Consent Mode v2 on an EU lead site: trust modeled conversions or the observed count?

Picture a lead gen site with a lot of EU traffic and a consent banner accept rate around 45%. Consent Mode v2 is on basic, so GA4 models the conversions it can't see. The real question is how much of that modeled number you report, and whether you let it drive bidding. Trusting it blindly is risky when the model has little to learn from. Below a few hundred conversions a month, I'd keep the observed count as the headline and show the modeled gap as a range, not a fact. Above that volume the model gets more reliable, but I still wouldn't set a tCPA target off it alone. How do you split the two in your reporting?

I don't think volume is the right threshold either. A few hundred conversions a month tells you little, because the model learns from consented conversions with usable signals, not the total. At a 45% accept rate, roughly half your data is invisible, and that share matters more than the headline count.

The observed number isn't a neutral baseline, either. It moves with the banner. Change the consent copy and the accept rate can drop, so the observed count falls even though nothing in the business changed.

My advice is to track observed and modeled totals side by side over time. When the gap shifts, check consent rate before you conclude performance changed.

The model doesn't just fill in for the users who declined. It learns from consented sessions, so how accurate it is depends on who accepts. If a new banner pulls in a different mix, say more mobile users who tap accept fast, what it learned stops matching the people it's modeling.

I'd log the consent state with every conversion event, using a dataLayer value registered as a custom dimension, and then watch conversion rate among consented sessions. That's a cleaner denominator than the modeled total, and I'd trust it before touching any tCPA setting.

GA4 only applies a custom dimension to events collected after you register it, so you won't get a clean before-and-after on the banner change. Register it before the new banner goes live, or accept that the older period can only be compared at the total level.

Consented-session conversion rate is a cleaner denominator only if the mix of people who accept stays put, and a new banner can change that mix. A clinic that swaps to a mobile-heavy tap-to-accept layout could see consented conversion rate shift even with flat demand, so I'd split by banner version and device before reading the trend.

Agreed, and the timing bites harder than most people expect. New GA4 custom dimensions can take a day or two to show up in reports, so register yours a week before the banner change if you want a usable pre-change baseline.

Iris, the lag matters, but I think it matters less than where the modeled numbers end up. Whether you need that clean baseline comes down to whether Smart Bidding reads the same conversion action you report on.

If you report observed leads but bid on a primary action that leans on modeled conversions, the banner-driven shifts go straight into your tCPA, and a pre-change baseline won't protect you. If the bidding action is a separate observed one, the model mostly shapes your reporting, so the custom dimension becomes a diagnostic tool instead of a safety net.

So before you register anything, check which action is primary for each campaign. Campaigns with only a few dozen observed leads a month are a separate problem, because the bidding model has little to learn from regardless of your consent rate.

It also depends on how the bidding action is collected, not just which one is primary. A conversion tag firing directly in Google Ads gets Consent Mode modeling on its own, so a banner-driven shift can reach tCPA even if your reporting sits on a different event. An action imported from GA4 carries whatever GA4 reports. So don't assume the split is clean until you've opened the conversion settings for each bidding action and checked the source.

I don't think the conversion settings tell you how much of the bidding signal is modeled. For a tag firing directly in Google Ads, the source only tells you where the event comes from. Modeled conversions go into the same count as the observed ones, so a 45% accept rate can shift the number tCPA sees with nothing marking it as modeled.

The more direct check is to compare each bidding action's count against your own lead records for the same days. The gap between the two is the part the model is filling in.