· Updated · 4 min read
The template I use to map AI workflows before writing a spec.
No spam, unsubscribe anytime. I only email teardowns.
A deep dive that changes metrics has two phases. First, instrument the funnel until you can name each drop-off. Second, form ranked hypotheses. At Sierra, 480K sessions became three fixable problems in two weeks — because we never diagnosed an aggregate, only a funnel.
"Cart abandonment is too high." That sentence sat in our metrics review for months, and everyone nodded, and nothing moved — because an aggregate number is unanswerable. When we finally stepped through 480K sessions properly, "abandonment" turned out to be three completely different problems wearing one trench coat. The failure happens four steps before the diagnosis: no events, no recordings, no funnel, no ranked hypotheses — and then a hunch gets a roadmap slot.
You cannot fix a funnel you cannot step through. The deep dive starts with one ask: name every step.
Only when you can answer "which step loses the most, and which looks like a single cause" do conclusions become credible.
An event-per-step pattern looks roughly like this — the exact fields vary by funnel, but the principle doesn't:
// Illustrative — one event per checkout step, not one "checkout started" aggregate
dataLayer.push({ event: 'checkout_start', step: 'cart', step_index: 0, value: cartTotal });
dataLayer.push({ event: 'checkout_step', step: 'shipping', step_index: 1, option: shippingMethod });
dataLayer.push({ event: 'checkout_step', step: 'payment', step_index: 2, option: paymentMethod });
dataLayer.push({ event: 'checkout_step', step: 'review', step_index: 3 });
dataLayer.push({ event: 'purchase', value: orderTotal, currency: 'USD', transaction_id: orderId });
Every step gets an event. No "checkout started" aggregate — you need the staircase.
Microsoft Clarity is free and unsampled, which matters more than any premium feature:
Don't just chart in GA4. Stream the raw events to BigQuery via the GA4 BigQuery export so you can interrogate them directly — a drop-off-by-step query, segmented by device, looks roughly like this:
-- Illustrative — drop-off by step, segmented by device
SELECT
step,
device.category,
COUNT(DISTINCT user_pseudo_id) as users,
COUNT(*) as events,
COUNT(DISTINCT user_pseudo_id) / LAG(COUNT(DISTINCT user_pseudo_id)) OVER (ORDER BY step_index) as dropoff_rate
FROM `project.analytics_events`
WHERE event_name IN ('checkout_step', 'purchase')
GROUP BY step, device.category, step_index
ORDER BY step_index;
SQL like this is the deep dive. The dashboard is the summary.
Aggregate data lies politely; a funnel stops lying. Take the event tables and rank drop-offs:
| Rank | What you're testing |
|---|---|
| 1 | The step with the biggest unique loss |
| 2 | The step most likely to be one cause |
| 3 | The step that is easiest to measure after |
At Sierra this drill surfaced three distinct checkout problems — shipping surprise, promo rage-click, payment dead-end — that one aggregate percentage had been hiding. Each got an independent hypothesis, an independent fix, and an independent measurement window, so when the number moved we knew exactly which change owned it. The full breakdown — what each problem looked like in the session recordings, what shipped, and the before/after numbers (73.1% → 53.9% abandonment, +47% mobile CVR, +49% checkout completion) — is in the checkout teardown.
| Hypothesis | Owner | KPI | Window | Evidence needed | Result | Decision |
|---|---|---|---|---|---|---|
| (one sentence, tied to a specific step) | (name) | (the one metric that proves it) | (fixed, e.g. 2 weeks) | (what the recording/event data must show) | (fill after the window closes) | (keep / kill / iterate) |
Fill this before you write a single event. If you can't fill the "Evidence needed" column, you don't have a hypothesis — you have a hope.
If your dashboard and your deep dive agree, your deep dive failed. The whole point of disaggregating is to find what the average was hiding.
The same instrument-first discipline is what found the single worst field in the lead-form fix and the intent-segmentation insight behind the category-page redesign — different funnels, same method. The full checkout numbers live in the cart & checkout case study.