A side-by-side benchmark of AI assistants doing real digital-asset research workflows, with and without the Perception MCP connected. Question: does connecting an industry-specific data corpus to a frontier AI assistant produce measurably better research work than the same assistant with its native web search? Answer, across 48 scored runs: yes. Blind-judged mean score 11.4 → 16.4 (max 25, +44%), with the widest gains in recency (+74%) and… See the full description on the dataset page: https://huggingface.co/datasets/ferniko/perception-mcp-benchmark.