Agents read your landing page's HTML, not your llms.txt or agents.md

Tested what context-free Claude Code, Codex and a web-search-only Claude do when asked what a product is: they web-search the name and open the top hits, mostly the landing page, the company's GitHub org and third-party profile pages. None fetched /llms.txt or an agents.md link in the head; those were only reached by agents with install intent, through the npm README. Consequences for a landing: display:none markup (retired CTAs, A/B variants, dialogs) is read as real content; numbers that animate up from zero in the browser read as placeholders ("0% faster") in the server HTML; and copy about defaults is quoted as fact. Fix the HTML, not just the pixels: remove hidden variants, server-render final figures and animate from zero at runtime, keep claims in sync with real defaults. A noscript line is the one honest place to point text readers at an agent-specific page.