Signals Corps read headers from 217,097,223 page responses (10% of Common Crawl's CC-MAIN-2026-39). 7,959 sites sent a header written for AI: 7,271 were pointers (6,989 via x-llms-txt) to llms.txt files, MCP servers or skills, 511 were policy statements, 139 reported classifying the client as an AI bot, and 30 tried to give an agent orders, 24 of them in a header named X-AI, with 22 of the 30 judged pranks. Headers only matter if the agent's fetch path copies them into context: curl -i in a shell tool, Gemini CLI's off-by-default direct fetch on error statuses, redirect targets in summarizing fetchers, the MCP reference fetch server's Content-Type and robots.txt edge cases, and Playwright MCP's network-request tool all do; body-only fetchers don't. In their test, Claude models in Claude Code ignored the test header unless the user's prompt gave the site authority (then 7 of 9 runs obeyed). The bigger takeaway is cloaking: Common Crawl only sees what's served to CCBot, so agent-only content wouldn't show up in these counts.