cro-optimizer

CRO specialist that pulls live analytics data via the Humblytics MCP, analyzes conversion funnels, identifies drop-off points, and generates prioritized A/B test hypotheses. Use when analyzing convers

SKILL.md

GitHub

CRO Optimizer

Purpose

Analyze conversion funnels using live Humblytics analytics (via the Humblytics MCP), identify the highest-impact drop-off points, and generate prioritized A/B test hypotheses with expected conversion-rate / volume impact. This skill turns raw analytics into a ranked optimization roadmap.

When to Use

  • Diagnosing why a funnel is underperforming
  • Identifying the biggest conversion bottleneck across a user journey
  • Generating a prioritized list of A/B test ideas
  • Preparing a CRO sprint plan or quarterly optimization roadmap
  • Analyzing page-level or step-level drop-off rates
  • Comparing conversion performance across segments (device, source, geography)

Credentials

Live data comes from the Humblytics MCP (server humblytics). Connect it once — see the repo README — and this skill calls mcp__humblytics__* tools directly. No API keys, base URLs, or .env plumbing live in the skill.

  • Never paste API keys into chat — they persist in transcripts and logs. Your key lives in the MCP connection headers, set once at connect time, and is never committed to the repo.
  • Property: the MCP auto-resolves your property for a single-property key (the common case). For a multi-property key, call list_properties and pass the propertyId you want to analyze.
  • Docs: https://docs.humblytics.com/api

Before You Start

  1. Confirm the property — The MCP auto-resolves the property for a single-property key. If the key covers multiple properties, call list_properties and confirm which one to analyze.
  2. Identify the funnel — Clarify which conversion flow to examine (e.g., homepage > pricing > signup > onboarding)
  3. Check for context — Look for existing project docs, AGENTS.md, or product briefs that describe the business model, target audience, and current conversion goals
  4. Establish the time range — Default to last 30 days; ask if the user wants a different window
  5. Confirm MCP access — Verify the Humblytics MCP (server humblytics) is connected before pulling data

Core Workflow

Step 1: Pull Funnel Data

Live data comes from the Humblytics MCP — the relevant tools here are get_pages_breakdown, get_page_details, query_funnel, get_funnel_sankey, get_forms_breakdown, and get_clicks_details. Retrieve:

  • Page views and sessions for each step in the funnel
  • Event data for key conversion actions (signups, clicks, form submissions)
  • Device and source breakdowns to identify segment-specific issues
  • Scroll depth via get_page_details and click density via get_clicks_details (there is no heatmap tool) for high-traffic pages

The analytics tools take start, end (ISO-8601), and timezone:

  • get_pages_breakdown — Page-level traffic across the site
  • get_page_details (page: "/path") — Single-page deep dive (UTM, device, country breakdowns, scroll depth)
  • query_funnel (steps: [{ page: "/" }, ...]) — Funnel step data. Optional: mode: "unbounded" | "sequential", breakdownBy
  • get_funnel_sankey (same steps) — Sankey path diagram for the same funnel

Fallback when funnels are down. query_funnel and get_funnel_sankey can return HTTP 500. If they fail, approximate the funnel from the tools that do work: pull per-step page volume from get_pages_breakdown (and get_funnel_suggestions with page: "/path" for the ranked next-page sequence), and pull conversion-event volume for the final step(s) from get_forms_breakdown. Compute step-to-step conversion / drop-off from these unique_sessions (pages) and submission counts (forms). Note in your output that the funnel is an approximation from page + form breakdowns because the native funnel tool was unavailable.

  • get_forms_breakdown and get_forms_details (page: "/path") — Form/conversion event data (there is no generic events tool)
  • get_clicks_details (page: "/path") — Click heatmap data for a specific page (no heatmap tool exists; click data is the closest analogue)
  • get_clicks_breakdown — Cross-page click comparison

Step 2: Map the Funnel

Build a complete picture of the user journey:

Traffic Source → Landing Page → Key Action → Conversion → Retention

For each step, calculate:

  • Volume: How many users reach this step
  • Conversion rate: Percentage who proceed to the next step
  • Drop-off rate: Percentage who abandon at this step
  • Absolute drop-off: Raw number of users lost

Step 3: Identify the Biggest Leak

Apply the Largest Leak First principle:

  1. Calculate the absolute number of users lost at each step
  2. Rank steps by absolute drop-off (not percentage)
  3. The step losing the most users in absolute terms is your highest-priority optimization target

Why absolute over percentage: A 50% drop-off at a step with 100 visitors loses 50 people. A 10% drop-off at a step with 10,000 visitors loses 1,000 people. Fix the 1,000-person leak first.

Step 4: Diagnose Root Causes

For each high-drop-off step, investigate:

  • Page load time — Slow pages kill conversions. Check if the step has performance issues.
  • Mobile vs desktop — Is the drop-off concentrated on mobile? Layout/UX issue.
  • Traffic source — Do certain acquisition channels show higher drop-off? Expectation mismatch.
  • Scroll depth — Are users seeing the CTA? Check scroll depth via get_page_details and click density via get_clicks_details (there is no heatmap tool).
  • Click patterns — Are users clicking non-interactive elements? Confusing UI.
  • Form fields — For forms, which field has the highest abandonment rate?

