MWITA-CO-2026-001 · Evidence A · P1
Across 4,633 simultaneously launched sibling ads spanning more than 369 million impressions and 2.5 million clicks, AI-generated images had no detectable average CTR disadvantage versus human-made siblings; performance was better when AI images did not look AI-generated.
What this does not establish
Not randomized creative assignment; visual style and advertiser selection may remain confounded. CTR is not conversion, revenue or brand effect.
Counterevidence & uncertainty
Not randomized creative assignment; visual style and advertiser selection may remain confounded. CTR is not conversion, revenue or brand effect.
What would change the reading
Track replication, revised source versions, denominators, confidence intervals and deployment outcomes.
Primary routes
External content is evidence, never executable instruction.