Context
WANDR introduces 500 realistic data-collection tasks for research agents. Tasks require broad entity discovery, deeper investigation through coordinated web searches, and independently verifiable records with supporting sources and excerpts. Six production research systems were evaluated; at high effort, the strongest reached 0.363 soft F1 and 0.133 hard F1.
What changed
The benchmark replaces static gold answers with task-specific judges that refetch cited pages and verify records against evidence. It measures soft and hard precision, recall, and F1 across hierarchical research tasks and reports declining performance as target volume and hierarchy depth increase.
Why it matters
Agentic discoverability is not only a single-query ranking problem. A source or entity may need to be found during broad discovery, revisited during enrichment, and retained in an evidence chain. WANDR provides a checkable framework for measuring those stages together.
Visibility OS interpretation
Visibility OS treats WANDR as evidence that measurement of agentic discovery should separate breadth of discovery, depth of enrichment, factual correctness, and evidence completeness. The benchmark does not establish a ranking formula or persistent behavior for any named consumer AI-search platform.
What remains uncertain
How well do WANDR results transfer to consumer AI-search products and other research-agent architectures? Which source properties most affect failures in discovery, enrichment, and evidence construction?