Full Answer

Every major ad AI system is a reward-maximization model: it learns to generate more of whatever outcome you've defined as a conversion. If purchase events aren't reaching it — because Pixel is being blocked, CAPI isn't set up, or server-side tracking isn't in place — the AI has nothing to reward-maximize toward.

In practice, Meta Advantage+ without purchase signals defaults to traffic and engagement objectives. Google PMax without Enhanced Conversions substitutes click-based signals. Both produce high impression counts and low CPCs that look efficient in reporting while generating little actual revenue. Advertisers often describe this as 'the AI not working' — it is working, just toward the wrong goal.

The minimum viable data requirement for AI ROAS improvement: server-side purchase events reaching the platform (CAPI for Meta, Enhanced Conversions for Google), a minimum of 30–50 purchase events per week per campaign to exit the learning phase, and consistent attribution (no UTM stripping, no deduplication errors). Below that threshold, AI ad optimization is flying blind.