The signal loss
Privacy changes have steadily removed the signals advertising relied on. Third-party cookies are being deprecated or restricted, browsers limit cross-site identifiers and cap cookie lifetimes, and consent requirements mean some users are not tracked at all. The result is that a growing share of conversions and audience behaviour can no longer be observed directly the way it once was.
This is not a temporary disruption but a structural shift to a more privacy-constrained environment, and it affects everyone at once. The old model, deterministic tracking of individuals across the web, is giving way to a world of partial observation, consent gates, and privacy-preserving techniques. Understanding this as the new baseline, rather than a problem to be worked around, is the starting point for measurement that lasts.
How AI fills the gap
The platforms' answer to the signal loss is modeling: using AI to estimate the conversions and audience behaviour that can no longer be directly observed. When a user does not consent or cannot be tracked across a journey, the platform models how many of those interactions likely converted, based on the patterns of users it can observe. Consent Mode's cookieless pings feed this modeling, and modeled conversions blend into the reported numbers.
This keeps automated bidding supplied with a complete-enough picture to optimize, which is the point: without modeling, bidding would be starved as observable signal shrinks. AI also powers privacy-preserving audience techniques that group users rather than tracking individuals. Modeling is genuinely useful and, in a privacy-constrained world, necessary, but it changes the nature of the numbers from observed to estimated.
What modeling can and cannot do
Modeled numbers are estimates, accurate in aggregate when the model has plenty of observed data to learn from, but not a row-for-row source of truth. They are least reliable on thin-volume accounts or where consent rates are very low, because the model extrapolates from a small observed base. And a reported total that blends observed and modeled conversions can obscure how much of it is estimated.
So modeling cannot fully substitute for real, observed signal. The more that has to be modeled, the more the numbers are inference rather than observation, and the harder they are to reconcile against a backend. This is why the goal is not to trust or dismiss modeling wholesale, but to reduce how much has to be modeled by strengthening the observed, consented signal underneath it.
