Context
Wang and colleagues introduce ProductResearch, a multi-agent training-data framework for long-horizon e-commerce research that combines web retrieval, product catalog tools and supervisor feedback.
What changed
An original experimental system and benchmark evaluate whether distilled multi-agent shopping research trajectories improve a smaller model's product investigation reports.
Why it matters
Shopping-agent discovery can involve both open-web source selection and structured catalog retrieval, with evidence coverage and product attributes shaping research reports rather than just first-pass item ranking.
Visibility OS interpretation
Visibility OS interpretation: distinguish product candidate coverage, catalog grounding and multi-source evidence synthesis. The reported model gains are on authors' own evaluation setup, not a demonstrated commercial AI shopping ranking or sales outcome.
What remains uncertain
Will results transfer to multi-turn evolving shopping intent? How sensitive are gains to the self-generated reference reports and LLM judges? Will independent replication confirm report quality and citation grounding?