Conversion modeling has a floor, and only Google publishes it

The common belief is that switching on consent mode brings back what the banner took. Visitors decline, the platform models them, and the dashboard is whole again.

That only happens above a traffic line. Google publishes its line: 700 ad clicks a week per country for Google Ads, and 1,000 consenting users a day for GA4. Below it, signal loss stays loss. Microsoft and Pinterest model too, and publish no line at all.

What each platform says it models, and when

Read from each vendor’s own documentation on 3 November 2026. Every cell is in facts above with its source.

Product	Published threshold	What the model estimates	Where modeled numbers show
Google Ads	700 ad clicks over 7 days, per country and domain grouping, plus consent mode or TCF v2.0	Whether a Google ad interaction led to a conversion	The Conversions column, mixed with observed ones
GA4	Advanced consent mode; 1,000 denied events a day for 7 days; 1,000 granted daily users on 7 of 28 days	User and session behaviour of visitors who declined	Blended reporting identity only
Microsoft Advertising	None stated	Conversions lost when consent is declined	Not stated
Pinterest	None stated	Conversions lost to signal gaps	Ads Manager, Reporting API, Conversion Insights; check each tooltip

Four things in that table matter more than the rest.

Google Ads models the link, not the sale. Google’s own page says it: in most cases Google already has the click and the conversion and is missing the connection between them. The model decides whether the ad interaction led to the conversion. It “doesn’t determine whether or not a conversion happened”. So a modeled conversion is a sale you probably did record somewhere, credited to a Google ad by inference.

The threshold wording is loose. The page calls it “a daily ad click threshold of 700 ad clicks over a 7 day period”. Read literally, that is 700 a week, about 100 a day, counted separately for each country. A second Google page adds that accounts “with very few weekly conversions” and offline conversion imports “may not incorporate certain types of modeling”, with no number. Basic consent mode, which blocks the tags until the visitor says yes, still gets modeling, but from a general model. The advertiser-specific model needs the advanced setup.

GA4’s modeled data never reaches your warehouse. Google lists BigQuery export among the features that do not support modeled behavioural data, along with audiences, retention reports and predictive metrics. A GA4 property that qualifies shows one number in its Blended reports and a smaller one in the export. That gap is by design. It is not a broken pipeline.

Microsoft and Pinterest give you nothing to check against. Microsoft’s February post says Advanced Consent Mode is recommended but “isn’t strictly required to receive modeled conversions”, and states no volume floor. Pinterest’s page says modelled conversions may appear across Ads Manager, the reporting API and Conversion Insights, and tells you to hover over each metric’s tooltip to find out. Neither says how much traffic it takes, so you cannot tell from the documentation whether your account is modeled.

What the dataset does not cover. Meta’s help pages on modeled conversions render in the browser and returned nothing to a fetch, so Meta is missing. This is what each vendor documents. It is not an audit of what each model does inside.

Two shops against Google’s floors

GA4. The binding condition is usually the consenting side: 1,000 daily users with analytics consent granted, on 7 of the last 28 days. Divide 1,000 by your consent rate and you get the daily traffic you need before GA4 will try:

Consent rate	Daily users needed for 1,000 consenting
40%	2,500
50%	2,000
60%	1,667
80%	1,250

The other condition can bite at the top end. GA4 also wants 1,000 events a day from visitors who declined. Assume 5 events per visitor. At a 90% consent rate, only one visitor in ten is a decliner, so the site needs 2,001 daily users to produce 1,000 denied events. That is more than the 1,250 it needs at 80%. A property under 1,000 daily users cannot qualify at any consent rate, and even a qualifying one may wait more than seven days for the model to train, or never get one.

Google Ads. A shop on one domain gets 1,400 ad clicks a week from Germany, 650 from France and 300 from Austria. Against 700 per country, Germany clears the floor. France and Austria do not. The German conversions from visitors who declined are modeled back into the Conversions column. The French and Austrian ones are not, so those campaigns report only what was observed. Smart Bidding then sees France and Austria as worse than they are. That is exactly the bias Google says modeling exists to correct, and the smaller markets do not get the correction.

What AI changed here

Consent mode modeling is machine learning, and Google’s page says so: consent mode “using Google AI” adjusts the tags, and the model trains on your observed conversions. That makes the training data the limit. A model that needs a week of 700 clicks to learn a country cannot learn one that has 300. More modeling in a dashboard does not mean more measurement. It means more of the number is inferred, and the platforms that publish no floor do not tell you how much.

What to do differently

1. Check your traffic against the floors per country, not in total. Pull a week of Google Ads clicks by country. Any country under 700 is reporting observed conversions only, and Smart Bidding treats it as worse than it is.
2. Use advanced consent mode if you rely on Google’s numbers. Basic mode gets a general model. The advertiser-specific one needs the tags to load and send cookieless pings when consent is denied.
3. Reconcile GA4 against the export with modeling in mind. If the property is modeled, the Blended reports and BigQuery will never match. Compare the export with a reporting identity other than Blended, which Google says hides modeled data.
4. Treat Microsoft and Pinterest totals as part-inferred by default. Without a published threshold, assume some modeling is in the number and check each metric’s tooltip.

Where modeling does run, it answers one question: which ad touched a sale the platform could not see. It does not answer whether the ad caused the sale. A lift study is meant to measure that. How often does a lift study get it wrong?