What Actually Broke When We Shipped Consent Mode v2

We rolled out Consent Mode v2 for a client ahead of the EEA enforcement deadline, tested it thoroughly in staging across three consent states, and watched Google Ads conversion volume drop by nearly 40% within six hours of the production push. Nothing in staging predicted this, because staging never had real cookie banner interaction data or real ad click IDs flowing through it.

The actual failure

The consent management platform's default state — before a user interacted with the banner at all — was set to deny for both ad_storage and analytics_storage, which is correct and expected behavior. The problem was that the CMP's consent banner itself had a render delay of almost two seconds on mobile due to an unrelated script loading race condition, and GA4 and Google Ads tags were firing in that two-second window on every single page load, correctly in denied mode, but modeled conversions depend on a reasonable volume of consented signal to calibrate against. With the banner rendering late and bounce rates high on mobile, a much larger share of sessions than expected never reached a consent decision at all before the user left.

Consent Mode v2 doesn't fail when people say no. It fails quietly when people never get the chance to say anything, and that's a UX bug wearing a compliance costume.

What we changed

• Fixed the actual root cause first: the script loading race condition delaying the banner render, unrelated to consent logic itself.
• Added a hard timeout fallback so the banner always renders within 300ms even if a dependency is slow to load.
• Instrumented consent_state as its own GA4 event parameter so we could see, in DebugView and BigQuery, exactly how long users spent in an undecided state.
• Set client expectations explicitly that some volume drop from modeled conversions is expected and not itself a bug — the goal was separating that expected drop from this unrelated regression.

The lesson that stuck with me: consent mode rollouts fail at the UX layer far more often than at the tagging layer. The GTM configuration was correct from day one. The thing that actually cost three weeks of clean data was a banner render delay nobody thought to test for, because it had nothing to do with consent and everything to do with how fast an unrelated script loaded on a slow mobile connection.