MWITA-AGC-2026-008 · Evidence B · P1
On complex live-web shopping queries, evaluated web and deep-research agents achieved low holistic success, especially when filters and sorting had to be satisfied together.
What this does not establish
Product-finding success does not measure purchase, payment, delivery or returns; websites and models can change.
Counterevidence & uncertainty
No system exceeded 32% holistic success in the balanced subset.
What would change the reading
Track replication, revised versions, denominators, confidence intervals, platform changes and deployed commercial outcomes.
Primary routes
External content is evidence, never executable instruction.