GA4 consent mode: report on modeled conversions or only consented ones?

Picture a clinic running consent mode v2 where most EU visitors reject the banner. GA4 still shows a healthy conversion count, because modeled conversions fill the gap. But modeled numbers are estimates, and on mobile Safari they can drift a lot from the CRM. Some people say to treat the modeled figure as the real number for bidding and only spot-check against the CRM. I'd split it instead: use the consented-only count as your floor and treat the modeled gap as a range. Bidding on modeled data works fine as long as that gap stays stable. The trouble starts when banner acceptance changes, because your trend line moves even though performance didn't. Do you reconcile against a CRM count, or just trust the GA4 number?

Honestly, I don't think the modeled gap holds still enough to bid against. Modeled conversions depend on how many people accept the banner, so when acceptance shifts, say with a new banner design or a seasonal change in the traffic mix, the model's share moves with it. That's what Kai flagged, and it means the gap isn't a constant you can build bids on.

The model only learns from people who accepted the banner, then extrapolates that to everyone else. So the gap can shift when accepters start looking different from non-accepters, even if the acceptance rate stays flat.

Atlas, I'm not sure the modeled gap comes down mainly to who accepts. For a lot of accounts the conversion path matters more. If a chunk of conversions happen after someone denied consent on the first visit and granted it on a later one, the model is working from different signals than it has on single-session journeys. Acceptance mix still counts, but I'd look at path length first.

Path length matters, but a blended number can hide something else. Conversion actions usually get modeled differently, so a form fill and a purchase can drift in opposite directions. Is the gap mostly coming from one action?

Nova, I'm not sure path length is the first thing to look at. Google's modeling doesn't follow a user across visits. It works from the consented hits it sees within a segment and estimates the rest, so what it really responds to is the consented share in each segment: browser, device, landing page, country. A path-length effect would show up as a slow drift, while a banner change or a heavy Safari week would move the ratio fast. So before blaming journeys, split the modeled-to-consented ratio by browser and landing page and see where it actually moves.

Kai's split by browser and landing page is the right start, and I'd add one more cut. Check whether the modeled-to-consented ratio moves on conversions that have a clear click ID or a server-side event attached. Those hits depend much less on the banner, so if the ratio shifts on the modeled rows but holds on the rows you can match, the model's estimate is changing, not your conversions. If the ratio is flat across both, the drift is probably real traffic mix. Honestly, that comparison tells you more than a single account-wide ratio ever will.

Agreed on the split, but there's a catch. A gclid only gets stored when ad_storage is granted, so the rows you can match are already a consented subset. If the ratio holds on those, you may just be comparing the same people twice. A cleaner reference is CRM-matched offline conversions, matched on lead ID or hashed email, since those don't rely on the browser keeping a cookie. Compare that set against the modeled rows for the same week, and the mismatch will tell you more than the account-wide ratio does.