FerroxLabs As of: 2026-05-16
funnel-diagnosis
Mode skill. Default-enabled on the Analyst specialist.
When to use
Use any time a user asks "why isn't this converting?", "where are we losing people?", "what's broken in our funnel?", or hands you a topline number that has dropped or stalled. Use before recommending any page change, copy test, or channel reallocation. If a teammate proposes a fix without naming the specific drop-step it addresses, run this mode first.
Trigger phrases:
- "Conversions are down."
- "Why isn't this working?"
- "We're getting traffic but no signups."
- "Should we redesign the landing page?"
- "What should we test?"
If a current drop-point and its denominator already sit in TEAM_MEMORY.md under ## Analyst, skip to step 4.
Procedure
1. Lay out the full funnel with denominators. Write each step on its own line with absolute count and rate-against-previous-step. Common shape: Visits → LP-action → Signup → Activation → First retained behavior → Repeat. If a step has no instrumented event, name the gap. Do not estimate over an uninstrumented step.
2. Compute relative drop at each step. A step losing 50 of 100 is worse than one losing 1,000 of 10,000. Sort steps by drop-rate descending. The top is the diagnosis target — a step further down cannot move the topline enough to matter until the bigger leak is plugged.
3. Check sample sufficiency. Count conversions through the suspect step in the window. If fewer than 30, refuse to call it a drop — call it "directional, n too small, recheck in N days." Compute the days needed at current run-rate, name the date, stop.
4. Segment the drop. Cut the suspect step by channel, device, audience, landing page, time-of-day. The cut with the largest between-segment gap is the lead. If no segment shows a meaningful gap, the issue is structural (the step itself), not selection (the audience mix).
5. Hand off the why. The diagnosis tells you where. The why requires session recordings, interviews, support tickets, exit-intent surveys. Name the segment and step; route to Research. Do not recommend a fix from numbers alone.
6. Stamp it. Write drop-step, denominator, window, segment cut, and sample-sufficiency note to TEAM_MEMORY.md under ## Analyst. Product, Copy, and Channels will read this before proposing changes.
Decision rules
- Biggest leak first, always. Even a brilliant fix to the second-worst step gives you a smaller topline lift than a mediocre fix to the worst.
- Denominator before numerator. A 40% conversion rate on 12 visitors is not a 40% conversion rate. State the denominator before stating the rate.
- One window per claim. Comparing a 7-day rate to a 30-day rate is not a comparison; it is a category error. Lock the window before reporting.
- Channel mix shifts before page changes are tested. If the channel mix changed in the window, the funnel rates changed because the audience changed. Hold mix constant, or segment by channel, before claiming a page or product issue.
- No fix without a why. The dashboard does not say why. Route to Research before approving a build.
Anti-patterns
- Reporting overall conversion rate as the diagnosis. Overall CR is a symptom, not a leak. The leak is at a step.
- Recommending a redesign based on "the landing page is bad." Without the drop-step and segment evidence, that is taste, not analysis.
- Comparing this week to last week with no significance test and no mention of sample size. Week-over-week swings on small n are almost always noise.
- A/B testing the second-worst step. Wasted variance budget — fix the biggest leak first.
- Hiding the gaps. If a step is uninstrumented, the diagnosis is not "we don't see a problem there." It is "we can't see there yet."
Before / after
Brief: "Conversions dropped 20% this week. What do we test?"
Before (no funnel, no segment):
Test a new headline on the landing page.
After (drop-step + sample + segment + handoff):
Funnel last 7 days: 18,400 visits → 4,200 LP-action (22.8%) → 1,900 signup (45.2%) → 410 activation (21.6%) → 92 retained-7d (22.4%). Biggest relative drop: visit→LP-action (77.2% loss). Segment cut shows paid-social traffic at 14% LP-action vs. organic at 31% — same week last month paid-social was at 27%. Sample on paid-social segment: 8,200 visits, 1,150 LP-action, n sufficient. Diagnosis: paid-social audience-message fit decayed this week. Routing to Research for ad-creative-vs-landing-page review before any page test. Next measurement: re-run cut after 7 days of held creative.