Ga4 Audit

GA4 Audit

A one-time audit of a Google Analytics 4 property — configuration and data quality. GA4 collects data the moment the tag fires, but most properties quietly leak trust: unmarked key events, PII in parameters, no Ads link, (not set) traffic. This skill finds those before they corrupt a quarter of reporting.

Requires: Cogny Agent subscription ($9/mo) — Sign up

Works with

Cookieless tools (Plausible, Fathom, Simple) and the store count what GA4 misses, so they turn "data quality" from guesswork into a measured gap.

Prerequisites Check

If no mcp__cogny__* tools exist (no list_integrations, no cogny_status), print this and stop:

Copy code
GA4 Audit needs your marketing data through Cogny's MCP gateway.

1. Sign up at https://cogny.com and connect Google Analytics 4 (Google Tag Manager optional).
2. Add Cogny to .mcp.json in your project:
   {"mcpServers": {"cogny": {"type": "http", "url": "https://app.cogny.com/mcp"}}}
3. Restart Claude Code; it opens a browser for the OAuth login on first use.

Nothing to run without Cogny: GA4's configuration is only readable through its API.

If Cogny is connected but mcp__cogny__mcp__Google_Analytics_4__tool_get_property isn't, say "connect Google Analytics 4 in Cogny" and stop.

Health check: call list_integrations, then tool_list_account_summaries as the cheap read (and GTM tool_list_accounts if GTM is connected). health_status misses revoked tokens, so a 401, Forbidden or scope error is 🔴 "reconnect "; don't retry it. If a tool named here is missing, call refresh_mcp_discovery once; new tools only load after the client reconnects, so use the step's fallback for this run and mark it ⚠ "needs reconnect".

Usage

/ga4-audit — full property audit /ga4-audit config — property + stream + link settings only /ga4-audit events — events, key events, custom dimensions only /ga4-audit data-quality — (not set), PII, anomaly checks only

Steps

1. Property discovery

Copy code
tool_list_account_summaries
tool_get_property        (property_id)

Record: time zone, currency, industry category, and property type.

Data retention: not verifiable via API — ask the user. tool_get_property has no retention field. Ask them to check Admin → Data collection and modification → Data retention, and flag it if it's at the 2-month default (it should be 14 months so year-over-year exploration works). Mark it "unverified" in the report until they confirm.

2. Data streams

Copy code
tool_list_data_streams   (property_id)

Flag:

• Multiple web streams for one site (splits data, double-counts users)
• Missing iOS/Android streams for a business that has apps
• Enhanced measurement settings: not verifiable via API — ask the user (the stream response has no enhanced-measurement data). Infer what you can from the event list in step 3 (scroll, file_download, view_search_results, video_start, form_start present or absent), and label it as inferred.

3. Key events (conversions)

Copy code
tool_list_conversion_events   (property_id)
tool_run_report   (property_id,
  date_ranges: [{"startDate": "30daysAgo", "endDate": "yesterday"}],
  dimensions: ["eventName"], metrics: ["eventCount", "keyEvents"])

The report shows which events actually fire and which key events recorded anything. Step 2's enhanced-measurement inference reads from it.

Flag:

• No key events configured — Ads bidding and GA4 reporting both lose their anchor
• Key events marked on noisy automatic events (page_view, scroll) — inflates the conversion count and misleads everyone
• Obvious business goals not marked as key events (e.g. generate_lead, purchase, sign_up)
• More than ~10–15 key events — when everything is a conversion, nothing is

Fix (write, asks first, if available): tool_mark_key_event (property_id, event_name, counting_method: ONCE_PER_EVENT or ONCE_PER_SESSION) for a real goal that isn't marked, tool_unmark_key_event (property_id, event_name) for a noisy one. Takes effect for new data only. Unmarking also removes it from any Ads import; say so before asking. These tools exist in the proxy but aren't exposed on every workspace. If they don't load, don't refresh in a loop: tell the user to mark/unmark the event in the GA4 UI (Admin → Events → toggle "Mark as key event") and give them the exact event name, case-sensitive.

4. Custom dimensions & metrics

Copy code
tool_list_custom_dimensions   (property_id)
tool_list_custom_metrics      (property_id)
tool_get_metadata             (property_id)

tool_get_metadata returns ~121k characters (every standard field). Don't read it whole: when the result is saved to a file, filter it with jq, e.g. jq -r '.dimensions[] | select(.apiName | startswith("customEvent:") or startswith("customUser:")) | .apiName'.

Flag:

• Event parameters collected by the tag but never registered as custom dimensions (the data is invisible in reports). With GTM connected, tool_list_tags on the live workspace and collect the event parameter names the GA4 event tags send; any parameter with no matching customEvent:<name> dimension is unregistered. Without GTM, ask the user for the tracking plan and say the check is partial
• Approaching the 50 event-scoped / 25 user-scoped dimension limits
• High-cardinality registered dimensions (timestamps, IDs) — they cause (other) rows

5. Traffic & data-quality report

Copy code
tool_run_report   (property_id, last 30 days,
  dimensions: sessionDefaultChannelGroup, sessionSource, sessionMedium,
  metrics: sessions, keyEvents, totalUsers)

Use keyEvents, not the deprecated conversions metric. If run_report times out with order_bys, retry once without them and sort the rows yourself.

