Consent Mode v2 modeled conversions: do you trust them for bidding?

Picture a store with a reject-by-default banner in the EU and opt-in sitting around 35 to 40 percent on mobile. Consent Mode v2 fills the gap with modeled conversions, so GA4 and Google Ads report more conversions than the consented ones alone. The catch is that the model needs enough consented traffic to learn from. On a small site with a few dozen conversions a week, you can't really tell whether the extra numbers reflect real demand or just the model filling in its own assumptions. I'd use the modeled numbers for trend lines, but keep bidding decisions on consented data until you've compared the gap against backend orders for a few weeks. Does anyone here compare modeled and actual order counts regularly, and how far apart do they drift?

Jun's split makes sense. The rule I'd use is to only bid on modeled conversions once their gap to backend orders has held steady for several weeks. A steady offset is easy to account for, but a drifting one quietly pushes your bids in the wrong direction.

The steady-gap test makes sense, but how strict it needs to be depends on how much of what your bidding sees is modeled. Take a store with 35 percent opt-in, where modeled conversions add roughly 40 percent on top of observed ones. That means almost three in ten conversions the algorithm learns from are estimates, so a 10 percent drift in the model shifts your tCPA target by a real amount. Now take a site with 80 percent consent, where modeling only tops up a sliver. The same drift barely moves the bid.

So check the modeled share first. When it's small, bid on reported numbers and just keep an eye on the backend gap. When it's large, wait until the gap holds steady over several weeks before letting automated bidding lean on it. The exact cutoff is a rule of thumb, not a standard, so test it against your own orders.

Checking the modeled share first makes sense, but I'd do it per conversion action, not per account. A store can have a purchase action that's 30 percent modeled while an add-to-cart or lead action runs on almost pure modeling. The account average hides that, and the thinnest action is where drift will hurt your bids most.

You should also confirm the model is running at all. If I remember Google's docs right, modeling needs roughly 700 ad clicks a day from the domain over the trailing week. A small site may be below that, so there's no modeled layer to judge, and any gap you see is coming from somewhere else.

Depends on whether that thin action is a primary goal. If purchase is your only primary conversion and add-to-cart is secondary, a shaky add-to-cart mostly messes up the reporting, not the bids. Once a modeled lead or cart action is primary, it starts steering tCPA, and that's when you want its gap checked.

On the 700 clicks figure, check the current docs rather than trusting anyone's memory, since those thresholds get changed.

Agreed on the primary/secondary split. One gap, though: backend orders can't validate an upstream action like add-to-cart, because nothing server-side records a cart. So you have to test it against something further down the funnel that you can count reliably, like checkout starts or orders, and see whether the modeled cart volume tracks what converts later. GA4 numbers can look internally consistent while drifting away from real buying, which is why this check is so easy to skip. Do you log checkout starts server-side, or are those browser-only too?