Bidding bias from signal loss

Most ad bidding systems are trained like a card counter who only sees half the deck. When browsers block third-party cookies and users run ad blockers, the system keeps getting clean feedback from a smaller, easier-to-track slice of people and starts bidding harder for them. (webkit.org, support.mozilla.org) That slice is not random. Safari has blocked third-party cookies by default for years, and Firefox’s Total Cookie Protection locks cookies to the site where they were created, so the people still fully visible to ad platforms are disproportionately the ones on Chrome-like setups, logged-in environments, or pages with fewer blockers. (webkit.org, support.mozilla.org, apple.com) Google added another twist in April 2025 when it said Chrome would keep its current third-party cookie approach instead of rolling out a new standalone prompt. That means the market did not become uniformly “cookieless”; it became patchy, with strong signals in some traffic and weak signals in other traffic. (iabuk.com, chromium.org) A bidding model treats that patchy visibility like truth unless someone corrects it. In machine learning, this is sampling bias: the examples that make it into training are not representative of the whole population, so the model learns the habits of the observed group instead of the market it is supposed to bid on. (developers.google.com, en.wikipedia.org) The ugly part is that the “visible” users can look dramatically better than everyone else even when they are not. Google says Consent Mode modeling can recover more than 70% of ad-click-to-conversion journeys lost to consent choices on average, which is another way of saying a large chunk of real outcomes disappears from direct observation before the bidder ever learns from it. (blog.google, support.google.com) Once that happens, the auction starts overpaying for the people it can still measure. If conversions from privacy-protected users go missing while conversions from trackable users stay visible, the trackable group can show apparent return-on-ad-spend lifts of 20% to 40% simply because more of their wins are counted. (x.com, support.google.com, facebook.com) Then the feedback loop kicks in. Higher measured performance wins more budget, more budget creates more logged conversions from that same easy-to-see segment, and the next training cycle becomes even more convinced that the visible minority is the whole market. (developers.google.com, arxiv.org) This is why “just optimize to observed conversions” breaks in privacy-heavy markets. In missing-not-at-random data, the chance that a label is missing depends on the user or outcome itself, so the absence of data is not noise; it is part of the pattern distorting the model. (arxiv.org, jonathantemplin.github.io) Ad platforms already use statistical patches to fill some holes. Google Ads reports modeled conversions for eligible Consent Mode setups, and Meta says it models conversions when user-level data is missing or partial because of privacy preferences. (support.google.com, facebook.com) But modeling missing conversions is not the same as fixing biased training data inside the bidder. If the auction still learns mostly from users with stable identifiers, clean browser signals, and complete post-click data, it will keep steering spend toward the people who are easiest to measure rather than the people most likely to buy. (developers.google.com, arxiv.org) The fix is less glamorous than a new identity graph. Teams have to reweight samples, audit performance by browser and consent state, separate “easy to observe” from “high intent,” and train models to account for the probability that a conversion was never seen in the first place. (developers.google.com, arxiv.org, iabtechlab.com) If they do not, signal loss turns into bidding bias, and bidding bias turns into fake efficiency. The dashboard looks cleaner, the tracked users look richer, and the campaign quietly learns to ignore everyone wearing the modern web’s privacy armor. (iab.com, webkit.org, support.mozilla.org)