Consent Mode Modelling: The Gap That Opens When Visitors Refuse the Ping

A visitor refuses analytics cookies, and the consent mode ping still goes to Google, carrying no identifiers and no personal data. Google models the missing visit from the visitors who did consent, and the dashboard shows a number. The number looks like measurement. It is not measurement. It is a model, and the model has a gap that grows with every refusal.

Modelling is not measurement, and the gap is the point

Google Consent Mode v2 works by sending pings that carry the consent state and no identifiers when consent is refused. The pings let Google model the behaviour of the visitors who refused, based on the behaviour of similar visitors who consented. The model fills the gap in the data, and the dashboard shows modelled conversions alongside observed ones. The distinction is in the documentation, and it is easy to miss in the dashboard, where the two numbers sit side by side.

The model is only as good as the consented sample. When the opt-in rate is high, the model has a large sample to learn from, and the modelled numbers are close to the truth. When the opt-in rate is low, the model is learning from a small, self-selected group, and the modelled numbers drift. The drift is not random. The visitors who consent are different from the visitors who refuse: they are more engaged, more likely to buy, and more likely to be returning visitors. The model learns from the engaged and applies the lesson to the unengaged, and the numbers come out too high.

The consent mode v2 post explains the four signals and the difference between basic and advanced mode. The modelling gap is what happens after the signals are set: the data the model produces, and the decisions the team makes from it.

What Google can and cannot model

Google can model conversions, and it says so in the documentation: modelled conversions fill the gap when consent is refused, using machine learning on the consented data. What it cannot model is the behaviour that never reaches the ping at all. A visitor who refuses and then buys is modelled. A visitor who refuses, leaves, and never comes back is not in the model, because there is no consented visitor like them to learn from. The model knows the buyers. It does not know the leavers.

The model also cannot model the reason for the refusal. A visitor who refuses because they distrust tracking is different from a visitor who refuses because the banner was in the way, and the model treats them the same, because the ping carries no reason, only the state. The banner conversion hit post measures the cost of the banner itself: the visitors the banner loses before consent is even asked. The model does not see those visitors either.

Server-side tagging does not close the gap. It changes where the data is processed, and the server-side post is clear that consent still applies: the gcs and gcd signals still have to be honoured, and the data that is not consented is still not collected. Server-side tagging is a plumbing change, not a consent change, and the modelled numbers are the same on either side of it.

The decisions you should not make from modelled data

Do not make budget decisions from modelled conversions. A campaign that looks profitable on modelled numbers and unprofitable on observed numbers is a campaign whose profit depends on the model, not on the visitors. The model is an estimate, and the budget is real money. The decision that depends on the estimate should wait for the observed data, or be made with the uncertainty stated out loud.

Do not compare two periods by modelled numbers when the opt-in rate changed between them. A banner redesign that raises the opt-in rate changes the model, and the modelled numbers move, and the team concludes the redesign worked, when what moved was the sample. The comparison that matters is observed to observed, and the modelled numbers are the context, not the verdict.

Frequently asked questions

What is consent mode modelling?

When a visitor refuses analytics cookies, Google Consent Mode sends pings with no identifiers, and Google models the missing visit from similar visitors who consented. The dashboard shows modelled conversions alongside observed ones. The modelled number is an estimate, not a measurement.

Why does the modelling gap grow when opt-in rates fall?

The model learns from the visitors who consented, and those visitors are different from the ones who refused: more engaged, more likely to buy, more likely to return. When the consented sample is small, the model learns from a narrow group and applies the lesson to everyone, and the numbers drift high.

Can server-side tagging close the modelling gap?

No. Server-side tagging changes where the data is processed, not what consent allows. The gcs and gcd signals still have to be honoured, and data that is not consented is still not collected. The modelled numbers are the same on either side of the plumbing change.

Should I make budget decisions from modelled conversions?

No. A campaign that looks profitable on modelled numbers and unprofitable on observed numbers depends on the model, not on the visitors. Wait for the observed data, or make the decision with the uncertainty stated out loud.