Where you are
You can build the GMV tree (b3) and you know what each funnel gate measures (b4). Today those two skills earn their keep: a real-looking drop lands, and you walk it down the tree until it has a name, a size, an owner, and a fix date. This is the second live-simulator session - the same funnel-deep tree from b3 returns as your diagnosis scratch pad.
Monday, 9:04am: GMV is down 12% 12 min live
The dashboard says Mango Lane's GMV fell 12% week over week. Marketing suspects the site redesign, product suspects the new ad audience, and the founder suspects everything. The worst response is to guess. The best response is a fixed walking order - the same five steps, every drop, every time - so the answer arrives before the opinions harden.
| Step | The question | The tool | What you get |
|---|---|---|---|
| 1 · Real? | Tracking change? Seasonal gravity? | Deploy log + a5's YoY comparison | Permission to proceed - or to stand down |
| 2 · Lever | Sessions, CR, or AOV? | Top tree + pass-through math | The guilty lever |
| 3 · Gate | Which funnel gate? | Deep tree (the b3 simulator below) | The owning team |
| 4 · Slice | Which channel / device / cohort? | Segment cuts (b6) | The location of the break |
| 5 · Write | So what, who, by when? | The finding template (Demo 2) | One paragraph that ends the thread |
LiveStep 1 - confirm the drop is real4 min▶
Before touching the tree, rule out the two impostors:
- Tracking changes. Did anything ship last week that touches the analytics tags - a site release, a consent-banner change, an app update? A drop that starts the exact hour of a deploy is a tracking suspect until proven otherwise.
- Seasonal gravity. a5's rule: compare year over year, not just week over week. If GMV falls 12% WoW every first week of the month (payday cycles, promo calendars), the drop is the calendar, not the business. Same week last year down a similar amount? Stand down.
Only when the drop survives both checks do you open the tree. Most Monday panics do not survive them.
LiveSteps 2-5 - walk the tree, then write4 min▶
Step 2 - top tree. GMV = sessions x session CR x AOV. Compute each driver's WoW percentage change. By b3's pass-through rule, the driver moves should roughly sum to the GMV move. One of the three will stick out.
Step 3 - deep tree. If CR is the guilty lever, open it into the four gates: PDP penetration, add-to-cart, checkout start, checkout completion. The gate that moved is the gate that names the owning team.
Step 4 - segments. Cut the guilty gate by channel, device, and new vs returning to find WHERE it broke. That is b6's whole session - today you just note the pointer.
Step 5 - size it and write the finding. One paragraph: what fell, by how much, traced to what, suspected cause, owner, action, re-check date. Demo 2 gives you the template.
Self-studySizing: turn the percent into dollars3 min read▶
"GMV fell 12%" gets attention. "This costs us about $79,000 a month if it holds" gets a fix scheduled. Sizing is one multiplication: the percentage fall x the baseline. At Mango Lane, 12% x $660,960 = $79,315 per month of run-rate GMV at risk.
- Sizing sets priority. A $79K/month leak justifies pulling an engineer off the roadmap tonight; a $2K/month leak joins the backlog. Without the dollar figure, urgency is decided by whoever shouts loudest.
- Size the fix window too. If the drop started Friday and gets fixed Thursday, the realized cost is roughly a week's slice, not the monthly figure. Say both numbers - "about $18K lost so far, $79K/month pace if unfixed".
- Round honestly. $79,315.20 pretends to a precision the estimate does not have. "About $79K" is the analyst's number.
Recreate the drop in the live tree ★ 15 min · everyone builds
The funnel-deep tree from b3 returns, at Mango Lane's healthy baseline. Your job today is the reverse of b3: there you broke the tree to see what happens; here you are handed what happened and must find the break that produces it. That reversal is diagnosis.
{
"unit": "$",
"root": {
"label": "GMV", "op": "x",
"children": [
{ "label": "Orders", "op": "x", "unit": "orders", "children": [
{ "label": "Sessions", "value": 360000, "unit": "sessions" },
{ "label": "PDP penetration", "value": 0.40, "pct": true },
{ "label": "Add-to-cart rate", "value": 0.25, "pct": true },
{ "label": "Checkout start", "value": 0.45, "pct": true },
{ "label": "Checkout completion","value": 0.60, "pct": true }
] },
{ "label": "AOV", "value": 68, "unit": "$" }
]
}
}
Drill 1 - the single-lever drop. Recreate Monday's report: cut Checkout completion from 60 to 52.8 (that is exactly -12%). Confirm GMV shows -12%. One lever, full pass-through - if the whole 12% traces to one gate, your diagnosis is already 80% written.
Drill 2 - the compound drop. Reset. Now cut Sessions by 5% (360,000 to 342,000) AND Add-to-cart from 25 to 23.25 (that is -7%). Confirm GMV shows about -11.7%: 0.95 x 0.93 = 0.8835. Real drops are often two medium moves, not one big one - the pass-through math is how you check your explanation actually adds up.
Drill 3 - the reflex builder. Reset, then press Simulate a drop repeatedly. Each time, race to name the guilty gate out loud before reading the note. Ten rounds of this and the walk from "GMV fell" to "checkout completion fell" stops being a procedure and becomes a reflex.
