# Google Ads Measurement

> >

- **Type:** Skill
- **Install:** `agentstack add skill-chanktb-claude-google-ads-measurement`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [chanktb](https://agentstack.voostack.com/s/chanktb)
- **Installs:** 0
- **Category:** [Search](https://agentstack.voostack.com/c/search)
- **Latest version:** 0.1.0
- **License:** MIT
- **Upstream author:** [chanktb](https://github.com/chanktb)
- **Source:** https://github.com/chanktb/claude-google-ads/tree/main/skills/measurement
- **Website:** https://khuetran.com

## Install

```sh
agentstack add skill-chanktb-claude-google-ads-measurement
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# Google Ads — Measurement (tracking gate)

Prove conversion tracking is correct and trustworthy, or block `plan`/`builder-*` until it is. Reuse the
live MCP detection from `setup`. **Never raise an alarm from the action list alone.**

## STEP 0 — Load context
**If `account-context.yaml` is missing, run `setup` first — don't validate tracking on an unconfigured account.**
Read `account-context.yaml`: `google_ads.customer_id`, `conversion_actions`, `measurement.*`,
`data_source` (for value sanity), `business.model/vertical` (lead vs purchase expectations). Output goes
to the working directory; update the context's `conversion_actions` + `measurement.*` as you verify them.

## Model dispatch (run cheap, decide expensive) — see `${CLAUDE_PLUGIN_ROOT}/references/model-tier-dispatch.md`
- **Scout (`haiku`)** — STEP 0 context read; the conversion-action list pull.
- **Routine (`sonnet`)** — STEP 1 per-campaign conversion segmentation (`segments.conversion_action_name`), STEP 3 Ads/GA4/store value pulls. Dispatch as `general-purpose` sub-agents; **return raw, don't conclude**.
- **Judge (main session)** — the WHOLE-picture verdict, every is-it-a-problem call, the double-count judgment, the PASS/WARN/FAIL that gates plan/builder. **Never let a cheap tier raise the alarm** — that holistic judgment is the entire point of this skill.

## STEP 1 — Build the WHOLE picture before concluding (mandatory)
A conversion list read in isolation lies. The same flag is fine or fatal depending on context. Do all four:

1. **Actions** — enabled conversion actions: name, category, primary_for_goal, counting_type,
   value_settings.default_value (value tracking on/off), source. (Omit `conversion_action.type` if the MCP
   rejects it; see ${CLAUDE_PLUGIN_ROOT}/skills/audit/references/gaql-notes.md.) **Decode the enums:**
   `category` 4=PURCHASE · 6=LEAD · 5=SIGNUP · 3=PAGE_VIEW · 8=ADD_TO_CART · 9=BEGIN_CHECKOUT ·
   11=PHONE_CALL_LEAD · 13=SUBMIT_LEAD_FORM · 16=GET_DIRECTIONS · 18=CONTACT · 19=ENGAGEMENT · 20=STORE_VISIT;
   `counting_type` 2=ONE (one per click) · 3=EVERY (many per click). For ecommerce, the working primary
   should be category=PURCHASE with EVERY counting + value tracking on.
2. **Usage** — which actions each ACTIVE campaign actually fires/optimizes for: segment conversions by
   `segments.conversion_action_name` per campaign over 30-90 days (explicit `YYYY-MM-DD` dates). An action
   no campaign fires is inert.
3. **Source & double-count** — map each value-carrying action to a source
   (store channel, e.g. Shopify Google&YouTube "App Purchase" / WooCommerce / GA4 import / gtag /
   Merchant Center). Two value
   purchases both PRIMARY and both firing = double count. One primary + one secondary = correct.
   See `${CLAUDE_PLUGIN_ROOT}/references/conversion-tracking-logic.md`.
4. **Verdict** — only now, per the rules below.

## What is / isn't a problem
- ✅ Campaigns optimize toward the real working purchase action → healthy, even if other purchase actions
  exist as secondary. A second purchase action kept secondary (e.g. GA4 web purchase alongside the channel
  App Purchase) is the CORRECT anti-double-count setup, not a bug.
- ✅ "Junk" actions (menu views, directions, calls, store visits, app micro-events) flagged primary but
  fired by NO active campaign → harmless. Note for awareness; do not alarm.
- ✅ Cross-brand action sitting secondary and unused → harmless; note it.
- 🔴 Two value-carrying purchase actions BOTH primary and BOTH firing in the same campaigns → real
  double-count. Flag.
- 🔴 The action campaigns optimize for is a low-value engagement action, not the purchase action → flag.
- 🔴 No working purchase/lead conversion at all, or value tracking off for ecommerce → flag.

## STEP 2 — Other checks
- **Counting type**: `one` for leads (avoid form-spam inflation); `every` for ecommerce purchases.
- **Value tracking**: on for ecommerce; sane currency.
- **Enhanced Conversions / Consent Mode v2 / server-side (CAPI)**: the API/MCP usually CANNOT confirm
  these — mark them **verify-in-UI**, not FAIL. State the ceiling if confirmed.
- **Conversion lag**: are conversions still trickling in (long lag windows)? Note if attribution looks cut.

## STEP 3 — Value sanity cross-check
When more than one source is available, compare and report the gap (don't silently pick one):
- **Google Ads** attributed value/conversions for the primary purchase action (30-90d).
- **GA4** purchase value for the same window (via GA4 MCP / API / a connected GA4 client).
- **Store** net sales (Online Store channel only — exclude draft/POS/TikTok/marketplace;
  see audit/references or aov-and-sales-sourcing.md).
A large divergence (>35% Ads vs GA4) is itself a finding (attribution, tagging, or channel-mix issue).

## Data sources & fallback (graceful, but NO fabrication)
Read the context `connections` block. GA4/store via MCP → API creds → a connected client/script. Missing
GA4/store does NOT block measurement — you can still judge the Google-Ads-side setup. **But report only the
Ads number and mark the value cross-check `UNVERIFIED — connect GA4/store`; NEVER estimate the missing
side to "complete" the comparison** (a fabricated GA4/store figure is worse than an honest gap). If the
source is merely unwired, prefer guiding the user to connect it over skipping.

## Output — verdict + notes (to the working dir)
- **Notes for awareness** — the harmless-but-worth-knowing items (inert primaries, unused cross-brand
  actions). Frame as notes, never alarms.
- **Verdict**: PASS / WARN / FAIL with specific fixes. Only genuine 🔴 problems gate downstream.
  **FAIL blocks `plan` and `builder-*`.**
- Update `account-context.yaml`: refresh `conversion_actions` (with the verified primary + source) and
  `measurement.*` flags (enhanced_conversions, server_side, ga4.enabled).

## To build / refine later
- [ ] A runnable helper to pull the per-campaign conversion segmentation + format the verdict.
- [ ] Reuse the shared HTML report module (see DECISIONS) when it exists.

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [chanktb](https://github.com/chanktb)
- **Source:** [chanktb/claude-google-ads](https://github.com/chanktb/claude-google-ads)
- **License:** MIT
- **Homepage:** https://khuetran.com

Install and usage instructions live in the source repository linked above.

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-chanktb-claude-google-ads-measurement
- Seller: https://agentstack.voostack.com/s/chanktb
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.