them, and allocation is the part every finance team does differently and half of them do badly. If shipping and support costs are allocated monthly rather than per order, the daily series is an artefact and every changepoint we find is a bookkeeping event. That is not a detection problem we can solve with better statistics, and it means onboarding is real work rather than a connector click. The plausible outcome is that we spend the first month of every account fixing allocation, which is consulting, and consulting does not scale to the price on this page.
Market
Priced against the margin points it recovers, not per seat
A $50M revenue e-commerce business at 40% blended gross margin makes $20M of gross profit. A segment that is a fifth of orders losing a third of its contribution margin for a quarter is a seven-figure question. Three thousand companies at the Operator tier is $25M ARR, and every marketplace and multi-channel retailer has segments that diverge.