Cm analytics audit
Skill buntysomroy/compound-marketing/skills/cm-analytics-audit
Use when you say '/cm-analytics-audit', 'audit the tracking', 'is the purchase tag firing', 'verify the conversion tracking', 'why is All-Conv way higher than Conv', 'check gclid/utm capture', 'GTM audit', 'conversion double-count', 'tracking is broken', 'conversions look off', or when a plan action is hard-blocked behind an LP/conversion-tracking gate before any pause/restructure. Compound Marketing — the ANALYTICS-DEBUGGING audit. Verify a client's conversion-tracking end-to-end in the live systems (storefront tags → GTM container → Google Ads/GA4 conversion actions), diagnose double-counts / dead actions / gclid-capture gaps, and package the exact fix.From its SKILL.md
npx -y skills add buntysomroy/compound-marketing --skill cm-analytics-auditAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
10.3 KB, as published. Nobody here has run it
/cm-analytics-audit — Compound Marketing: analytics-debugging audit
Where this sits. A diagnostic sibling of the CM pipeline (
/cm-audit→/cm-analyze→/cm-plan→/cm-review→/cm-execute). It runs when a/cm-planaction is hard-blocked behind a conversion-tracking gate ("verify the canonical purchase action / gclid capture before pausing anything"). Its output feeds the blocked plan action, and its artifact is a<!-- cm:solution -->doc — the same shape/cm-plansingle-problem mode emits.Stage contract (read FIRST, every run):
reference/protocol-cm-stage-contract.md— the five contract steps (decisions recall, findings confirmation, quantitative-claim rule, handoff block, decision-time logging) are mandatory for this stage. This skill is the thin driver; do not improvise contract mechanics from memory.Full CM pipeline reference:
reference/sop-cm-pipeline.md
This is the audit stage for measurement itself — turn "is the tracking right?" into a verified, sourced diagnosis and a packaged fix. Verify every finding in the live system; never infer a tag/action state from the storefront alone or from memory.
Core principle
A tracking claim you did not see in the live GTM container + the live Google Ads conversion-actions screen is a guess. The storefront tells you what loads; only the container + the Ads goals screen tell you what counts. Trace each purchase from dataLayer/pixel → GTM tag (ID + label + trigger) → Google Ads conversion action (Primary/Secondary + count) before concluding anything.
Global rules (read first)
- Confirm access before driving. Client browser work should run in a browser profile authenticated to the client's ad/analytics accounts; confirm the logged-in Google account first (wrong profile reads as a false "no access").
- Verify each finding live. Storefront inspection is necessary but NOT sufficient — the double-count, dead actions, and Primary/Secondary status only exist in GTM admin + Ads. Flag any un-verified inference with "⚠️ INFERENCE".
- Prefer reversible + observable over irreversible. Pause a tag and publish a documented GTM version (revertible via version history) rather than hard-delete. Do NOT remove a live Google Ads conversion action until you've observed the fix held (All-Conv drops while the Primary holds) over ~1 week — removing it now erases the confirmation signal and is less reversible. Optimize for the next debugger.
- Never place a real purchase. The final end-to-end "one conversion with gclid" proof needs a real/test order — that is a site-owner action, not something you should execute yourself. Everything short of the order is verifiable by us.
- Resolve IDs to names. Name conversion actions, tags, and containers; keep the raw IDs (label,
AW-…,GTM-…) alongside as copy-paste anchors. - Recommend the root-cause option (repo CLAUDE.md global rule): fix the duplicate fire at source, not just the symptom number.
The phase flow (0 → 1 → 2 → 3 → 4 → 4.5 review gate → 5 → 6)
Phase 0 — Intake
Capture: client slug, the tracking question, the blocked plan action (if any), and the platforms (Shopify? GHL? WordPress? custom?). Confirm admin access + the RPD Chrome profile.
Phase 1 — Storefront tag-layer audit (what LOADS)
Load the storefront with a test gclid + utm (?gclid=TEST_AUDIT_123&utm_source=google&utm_medium=cpc&utm_campaign=audit_test). In one JS pass extract:
- Platform (
window.Shopifyetc.); GTM container IDs (gtm.js?id=GTM-…); gtag IDs (AW-…,G-…,GT-…); inlinegtag('config',…)+AW-…/labelrefs. dataLayerevents; gclid capture — does_gcl_aw/_gcl_au/_gcl_lsget written? (proves the conversion linker fires on ad-click).- Other pixels (Meta
_fbp, Klaviyo, TikTok) and whether GHL is present at all. - Flag multiple GA4 streams (more than one
_ga_*cookie = more than one property firing).
Phase 2 — GTM container audit (what is CONFIGURED)
In GTM admin for the container found in Phase 1, enumerate every tag. For each conversion/purchase tag read: Tag type, Conversion ID + Label, Value / Transaction ID (dedup key) / Currency variables, Conversion Linking, and firing trigger. Note duplicates (two purchase tags, two Conversion Linkers) and each tag's age/enhanced-conversions setting.
