Context
Wadi and Ma report controlled Tool-Lab experiments across 16,400 sessions and eight LLMs from three providers. Vague shopping goals and costly product-attribute inspection often caused agents to skip information needed for unit-price comparisons, producing suboptimal choices under complete-information criteria.
What changed
The study introduces observable tool-call process tracing to separate failure to acquire diagnostic product facts from failure to use facts already acquired, with full-information and partial-information validation conditions.
Why it matters
AI shopping discoverability includes whether agents inspect relevant product attributes, not only whether a product is retrieved or recommended. Product data accessibility and instruction specificity may affect which evidence is used during selection.
Visibility OS interpretation
Visibility OS interpretation: evaluate the attribute-acquisition stage of shopping-agent journeys separately from product retrieval and final recommendation; straightforward price, size and discount facts may reduce omission risk. This is not evidence of a commercial search ranking factor.
What remains uncertain
Do the effects generalize to live storefronts, non-coffee categories, real token/latency costs, more products, and conversational shopping products? How stable are results across model versions?