Ecommerce & Conversion CRO Audit

Find where a funnel leaks and why, and turn that into evidence-tagged hypotheses, not a list of opinions. The audit's output is a prioritised research-and-test plan with a few safe "just do it" fixes.

Why the method looks like this (evidence)

• AI page reviews alone are unreliable. In Baymard's test, 80% of GPT-4's UX suggestions were false positives; current AI UX tools score 50–75% (https://baymard.com/blog/gpt-ux-audit, https://baymard.com/blog/ai-heuristic-evaluations). AI does well when it only classifies against fixed rules, so the page review here is a binary checklist on rendered screenshots.
• Most changes don't move revenue. 10–33% of tested ideas win at mature companies and median lifts are near zero (${CLAUDE_PLUGIN_ROOT}/skills/experiment-design/references/base-rates.md). Impact claims must be humble and evidence-weighted.
• Research-backed ideas are the industry standard. CXL/Speero ResearchXL and PXL both require analytics, customer and usability evidence behind a test idea. (Neither has published win-rate validation; treat them as sound process.)

Operating principles

1. Separate observation from recommendation. A checklist failure or something you noticed is an observation. It becomes a recommendation only with corroborating evidence.
2. Compare the store with itself first. Its own step rates by device, source and template over time beat any industry benchmark. Benchmarks are context, never the diagnosis.
3. Locate the leak before prescribing. Which step, which segment, then which mechanism.
4. Behavioural levers beat cosmetics. Across 6,700 ecommerce tests, social proof, genuine scarcity/urgency and abandonment recovery outperformed colour, button and CTA-text changes, which averaged about zero (Qubit 2017).
5. Say how sure you are. Every hypothesis carries its evidence sources and a confidence level.

Workflow

1. Brief and instrumentation. Read ./cro-engagement.md. Check analytics hygiene (bots, "direct" share, device mix, a device converting ~0) per references/diagnostics-and-benchmarks.md. If the testing tool is new, recommend an A/A test. Broken measurement is finding #1 when present.
2. Baseline the funnel (first-party data; ShopifyQL queries in references/shopify-specifics.md): sessions → add-to-cart → reached checkout → completed, split by device, source, new vs returning, top templates; compare 90-day vs prior period and vs last year. Name the leaking step and segment.
3. Customer evidence, in parallel. Run voc-research (reviews, tickets, polls, post-purchase survey, user-test plan); delegate bulk coding to the voc-coder agent.
4. Inspect the pages as shoppers see them. Run python3 ${CLAUDE_PLUGIN_ROOT}/skills/ecommerce-cro-audit/scripts/capture_pages.py <urls> --out audit-captures (desktop + mobile, above-fold + full page). Score references/heuristic-checklist.md from the screenshots and facts.json. If capture is unavailable, ask for screenshots; do not score from fetched HTML.
5. Synthesise hypotheses using the card below. Combine checklist failures, funnel data and VoC themes; one hypothesis can cite several. Sort each into a bucket:
   • Instrument — measurement is broken or missing.
   • Just do it — a ★ checklist failure or a bug, cheap and reversible, plus at least one other evidence source. Ship, annotate the date, monitor.
   • Test — 2+ evidence sources and enough traffic (python3 ${CLAUDE_PLUGIN_ROOT}/skills/experiment-design/scripts/power.py mde).
   • Investigate — single-source or heuristic-only; needs research before testing.
6. Prioritise within Test by evidence strength (number and type of sources), reach (traffic on the page), lever prior from the client's test log (python3 ${CLAUDE_PLUGIN_ROOT}/skills/experiment-design/scripts/test_log.py prior), and effort. Prefer PXL-style yes/no criteria over 1–10 gut scores.
7. Plan tests with experiment-design (pre-registration template, sizing, low-traffic path).
8. Deliver with report-information-design; run cro-verifier before it ships.

Hypothesis card (the unit of output)

H<n> · <one-line hypothesis>
Leak: <funnel step + segment + number, e.g. mobile PDP→ATC 2.1% vs desktop 5.4%, 90d>
Evidence: <each source with its finding> — analytics | VoC (theme, share, n) | checklist ID | user test | benchmark
Lever: Cost | Trust | Usability | Comprehension | Motivation
Confidence: High (3+ independent sources) | Medium (2) | Low (1 / heuristic only)
Bucket: Instrument | Just do it | Test | Investigate
Expected effect: likely / possible / speculative (no % promises; see base rates)
Next step: <ship + monitor | test plan id | research needed>
Effort: S | M | L

Reference map

• references/heuristic-checklist.md — the binary PDP / checkout / landing checks, with ★ items backed by large-sample research.
• references/diagnostics-and-benchmarks.md — traffic-quality checks, step-conversion, benchmark context (with caveats), the five-gate lens.
• references/conversion-playbook.md — the main levers and what "good" looks like, with sources.
• references/checkout-mobile-accessibility.md — Baymard checkout data, mobile, accessibility, search, retention flows.
• references/buyer-psychology.md — premium/considered-purchase psychology (generic); examples/premium-footwear-engagement.md shows one applied case.
• references/persuasion-and-mental-models.md — Cialdini, mental models, pricing psychology.
• references/shopify-specifics.md and references/liquid-product-template-checklist.md — Shopify data pulls, theme anatomy, build checklist.
• references/beyond-ecommerce.md — landing, lead-gen, SaaS signup.
• references/experimentation.md — pointer to the experiment-design skill.
• scripts/capture_pages.py — rendered desktop/mobile screenshots + page facts.

Engagement context

At the start of a task, read ./cro-engagement.md. Use its tier, ICP, surface, goal metric, traffic per template, research assets and test-log path. If it's missing and the work is more than a one-off, offer to capture one via getting-started.

Related skills

voc-research (customer evidence), experiment-design (sizing, readout, test log), deep-research-synthesis (external evidence), report-information-design (deliverable), conversion-copywriting (variants). Agents: cro-auditor runs steps 2, 4 and 5; voc-coder codes feedback; cro-verifier checks the output.

Guardrails

• Never present a heuristic-only observation as a recommendation or promise a % lift.
• Never use synthetic users or "simulated shoppers" as evidence.
• Don't recommend fake urgency or fabricated social proof.
• Report what is working too; don't invent issues to fill the list.
• Keep the traffic problem (acquisition quality) separate from the page problem.