Flag:

• A large Unassigned or (not set) share — broken UTMs or tagging gaps (point the user to /utm-builder)
• A Direct share above ~40% — usually lost referrer data or missing tags
• Self-referrals (your own domain as a source) — needs an unwanted-referral filter
• Login and payment referrers (accounts.google.com, checkout.stripe.com, paypal.com, OAuth providers such as login.microsoftonline.com, github.com/login, appleid.apple.com) — they steal credit for the conversion from the real source; add them to the unwanted-referral list
• Internal traffic visible in reports — no internal-traffic filter applied

5b. Cross-check against other sources (enhanced when connected)

Call list_integrations once. For each healthy source, compare the same 30 days:

Report each source's number next to GA4's, named. Don't expect a match: cookieless tools count visitors differently. A gap you can measure goes in the Data quality score; deep reconciliation belongs in /conversion-debug.

6. PII & policy check

Scan event and parameter names from get_metadata for anything that could carry personal data — email, phone, user_email, full names, raw addresses in parameters or in the page path / query string.

Then query the URLs themselves; names alone miss PII in query strings. This found real email addresses and magic-link tokens in a live run:

Copy code
tool_run_report   (property_id, last 90 days,
  dimensions: ["pageLocation"], metrics: ["screenPageViews"], limit: 50,
  dimension_filter: {"orGroup": {"expressions": [
    {"filter": {"fieldName": "pageLocation", "stringFilter": {"matchType": "CONTAINS", "value": "email="}}},
    {"filter": {"fieldName": "pageLocation", "stringFilter": {"matchType": "CONTAINS", "value": "%40"}}},
    {"filter": {"fieldName": "pageLocation", "stringFilter": {"matchType": "CONTAINS", "value": "@"}}},
    {"filter": {"fieldName": "pageLocation", "stringFilter": {"matchType": "CONTAINS", "value": "token"}}}]}})

Any row is a finding. Redact the values when you quote them (show the path and parameter name, never the address or token).

🔴 PII in GA4 violates Google's policy and risks the property being deleted. Flag it as critical and name the exact parameter.

7. Integrations

Copy code
tool_list_google_ads_links   (property_id)

Flag:

• No Google Ads link — conversions can't flow to Ads for bidding (an empty result means no link on this property; confirm with the user if they import via another property)
• BigQuery export and Search Console link: not verifiable via API — ask the user (the API exposes neither link). Ask them to check Admin → Product links. Recommend BigQuery export for any property that wants raw-data analysis (/ga4-bigquery-schema covers the schema). Mark both "unverified" in the report until confirmed

8. Score and report

Copy code
GA4 Audit — [Property Name]
GA4 Health Score: X/100

Score breakdown:
  Configuration ...... X/20   (streams, retention, time zone)
  Key events ......... X/25   (the reporting + bidding anchor)
  Custom dimensions .. X/15
  Data quality ....... X/25   ((not set), Direct, self-referrals)
  Compliance ......... X/15   (PII, internal traffic)

🔴 Critical   — PII collected, no key events, retention at 2 months
🟡 Important  — unregistered dimensions, high (not set), no Ads link
🟢 Optimization — BigQuery export, enhanced measurement tuning

Top 3 Actions:
1. [Highest impact]
2. ...
3. ...

9. Record findings

jsonCopy code
{
  "title": "Email address captured in GA4 'page_location' — policy violation",
  "body": "The checkout-success URL includes ?email=… and GA4 stores it in page_location and page_path. Google policy prohibits PII in Analytics and can delete the property. Strip the parameter at the tag, or redact it before the event fires.",
  "action_type": "data_quality",
  "expected_outcome": "PII removed from collection; property compliant",
  "estimated_impact_usd": 0,
  "priority": "critical"
}

On Cloud, create_ticket takes the same content with the fields mapped per section 2 of the routing guide.

Action types: tracking_configuration, conversion_setup, data_quality, custom_dimension, integration, compliance.

Recording findings

Use create_ticket when the workspace has it (Cloud) and create_finding otherwise (Solo). The fields differ; see section 2 of ../marketing-engineer/references/mcp-routing.md.

Critical rules

1. PII is always critical. Flag it loudly and name the parameter.
2. Key events are the anchor — for GA4 reporting and for Ads bidding. Weight them.
3. Quote real numbers: (not set) %, dimension counts. Settings the API can't read (retention, enhanced measurement, BigQuery and Search Console links) are "unverified", never guessed.
4. Read-only by default. The only writes are tool_mark_key_event / tool_unmark_key_event, only when the workspace exposes them, and only after the user says yes to the exact event. Otherwise point to the GA4 UI.
5. Name the source on every cross-check number. Never average GA4 with Plausible or the store.