Why checkout gets checked early: checkout is where the most machinery lives - payment providers, address validation, promo codes, 3-D Secure - and industry-wide about 70% of carts are already abandoned there in the best of times (Baymard Institute's running average is 70.22%). A gate that fragile plus a Friday-evening payment deploy is the single most common true answer to a Monday GMV drop.
Self-studyWhy compound drops fool smart people3 min read▶
Drill 2's drop is sneakier than drill 1's, and it is worth naming why. A single -12% lever screams from the dashboard. Two levers at -5% and -7% each look like ordinary weekly wobble on their own charts - neither triggers an alert, neither gets a Slack thread - yet together they produce almost the same top-line damage.
- Do not stop at the first suspect. Finding one lever down 5% does not explain a 12% drop. Run the pass-through check: do the falls you found multiply out to the fall you are explaining? 0.95 alone = -5%. Not enough. Keep walking.
- Compound drops often mean compound causes. Sessions -5% (a paused campaign) and add-to-cart -7% (a PDP image bug) can land the same week by pure coincidence. One finding per cause, two owners, two re-check dates.
- The tree keeps you honest both ways: it stops you blaming one lever for everything, and stops you declaring victory at half an explanation.
Write the finding ★ 10 min · everyone writes
A diagnosis that lives in your head ends nothing. The deliverable is one paragraph, posted where the argument is happening. Same skeleton every time - fill the brackets and hit send.
LiveThe one-paragraph finding template4 min▶
"GMV fell [X]% WoW. The drop traces to [lever/gate], which fell [Y]% ([baseline] -> [now]). [Slice detail if known]. Suspected cause: [hypothesis]. Owner: [team]. Action: [fix]. Re-check: [date]."
Every bracket is a claim you can defend from the tree. No adjectives, no blame, no "it seems" - numbers, a name, and a date.
LiveThe template filled in - Monday's drop4 min▶
"GMV fell 12% WoW. The drop traces to checkout completion, which fell 12% (60% -> 52.8%). Concentrated in mobile sessions paying by card. Suspected cause: payment provider latency spike after Friday's gateway update. Owner: payments team. Action: roll back the gateway config and re-test the card flow. Re-check: Thursday's numbers, posted here."
Notice what this paragraph did to the Slack thread: marketing's redesign and the new ad audience are both off the hook without anyone arguing, and there is exactly one next step with exactly one name on it.
Self-studyKeep a finding log3 min read▶
Post the finding where the argument is happening, then copy it into one running document - a simple dated list of every finding you have written. Three payoffs arrive within a quarter:
- Repeat offenders surface. If checkout completion stars in four findings in three months, the payment stack needs a project, not another rollback. The log turns incidents into a pattern.
- Re-check dates actually get checked. A finding with a date in a log is a finding someone follows up on. A finding in a scrolled-away Slack thread is a wish.
- Your hit rate becomes visible. When the Thursday re-check confirms the fix, say so in the log. A track record of confirmed findings is what makes the next one believed on sight - and it feeds directly into b7's weekly review pack.
False positives self-study
Self-studyWhen the tracking breaks, not the business4 min read▶
The nastiest false positive: a tracking outage masquerading as a conversion drop. The purchase events stop arriving - a tag broken by a release, a consent change, an SDK update - but the orders keep landing in the payment and order systems just fine. Your dashboard shows checkout completion cratering; your warehouse is shipping at full speed.
- The cross-check: before declaring a business problem, compare event-counted orders (from analytics) against order-system orders (from the platform's order admin or finance) for the same period. If the order system shows normal volume while analytics shows a cliff, you have a measurement problem, not a demand problem.
- The tell: tracking breaks tend to be sharp cliffs starting at a deploy timestamp, often on one platform (app vs web) only. Real demand problems usually slope.
- The payoff: this one check has saved more analyst credibility than any model ever built. Declaring a fake drop real burns trust in every future finding you write.
Try it yourself - this week ◐ 20-30 min total
- Open your own store's numbers and find last month's worst day. Run the five steps on it for real: real check, top tree, deep tree, one segment cut, sized finding.
- Write the finding paragraph using the template - even if the honest answer is "seasonal gravity, no action". Practicing the write-up on a boring answer is still practicing the write-up.
- In the live tree above, invent a compound drop of your own (two levers, different sizes), predict the GMV move with 0.9-something x 0.9-something arithmetic first, then check yourself against the simulator.
- Find out where your company's order-system order count lives (Shopify admin, finance report, wherever). You want that cross-check bookmarked before you need it at 9am on a Monday.
Sources this session draws on
Diagnosis is a craft workflow built on the course's own earlier machinery plus one industry benchmark. This page draws on:
Three questions before you go 🎯 ◐ 90 seconds
1 · GMV is down 12% WoW. What is the very first thing you do?
Step 1 always comes first: rule out tracking changes and seasonal gravity (a5's YoY rule) before opening any tree. Most Monday panics die at step 1 - and diagnosing a drop that never happened costs a week.
2 · Sessions fall 5% and add-to-cart falls 7% in the same week. GMV falls about...
Multiply the survival rates: 0.95 x 0.93 = 0.8835, so GMV lands at 88.35% of baseline - a fall of about 11.7%. Small percentage moves roughly add (5 + 7 = 12 is close), but the multiplication is the exact check.
3 · Which of these is a complete finding?
A finding names a lever, a size, an owner, and a re-check date. Option B has all four. The others are vibes - no owner, no size, no date, no end to the Slack thread.