Phase 3 — Google Ads (+ GA4) conversion-actions audit (what COUNTS)
In Google Ads → Goals → Conversions, enumerate all actions. Per action: source (Website / GA4 import / UA), "Included in account-level goals" (= Primary vs Secondary), count + All-conv, last seen (dead vs live), attribution/window. Identify: the single Primary purchase action (canonical), redundant Secondary purchase actions, and dead/UA actions.
Phase 4 — Diagnose
Map GTM tags ↔ Ads actions by conversion label. Determine:
- Canonical action = the single Primary purchase action, and which GTM tag/pixel feeds it.
- Double-count = 2+ actions counting the same order (usually a GTM tag + a GA4-import, or two GTM purchase tags with different labels). This is the "All-Conv ≫ Conv" gap.
- Dead weight = UA-source or last-seen-months-ago actions; redundant Conversion Linkers; multiple GA4 streams.
- gclid/utm health = confirmed by Phase 1's
_gcl_awwrite + tag Conversion Linking.
Phase 4.5 — Adversarial review (HARD GATE — do NOT skip before any live change)
The diagnosis is a set of claims, and a first-pass reading is routinely wrong (e.g. "the double-count is in the reported Conversions" — often it is only in All-Conversions). Route the diagnosis through /cm-review's lenses before Phase 5 touches anything live — this is the same Stage-4 gate /cm-plan runs, and it is what makes "verify every finding" a mechanical step rather than something the operator has to remember to ask for:
cm-lens-evidence(the critical one here): for every finding, was it verified against the authoritative live source (GTM admin tag config + the Google Ads conversion-actions screen), or is it inferred from the storefront alone / from memory? Default to "assumption until proven in the live system." A finding that can't cite the live screen it came from does not pass.cm-lens-measurement: every proposed fix names a success signal + where it's observed (All-Conv drop, Primary hold, test-order confirmation).cm-lens-ownership: every change has the right owner (RPD-admin vs site-owner test order) and a rollback.cm-lens-brand-client: any number that could reach the client is framed correctly (internal-only until reconciled).
Fold confirmed findings in; unresolved ones become Open questions. No live GTM/Ads change in Phase 5 until the evidence lens has cleared each finding it touches. (Origin: a prior audit run where the operator had to manually request "verify each finding," which is this lens; baking it in as a gate removes that dependency on memory.)
Phase 5 — Package + apply the fix
- Apply now (reversible): pause the duplicate purchase tag(s) in GTM; publish a documented version (name + full description: what/why/reversibility/follow-ups).
- Defer (irreversible / signal-destroying): removal of the live redundant Ads action → hold ~1 week to observe All-Conv drop; dead-action archiving; GA4-stream consolidation; the site-owner test order.
- State exactly what each owner does, and the gate impact on the blocked plan action(s).
Phase 6 — Write the artifact + gate + hand off
Write the <!-- cm:solution --> doc (template below). Render the verified findings + fix inline in chat first (Report-Findings-Before-Output). Present the apply/defer decision via AskUserQuestion (Apply / Document-only), recommending the reversible root-cause fix. Hand off deferred items as tracked tasks.
Artifact template
<!-- cm:solution -->
# CM Solution — <Client> — <Tracking topic> (<blocked gate ref>)
## 1. Problem <the gate + the symptom number>
## 2. Evidence base <one row per VERIFIED fact: storefront / GTM / Ads — source named>
## 3. Root cause <the double-count / dead-weight, + any first-pass correction>
## 4. Fix <applied now (reversible) vs deferred (with why); owner each>
## 5. Measurement plan <signal · where · when — incl. the ~1-week observation + test order>
## 6. Gate impact <how this unblocks each downstream plan action>
## 7. Open questions
## Lens Review Summary
Close — session-wrap offer (R10)
If this session settled a durable marketing decision, produced a CM artifact the user reworked before approving, or surfaced a methodology learning worth carrying forward, offer /cm-session-review as the wrap step: "This session settled something worth capturing — run /cm-session-review to mine the learnings and close the CM loop."
Self-update directive
When a run surfaces a recurring tracking failure mode (Shopify checkout-pixel vs theme-dataLayer double-fire, GA4-import-over-gtag double count, multi-stream GA4, a new platform's capture quirk) or a better verification step, update this file or propose the improvement before finishing.
Appendix — Red Pine reference implementation (optional)
This skill's origin run at Red Pine Digital was a Shopify client account, documented at documents/clients/sprinkler-supply-store/SSS-Analytics-Tracking-Audit-2026-07-02.md. That deployment runs client browser work in an RPD-named Chrome profile (confirming the logged-in Google account first avoids a false "no access"), and the final real/test-order step is owned by the client's own site team.