What consent modelling is, and when Google actually switches it on

All posts

Consent modelling is Google estimating what it could not observe because a visitor declined. In Google Analytics it is behavioural modelling, and it starts only once the property runs advanced consent mode and clears Google's published bar: 1,000 denied events a day for seven days, plus 1,000 daily consented users on seven of the last twenty eight. Google Ads models separately.

The word covers two different systems

Almost every argument about modelling is two systems discussed as one. They live in different products, estimate different things, and carry separate eligibility bars that Google publishes in separate places.

Behavioural modelling is the Analytics one. When a visitor declines analytics storage, GA4 cannot identify them across pages, so their journey would normally be absent. It fills that space using patterns learned from consented visitors, producing an estimate of users, sessions and conversions never directly observed.

Conversion modelling is the Ads one. Its problem is narrower: an ad click happened, a conversion happened, and without storage the two cannot be joined. It estimates that join, and the result arrives blended into your ordinary conversion figures.

A site can qualify for one and miss the other. If that is the shape of your problem, why Google Ads and GA4 report different conversions goes through the rest of it.

The thresholds, stated plainly

For Analytics, Google lists three conditions. Consent mode has to be enabled across all pages, implemented so that tags load before the consent dialog appears and load in all cases rather than only after consent, which is the advanced implementation. The property has to collect at least 1,000 events a day with analytics_storage set to denied, for at least seven days. And it has to have at least 1,000 daily users sending events with analytics_storage granted, on at least seven of the previous twenty eight days. Google adds that it may take more than seven days at that volume for the model to train.

For Ads, there is one published number: 700 ad clicks over a seven day period, per country and domain grouping. That last clause does more work than it looks: a business spread across five countries is measured on each country separately, so a respectable overall click volume can still fail everywhere at once.

Read the two bars together and the conclusion is unwelcome. A thousand declining visitors a day is a substantial site. This is not a formality you pass on the way to a working setup, and if you run advanced consent mode because modelling was the promised payoff, it is worth knowing whether that payoff is available to you.

How to tell whether it is actually on

1. Check the reporting identity in Analytics

   In the GA4 admin panel, open Reporting identity. Modelled data reaches your reports only under Blended, which is the default; the Observed and Device based options use observed data alone. Selecting Blended does not by itself qualify a property, so treat it as a precondition rather than as proof.

2. Confirm you are actually in advanced mode

   Analytics eligibility depends on it, so prove it rather than assume it. A denied visitor should still generate a request; if declining produces silence you are in basic mode and no volume of traffic will qualify the property for behavioural modelling. Ads conversion modelling is the exception here, since it does not carry the same advanced requirement. Ours is the check for which one you are running.

3. Read what the signals actually say on the wire

   Eligibility is counted from the denied pings, so it matters that they carry the state you think. The parameter that carries the two signals Consent Mode v2 added is covered in what the gcd parameter means.

4. Expect Analytics and BigQuery to disagree

   Modelled data reaches the reporting interface and the API, and it is not included in the BigQuery export. Both numbers are correct for what they measure. Anybody reconciling one against the other without knowing this will conclude something is broken when nothing is.

What to do when you are below the bar

This is where the published answers stop, and it is the more useful half of the question. Three things are worth doing, and none of them is waiting.

Run advanced mode anyway. Its value does not begin at the modelling threshold: the cookieless signal from a declined visit is what keeps the visit represented at all, and those denied pings are the very thing counted against the 1,000 a day bar, so sending them is how a property becomes eligible in the first place. Choosing basic mode because modelling is out of reach removes the only trace you were still getting.

Then attack the acceptance rate rather than the estimate. Modelling only ever approximates the hole; the acceptance rate decides how big the hole is. Across Amplio Data client implementations, consent hides around 34% of sessions, which is a measured range and not a guarantee, and moving that number is the one intervention that changes the underlying measurement rather than the reconstruction of it. Doing so without dark patterns is its own piece of work.

Velo runs advanced consent mode by default, so the denied pings that eligibility depends on are sent from day one. It cannot lower Google's bar, and no consent tool can; what it can do is make sure the signals underneath were correct when you clear it. The mechanism is on the product page.