Modern AI-driven advertising systems, including Google Performance Max and Meta Advantage+, rely on conversion event data as their primary training signal. This paper demonstrates that adversarial contamination of these signals constitutes the binding constraint on algorithmic advertising performance, a constraint that no amount of computational sophistication can overcome. We develop an information-theoretic framework showing that the data processing inequality imposes a hard ceiling on optimization quality: when conversion data is poisoned, algorithmic decisions inherit and amplify the corruption. We formalize a four-generation taxonomy of advertising fraud evolution, from volumetric bot traffic (Generation 1) through sophisticated invalid traffic (Generation 2), conversion-level data poisoning (Generation 3), to an emerging class of autonomous agent collusion (Generation 4). Applying results from PAC learning theory, we prove that at empirically observed fraud rates of 25% in lead generation (rising to 67-85% in audience networks), sample complexity inflates by 4x to over 100x, pushing many campaigns past the threshold of learnable signal. Economic analysis reveals that the total cost extends well beyond the estimated 84 billion in direct annual losses (projected to 172 billion by 2028): corrupted optimization data creates a compounding feedback loop in which platforms allocate increasing budget to fraudulent sources, producing a measurable divergence between reported and verified returns. We propose signal hygiene, a verification paradigm that validates the event rather than surveilling the user, as the structurally privacy-compatible alternative to surveillance-based fraud detection.
Ivitskiy et al. (2026) studied this question.