Anatomy of a GA4 audit

Symptom

Reported purchases run consistently higher than the order platform, and the gap grows on days with more traffic.

Diagnosis

GTM Preview and DebugView show two purchase events per checkout: one from the GTM container's GA4 event tag, one from the platform's built-in GA4 integration that nobody remembered was still switched on. A BigQuery count of purchase events per transaction_id confirms it — the same order IDs appear twice. GA4's transaction_id deduplication is best-effort, and it does not reliably reconcile hits arriving from two independent sources.

Impact — in this example

In this composite example: 1,000 real orders produce roughly 1,800 reported purchases once the second pipeline and confirmation-page reloads stack up — revenue reads nearly double, and every ROAS decision made on it inherits the error.

Symptom

EU conversion counts look implausibly weak versus non-EU, and Google Ads reports fewer conversions than GA4 for the same campaigns.

Diagnosis

The consent default snippet lives in a tag fired on Page View — after the Google tag has already sent its first hit. On every new EEA visit, the first pageview (and any early conversion) leaves without consent state. Only two of the four v2 signals are set, so ad_user_data and ad_personalization are missing entirely, degrading Google Ads measurement for EEA traffic.

Impact — in this example

In this example, with roughly 40% of sessions from the EEA, a meaningful slice of paid-traffic conversions never enters measurement at all — the channel looks weaker than it is, and budget quietly shifts away from what's actually working.

Symptom

Every deeper question — deduplication, reconciliation, cross-device behavior — dead-ends at the GA4 UI's sampling and thresholding.

Diagnosis

The property has never had its BigQuery export enabled. There is no raw, unsampled record of events, so claims like 'revenue is double-counted' can be suspected but not proven, and historical data before the export is switched on is gone for good.

Impact — in this example

In this example the audit itself is constrained: without raw events, quantifying the duplicate-purchase rate requires manual test orders instead of a five-minute query. Every future analysis inherits the same ceiling until the export exists.

Symptom

A payment provider's domain ranks among the top 'referral' sources, and paid campaigns under-report conversions they clearly drove.

Diagnosis

Checkout redirects through the payment gateway; because the gateway's domain was never added to GA4's unwanted-referrals list, the return registers as a new referral traffic source — and last-non-direct-click attribution hands the purchase's credit to the gateway instead of the campaign that actually won the customer. (If the payment step outlasts the 30-minute session window, the purchase can also land in a fresh gateway-attributed session.)

Impact — in this example

In this example the distortion is structural rather than a percentage: the channel report shows a payment processor as a top revenue 'source' while the campaigns that actually won the customer are marked weaker — precisely backwards for budgeting.

Symptom

Finance reconciles against Stripe; marketing reports from GA4 — and the two never agree, so neither team trusts the other's number.

Diagnosis

After the duplicates are fixed, a residual gap remains: client-side tracking loss from Safari's ITP, ad blockers, and consent declines means GA4 will always see fewer purchases than the payment processor. That part is physics, not misconfiguration — what's missing is that nobody has measured the gap, explained it, or agreed on a tolerance.

Impact — in this example

In this example the gap itself is normal; the cost is organizational — hours lost re-litigating 'whose number is right' in every reporting cycle, because the divergence was never documented.