What are modeled conversions

An advertising report rests on a simple idea: someone clicks, someone buys, and the system joins the two through an identifier. That chain breaks every time the permission is missing or the identifier is missing. If the visitor declines advertising cookies, if the browser deletes the identifier before the purchase happens, or if the click occurs on the phone and the order is completed on the laptop, the sale exists but the system does not see it.

Modeled conversions are Google's answer to that gap. Instead of leaving the cell empty, the platform calculates how many conversions probably occurred in the unobservable part of the traffic and adds them to the report. They are not individual records recovered later. They are an aggregate figure built from the behaviour of the traffic that can be measured.

The practical distinction is this: an observed conversion has a concrete event with a timestamp behind it, while a modeled conversion has a probability calculation behind it. Both appear together in the same report column, with no label separating them, and that silent blend is what confuses most teams.

Google does not invent the figures, it derives them from comparable slices of traffic. The documentation describes the method as using “observable slices” whose behaviour resembles that of the slice which cannot be measured. If the conversion rate of a campaign is known in a browser where consent was granted, that pattern serves as the basis for estimating what happens in a browser that blocks measurement.

The model does not rest on a single factor. Among the signals feeding the calculation are device type, time of day, geographic location and operating system, plus the account's own conversion history. The more alike the measured slice and the estimated slice are, the more reliable the extrapolation becomes.

There are minimum conditions. Google applies confidence thresholds and only publishes modeled conversions when it considers it highly likely that those conversions actually happened. An account with very few weekly conversions may not receive certain types of modeling at all, simply because there is not enough observable base. In Google Analytics 4 the thresholds are published: behavioural modeling requires at least 1,000 daily events with analytics_storage='denied' over seven days and at least 1,000 daily users with consent granted on seven of the previous 28 days. Even when an account meets them, the model applies its own quality criteria and may not kick in.

The calculation is not instant either. The figures need up to five days to process and stabilise, so yesterday's number is almost never the final number.

The point where this hurts is not theoretical. A modeled conversion is an aggregate estimate and cannot be traced back to a single click. There is no record row, no click identifier, no user. There is only a probability that n conversions occurred in that slice of traffic.

When someone takes the conversion column, which already holds observed and modeled values together, and sets it against a campaign's spend, they are working with two different quantities inside the same figure. One part comes from events that happened and were recorded, the other from a statistical calculation whose margin of error appears nowhere in the report. The resulting ROAS looks like an exact number and is not one.

The effect gets worse as you drill down. The finer the segmentation (by campaign, by ad group, by keyword), the less volume remains in each cell and the more relative weight the estimated part carries. Decisions to pause or scale are taken at exactly that level, and there a difference the size of the model's uncertainty is enough to flip the ranking of two campaigns.

• Record which part of your measurement depends on modeling before results are discussed. Without that context, any month-on-month comparison mixes real business change with change in measurement coverage.
• Implement consent mode correctly and keep it stable. Modeling needs consistent consent signals, and a badly configured CMP degrades the estimate instead of improving it.
• Strengthen the observable part with first-party data: enhanced conversions, server-side measurement and clean campaign parameters shrink the gap the model has to fill.
• Always cross-check against an independent source, usually the orders in the billing system or the CRM, and work with the known deviation instead of pretending it does not exist.
• Wait at least five days before drawing conclusions from a recent period, because the figures are still consolidating.
• Decide budgets on wide time windows and sufficient volume, not on the daily view of a small ad group.

• Calculating a campaign ROAS from the conversion column as it stands and presenting it as an accounting figure, without flagging that it includes estimates.
• Reading the rise in conversions after switching on consent mode as a performance improvement, when what changed is the coverage of the measurement.
• Promising the client that platform conversions will reconcile with backend orders down to the last decimal, a goal that modeling makes impossible by design.
• Switching off consent mode or blocking signals in order to “see only real data”, which also removes the signal that feeds automated bidding.
• Analysing tiny segments, where the observed volume is so low that the estimate dominates the result and any conclusion is noise.