Since March 2024, advertising to new users in the European Economic Area without sending Google four consent signals has meant no personalization, no remarketing audiences and degraded measurement. That was the Consent Mode v2 mandate, pushed by the Digital Markets Act, and it added two parameters to the consent state: ad_user_data and ad_personalization. More than two years later, every EU account I touch reports a blend of observed and modeled conversions, and most people reading those dashboards cannot tell which is which.
I implement Consent Mode v2 through server-side GTM for clients and I have watched the modeled share of reported conversions become material. This is about what those modeled numbers actually are, and how I keep them honest.
Basic and advanced mode are different products
Basic consent mode blocks Google tags entirely until the user consents. Decline means silence: no ping, no cookie, no event of any kind. Google then models your lost conversions from a general picture of how consenting and non-consenting users behave, because your account gives it nothing to learn from.
Advanced mode loads the tags regardless, and on decline sends cookieless pings: no identifiers, but evidence that a visit happened, that a conversion fired, that an ad click preceded it. The modeling in advanced mode is trained on your own traffic patterns instead of a category average. If you are going to live with modeled data anyway, advanced mode gives the model something real to work with. It also carries more legal nuance than basic, so it is a conversation with your privacy counsel, not a unilateral tagging decision.
If you do not know which mode your site runs, load it in a clean browser with the network tab open and decline everything. Requests to Google’s ad and analytics endpoints after the decline, stripped of identifiers, mean advanced mode. Total silence means basic. I have audited sites whose owners were certain of one answer and wrong.
What conversion modeling claims to recover
Google’s published figure, from when consent-based conversion modeling launched in Google Ads, is that modeling recovers more than 70% of ad-click-to-conversion journeys lost to consent choices. Handle that number with gloves: it dates to 2021, before the v2 enforcement wave, it is an average across advertisers, and it comes from the party selling the ads.
Google says conversion modeling through consent mode recovers, on average, more than 70% of the ad-click-to-conversion journeys lost to consent choices. The claim is plausible in aggregate and unverifiable for your specific account, which is exactly why you need your own check.
The mechanics matter more than the headline. Google’s documentation notes that consented users typically convert at 2 to 5 times the rate of unconsented users. People who click accept-all are simply more engaged. A naive model that projected consented behavior onto everyone who declined would overcount badly, and correcting for that selection effect is most of the modeling work. The same help page walks through an advertiser with a 50% consent rate ending up with a 19% drop in reported conversions after modeling uplift, rather than the 50% you might fear.
The consent rates feeding the models
How much of your data is modeled depends on how many people decline, and the European picture is better than the horror stories suggest. Didomi’s benchmark of consent interactions across European markets in 2025 puts consent rates between 75.1% in Western Europe and 89.3% in Eastern Europe, with France at 71%. Count the visitors who never touch the banner at all and the effective opt-in rate drops to between 55.7% and 67.6% depending on region. Industry matters too: in the same dataset, media and publishing sites lead at 82.7% while energy and utilities trail at 69.6%.
Run the arithmetic on your own market and you will usually find that a third or more of your EU conversion journeys ride on the model. That is too large a share to accept on faith.
How I audit modeled conversions
The only sanity check that has held up for me is a server-side source of truth. Backend orders, CRM-qualified leads, reconciled revenue: pick the hardest number your business produces and compare it to platform-reported conversions every month, split by market. In BigQuery the whole check is one scheduled query, daily platform conversions joined to timestamped backend orders on date and market. It takes an afternoon to build and it has caught more measurement drift than any monitoring product I have paid for.
Three patterns to watch. If the ratio of platform conversions to backend truth is stable, the modeling is at least consistently wrong, which is workable for bidding. If the ratio jumps after a banner redesign, a CMP migration or a tagging change, the model is reacting to your instrumentation rather than your customers. And if modeled share grows while backend revenue stays flat, treat any platform-reported CPA improvement as fiction until proven otherwise.
A client of mine redesigned their consent banner last year and platform conversions rose within two weeks while backend orders did not move. The uplift was the model relearning a new consent mix, nothing more. Had I taken the dashboard at face value, I would have credited a creative refresh that shipped the same week and doubled down on it.
Modeled conversions are a forecast presented as a measurement. I use them, I bid on them, and I trust them exactly as far as my backend numbers allow. Two years into Consent Mode v2, that discipline is the whole game.