# Tracking health screen

A stack audit screens tracking; it does not repair it. The screen answers one question per traced workflow: is tracking healthy enough for the decisions people make with it today? Anything deeper goes to the tracking skills, and their verdict comes back here as a finding with its own evidence.

## 1. Ratio against an independent source

- For each workflow, pick the outcome event and the system that commits the fact: purchases against the order or billing system, sign-ups against the user table, leads against accepted CRM records.
- Compute analytics outcomes divided by independent outcomes for each completed day over the last 28 completed days. Compare ratios, not totals.
- When GA4 is the analytics side, end the 28 days 4 days before the audit, the same window as marketing-data-audit: Google updates a daily table through the third day after its date, so the table dated 3 days back can still change on the audit date.
- The baseline band is the ratio's range over comparable days (same weekday type, holidays excluded), not 100%.
- Read it:
  - Inside the band for the whole window: PASS for the screen.
  - At least 2 consecutive completed days outside the band, or 3 of 5: hand off to tracking-break-map.
  - Above 100% for a once-per-outcome event: over-count; hand off to tracking-break-map, and to purchase-reconciliation for purchases.
  - Under about 30 outcomes a day: use weekly buckets before reading a band.
- Never demand equal totals across GA4, ad platforms and billing. Identity, consent, reporting date, attribution and population differ. Explain gaps by category.

## 2. Purchase and lead gaps

- Unexplained gap up to 2% of eligible orders and value: pass. Above 2% and up to 5%: investigate the top suspect. Above 5%: defect. These are the purchase-reconciliation skill's lines; hand off above 5%.
- Duplicate rate above 0.5% of observed orders: defect; hand off to purchase-reconciliation.
- Under 200 eligible orders, extend the period before quoting a rate.
- For leads and sign-ups, keep three counts apart: forms submitted, records accepted by the CRM or user table, and qualified opportunities. An unexplained gap above 5% between analytics and accepted records goes to tracking-break-map, not purchase-reconciliation, which only reconciles orders.

## 3. Page senders by consent state

- In a fresh browser profile with no extensions, load the key pages of each workflow in three consent states: no choice, rejected, accepted. Only the no-choice load needs no click. Clicking reject or accept on a live banner records a consent choice and can send data, so it needs the owner's approval for that test.
- Record every request to Google tag, GTM, Meta, LinkedIn, TikTok, Microsoft and other pixel endpoints, with the IDs they carry.
- Compare the IDs with the register and with retired IDs. A sender that is not in the register is an INV finding; a sender that fires before a choice or after a rejection is a consent finding.
- Hand off to consent-debugging and gtm-container-audit when tags fire before consent, unknown senders appear, or retired IDs still receive data.

## 4. Consent default owner

- Find who sets the consent default: the CMP template, page code before the container, the Google tag's consent override, or the container settings.
- One owner is required. Two owners writing different defaults is a FAIL; hand off to consent-debugging.

## 5. Conversion definitions

- Compare each ad platform's conversion settings with the tracking plan before judging a count: which event, which value, counting one or every, and the attribution window.
- No tracking plan, or events nobody can explain: hand off to tracking-plan-builder for a retrofit.
- Rising Unassigned or (not set) traffic: hand off to ga4-unassigned-not-set-diagnosis.
- A tracking plan exists but events have drifted from it: hand off to tracking-plan-audit.
- The screen fails on more than one workflow, or the senders and consent gates need a full inventory: hand off to tracking-audit. One GA4 property's configuration: ga4-audit.

## What to record back

For every hand-off, add a finding in the audit report with:

- The trigger that sent it (for example "purchase ratio above 100% on 4 of the last 5 completed days").
- The receiving skill and the date.
- Its verdict and evidence, once returned, rated FAIL, WARN or PASS like any other finding.
- Until the verdict returns, the item is listed under not checked with the date it was handed off.

## What not to do

- Do not change tags, filters or channel rules to make numbers match.
- Do not submit forms, place orders or click consent banners on a live site without the owner's approval and agreed test data. Preview modes send real requests to live destinations.
- Do not average a tracking verdict into any score. A broken suppression flow is not offset by healthy page views.