What went right, what went wrong

Ten small charts, one per server, same axes, sixty months each. Eight of them are boring - and boring is the finding, because it tells you the break is not seasonal, not macro, not a platform-wide price move. Two of them step down in the same month. When ten units share a market and two break together, the cause is something those two share and the others do not: in this case, one fulfilment set-up.

Take the three months before the break, December 2023 to February 2024, and the three months after, April to June 2024. Orders: about the same. Average price per unit: about the same. Unit cost: up roughly 22 percent. Margin: down thirteen points. That is a fingerprint, and it rules things out. A demand problem moves volume. A pricing problem moves price. A cost problem moves neither - which is exactly why a growth dashboard cannot see it.

The seasonal trap here is real: compare March to December and you compare a quiet month to a peak. Comparing the three months either side of the decision holds the season roughly still and lets the decision show through.

Profit change between two periods decomposes exactly into four terms: how much came from selling more or fewer units, from charging more or less per unit, from paying more or less per unit, and from selling a different mix of products. On S07 and S08 the terms read volume -18k, price +6k, cost -54k, mix +5k. Cost is the whole story. Run the same bridge on the other eight servers and the cost term is zero - a second confirmation that the break was a decision, not a market.

Describing a break is not the same as pricing it. Take the margin the two servers earned before the shift, 42.4 percent, apply it to every month of revenue since, and the gap between that and the profit actually booked is the cost of the decision: 251k over thirty months, an average of 8.4k a month. The revenue itself kept growing in 2024, which is why nobody with a revenue chart felt any pain.

The number rests on one assumption and it must be said out loud: that the March cost change caused the margin drop rather than coinciding with it. Two things make the assumption defensible here. The other eight servers, same months, same market, show no cost term. And the two-event test in Part 2 shows the customer effect in both surviving-merchant and churned-merchant customers, so the change reached the customer experience itself. Neither is a counterfactual, and the analyst notes say so. A number with its assumption attached survives a hostile board; a number without one does not.

• State the counterfactual. "At the old margin" is a choice; write it on the chart.
• Say where it is tested. The eight quiet servers and the two-event split are the tests here.
• Do not add the retention loss yet. That is a second estimate with a second assumption. Stack them and the board discounts both.

A customer belongs to the quarter of their first order. Ask what share of each quarterly cohort came back within the next two quarters. On S07 and S08 that share ran at 28.8 percent for the 2023 cohorts and 17.2 percent by 2025. On the other eight servers: 27.4 percent, then 26.8. Same company, same products, same years - the only thing that differs is which servers the customer bought on. The first cohort to fall is the one acquired around September 2024, two quarters after the cost decision.

The platform's own customers are merchants. Count them the way you count customers: new, retained, churned, net, every quarter. The Vietnam servers were the growth engine from 2021 to 2023 - then the merchant churn bar grows through the second half of 2024. The survival view says the same thing more precisely: before June 2024 a key-server merchant had a 1.26 percent chance of leaving in any given month; after it, 2.39 percent. The hazard doubled in the quarter the merchant-success team was cut.

The method matters for one reason a chairman should know: a plain average of merchant lifetimes throws away every merchant who has not left yet, so it flatters the past and hides the present. Survival analysis counts the still-here merchants as "at least this long" and gives you a rate you can compare before and after a date.

Here is the honest problem with the story so far. Two decisions landed three months apart on the same servers: the fulfilment change in March and the merchant-success cut in June. The repeat rate fell. Which lever pulled it? Split the key-server customers by whether their merchant survived to 2026. If the drop lives only among customers of merchants who later left, the June cut did it - customers left because their shop did. If it shows in customers of surviving merchants too, the buying experience itself changed.

So the March decision reached the customer, and the June decision made it worse. That ordering is what c3 needs, because it means reversing or re-pricing the fulfilment change is the first lever and restoring merchant success is the second - not the other way round, and not either alone.

The store dimension says 342 stores exist. A store exists in that table until someone closes the record, and nobody closes records - roughly 65 percent of stores that have not sold in a year are still marked open. Define active as at least one order in the trailing 90 days and the share of dormant stores climbs through 2025, concentrated on the key servers. Dormancy is the earliest merchant-churn signal there is: a merchant stops selling long before anyone files paperwork.

• Count active, not existing. Existence in a table is a fact about the table.
• Status as events, not fields. A field that gets overwritten has no history; an event that gets appended does. c3 puts this in the tech table.
• Dormancy is a list, not a chart. Each dormant store on a key server is a merchant somebody should call this week.

Every product line exists twice: the platform's standard version and the merchant's customised version. The customised SKUs earn 8 to 10 margin points more. That alone would be a pricing note. The finding is what happens to the customers: group them by how much of their merchant's catalogue is customised, and customers of high-customisation merchants come back at 33.9 percent within twelve months against 28.1 percent for low-customisation merchants. Better margin and better retention on the same lever is rare; when you find one, it is the growth story.

Look at what customers put in the same basket without being asked. A mug with a t-shirt. A poster with a cushion. A notebook with a sticker pack and a pen set. Each pair co-occurs 1.3 to 1.6 times more often than chance, and the attach rate - how often the second item appears given the first - sits above the second item's base rate every time. Customers are already assembling these bundles one item at a time. The platform has never priced, promoted or packaged any of them.

One caution on reading lift: a pair with three co-purchases can show a huge lift and mean nothing. The bundle candidates here clear a support threshold first, then lift. Ask your analysts for both numbers, always.

Five categories, fifteen lines, five years. Drinkware and home spiked in 2020 and settled. Stationery has faded every year since 2023. Bags have risen steadily and nobody planned it. A category chart by year shows the shape; a slope chart of 2021 against 2025 by line shows which lines to lean into and which to stop defending. The catalogue is the one part of the business that changed for the better without a decision, which means the decision is now overdue.

Merchant churn in month thirty was predictable in month three. Take every merchant onboarded by 2023, so all of them have had time to leave or stay, and look at two early signals: months from onboarding to first sale, and revenue in the first ninety days. Merchants that later churned took a median 3.8 months to make a first sale; merchants still trading took 1.3. Median first-90-day revenue: 0 for the ones who left, 399 for the ones who stayed. The merchant-success team never had this list. It measured the funeral.