How to audit conversion tracking with an AI agent
Whether the numbers can be trusted: GA4 measurement health, platform pixels, key events, and where platform-reported and site-reported conversions diverge.
IntentEstablish whether measurement is sound before anyone optimises against it.
- 5
- Steps
- 11
- Tools
- 4
- Rules
- 3
- Failure modes
When to use it
Use it when
- 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.
Do not use it when
- As a substitute for a tag-manager debug session: this audits what APIs expose, and says so where that boundary bites.
What it needs first
Required
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
In order. Every tool named is one the gateway ships, and links to its reference.
Read GA4's own diagnostics: ecommerce, server-side tagging, audience health.
List key events and check each fires at plausible volume over the window.
Check event parameters: declared vs actually populated.
On each connected ad platform, read conversion wiring and pixel presence.
Compare platform-reported conversions to GA4 channel figures over the same window.
Decision rules
The observable condition, what it lets you conclude, and what takes the conclusion back.
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 every conclusion must carry
- Every gap is quantified: both figures, window, and the mechanism suspected.
- Diagnostics findings quote the API's own diagnostic, not an inference from symptoms.
Where the agent stops
- 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.
How it goes wrong quietly
The cases where the analysis is wrong and still looks right. Read them before trusting a number.
- 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.
What the answer contains
- 01Verdict
- Trustworthy, trustworthy-with-caveats, or not trustworthy, per system.
- 02Gaps, ranked
- Each with both figures, mechanism, and who inherits the error.
- 03Work order
- Fixes, where they live, who can do them.