When the leak appears concentrated in paid traffic (drop-off significantly worse for utm_source=google or utm_source=facebook than for organic), call get_ads_attribution (startDate, endDate as YYYY-MM-DD) to see which specific campaigns are landing on the underperforming page. A creative/landing-page mismatch on one campaign can drag down a whole step's conversion rate. Hand off to revenue-attributor for the full ROAS picture or ad-expert to fix the creative.

Step 5: Generate Test Hypotheses

For each identified issue, create a hypothesis using the ICE framework:

Format:

IF we [change], THEN [metric] will [improve/increase/decrease]
BECAUSE [evidence from data]

Impact: [1-10] — How much will this move the needle?
Confidence: [1-10] — How sure are we this will work?
Ease: [1-10] — How quickly can we implement and test this?
ICE Score: [average of three]

Step 6: Prioritize and Recommend

Rank all hypotheses by ICE score and present:

  1. Top 3 Quick Wins — High ease, decent impact (ship this week)
  2. Top 3 High-Impact Tests — High impact, may require more effort (sprint backlog)
  3. Strategic Bets — Lower confidence but potentially transformative (quarterly roadmap)

For each recommendation, include:

  • The specific page or funnel step
  • What to change and why
  • Expected impact on conversion rate
  • Suggested test duration based on traffic volume

Analysis Frameworks

The RICE Prioritization (for larger teams)

  • Reach: How many users per month does this affect?
  • Impact: Expected lift (minimal / low / medium / high / massive)
  • Confidence: Data quality supporting the hypothesis (low / medium / high)
  • Effort: Engineering/design time (days)

Score = (Reach x Impact x Confidence) / Effort

Segment Analysis Checklist

Always break down conversion data by:

  • Device type (mobile / desktop / tablet)
  • Traffic source (organic / paid / direct / referral / social)
  • Geography (if international)
  • New vs returning visitors
  • Entry page

Common Funnel Archetypes

Funnel Type Key Metrics Common Leaks
SaaS Free Trial Visit > Signup > Activate > Convert Signup form friction, activation failure
E-commerce PDP > Cart > Checkout > Purchase Cart abandonment, checkout form
Lead Gen Landing > Form > Thank You Form length, trust signals
Content > Conversion Blog > CTA > Signup CTA visibility, relevance match

Output Format

Present findings as:

  1. Funnel Overview — Visual step-by-step with volumes and rates
  2. Key Finding — The single biggest insight in one sentence
  3. Drop-off Analysis — Ranked list of leaks with absolute numbers
  4. Root Cause Diagnosis — What is causing each major leak
  5. Prioritized Test Roadmap — ICE-scored hypotheses ready for execution
  6. Expected Impact — If top 3 tests succeed, projected conversion lift

Related Skills

  • ab-test-generator — Take the hypotheses from this skill and generate actual test configurations
  • funnel-reporter — Pull comprehensive funnel reports with traffic & conversion volume (revenue only if a Stripe/ChartMogul revenue connector is attached)
  • page-cro — Deep-dive into a specific page's conversion issues
  • copywriting — Generate optimized copy for test variants

Shared Frameworks (REQUIRED reading)

Before producing recommendations, anchor your analysis against the shared primitives in skills/_shared/. Skipping these is the #1 cause of generic, low-confidence output.

  • _shared/frameworks/preflight-checklist.md — five context items to verify before scoring (URL, time range, goal, vertical, statistical reachability). If anything's missing AND would change the recommendation, ask one focused question; otherwise state assumptions explicitly.
  • _shared/frameworks/largest-leak-first.md — rank by absolute people lost, not by percentage drop. A 10% drop on 10,000 visitors outranks a 50% drop on 100. Always compute absolute loss per step before applying ICE.
  • _shared/frameworks/percentile-framing.md — report current metrics against vertical p25/p50/p75 bands from _shared/benchmarks/baselines.json. Replace "your CVR is low" with "your CVR is at p35 — meaningful headroom to p50".
  • _shared/frameworks/ice-confidence-rubric.md — anchor ICE.Confidence on evidence quality, not familiarity. 9–10 = ≥2 sources with n≥1000 in target vertical; 5–6 = general best practice; 1–3 = directional hunch.
  • _shared/frameworks/anti-patterns.md — counter-evidence for canon advice (customer logos LOST in most DoWhatWorks tests, "free" CTAs lose −16.8% platform-wide on Unbounce, hero video net-negative on mobile, etc.). Read before recommending the "best practice" version of any well-known pattern.
  • _shared/frameworks/base-rate-priors.md — realistic priors: only ~14% of CTA tests reach significance; ~31% of headline rewrites beat control. Anchor expectations against base rates, not best-case outliers.
  • _shared/benchmarks/patterns.json — 54 curated patterns with cited lift ranges, prerequisites, anti-patterns. When recommending a change, find the matching pattern_id and quote evidence[].lift_range_pct instead of guessing.