# How to audit conversion tracking with an AI agent

> Tier: deep. Full contract: quantified decision rules and evidence requirements.

Whether the numbers can be trusted: GA4 measurement health, platform pixels, key events, and where platform-reported and site-reported conversions diverge.

**Intent.** Establish whether measurement is sound before anyone optimises against it.

## When to use

- Before any performance judgement on a new account.
- After a site migration, consent banner change or tag rework.
- When two systems report conversions that cannot both be true.

## When not to use

- As a substitute for a tag-manager debug session: this audits what APIs expose, and says so where that boundary bites.

## Sources

Required: Google Analytics 4
Optional: Google Ads, Meta Ads, TikTok Ads, Pinterest Ads

## Preconditions

- GA4 property ticked and readable.
- A window long enough to compare: 28 days by default.

## Inputs

- `window`: Comparison window. Default: last 28 days.

## Procedure

1. Read GA4's own diagnostics: ecommerce, server-side tagging, audience health. — tools: `ga4_get_ecommerce_diagnostics`, `ga4_get_server_side_tagging_diagnostics`, `ga4_get_audience_diagnostics`
2. List key events and check each fires at plausible volume over the window. — tools: `ga4_get_key_events`, `ga4_run_report`
3. Check event parameters: declared vs actually populated. — tools: `ga4_get_event_parameters`, `ga4_get_custom_definitions`
4. On each connected ad platform, read conversion wiring and pixel presence. — tools: `google_ads_get_conversion_actions`, `meta_get_pixels`, `tiktok_get_pixels`, `pinterest_get_conversion_setup`
5. Compare platform-reported conversions to GA4 channel figures over the same window. — tools: `ga4_run_report`

## Decision rules

- **When** A key event used for bidding fired zero times in 7 days on a site with traffic → **conclude** Broken wiring, highest severity: platforms are optimising toward silence.
- **When** Ecommerce diagnostics report items missing on purchase events → **conclude** Revenue by product is unreliable; say which reports downstream inherit the gap.
- **When** Platform and GA4 conversions diverge more than 25% with consent mode active → **conclude** Attribute the gap to consent and modelling before blaming either system; the audit names the mechanism, not a culprit.
- **When** Server-side tagging diagnostics report a failing transport → **conclude** Everything measured after that hop is suspect; date the failure and bound the damage.

## Evidence requirements

- Every gap is quantified: both figures, window, and the mechanism suspected.
- Diagnostics findings quote the API's own diagnostic, not an inference from symptoms.

## Human approval boundary

- Tag and consent fixes happen in tag managers and CMPs, outside the gateway: the audit delivers the work order.
- The one thing never proposed is optimising against numbers the audit just flagged.

## Failure modes

- GA4 thresholding hides low-volume rows; absence in a report is not absence of events.
- Intraday GA4 data is incomplete; never audit today.
- Consent-gated tags fire differently by region; a country mix shift can look like a tracking break.

## Output

- **Verdict**: Trustworthy, trustworthy-with-caveats, or not trustworthy, per system.
- **Gaps, ranked**: Each with both figures, mechanism, and who inherits the error.
- **Work order**: Fixes, where they live, who can do them.

## Related

- Skill: [account-audit](https://www.getmcpads.com/skills/account-audit.md)
- Skill: [cross-platform-performance](https://www.getmcpads.com/skills/cross-platform-performance.md)
- Guide: [ga4-analysis](https://www.getmcpads.com/guides/ga4-analysis.md)