Consent Mode v2 modeled conversions: bid on them, or report only the consented count?

A cookie banner can hide a big share of EU conversions, and Consent Mode v2 is meant to win some of that back through modeling. The catch is that the model only learns from users who accepted. If the decline rate is high and skewed toward mobile or one country, the modeled numbers can drift a long way from what actually happened. Bidding on them is probably fine as long as you track the ratio of modeled to consented conversions every week. If that ratio swings a lot, the model is guessing, not measuring. What decline rate would you say is too high to trust the model at all?

Not sure there's a clean cutoff. What I'd check is whether the declines are spread evenly or piled up on one device or region. A random 40% decline is easier to model than 15% concentrated on iPhone Safari. For a local plumber with traffic from one city, that split often matters more than the overall rate.

I don't buy that a random 40% decline is easier to model than 15% piled onto iPhone Safari. The model doesn't care much whether declines are spread evenly. What matters is how many consented conversions it has to learn from in each segment. If that plumber gets 300 consented conversions a month from iPhone Safari, the model has plenty to work with, even if a lot of those users declined. If a random 40% leaves only 20 consented conversions a month across all devices, the model is thin everywhere, and the even spread doesn't rescue it.

So look at consented conversion counts per device and region before you look at the percentages. Concentration only becomes a real problem when the segment you're bidding on has too few consented conversions to anchor the estimate. What you want to know is how many real conversions the model is extrapolating from in each bucket.

Agreed on the counts, but I'd go one level lower and count consented ad clicks, not just conversions. From what I understand, the modeling starts from clicks that carry a gclid and were consented, then estimates conversions for the clicks that declined. So a plumber with 300 consented conversions could still have only a handful of consented clicks from iPhone Safari if the conversion rate there is low, and that click number is what anchors the estimate. I'm not sure what Google's current minimum thresholds are, so check the docs before assuming a small segment is being modeled at all.