Post

1. Markdown copies of your pages. He ran it across hundreds of thousands of chats. In 0% of them did an LLM cite the .md version over the HTML page. What you get instead: duplicate content, double the crawl, a diluted internal link graph.

2. Schema as a delivery mechanism for facts. Published test: when the page text and the schema disagree, the text wins. When the fact lives ONLY in schema, ChatGPT answered ~20% of the time and Claude 0%. Schema is a forcing function for data hygiene. Not a pipeline.

3. llms.txt as a visibility play. He's never seen one lift visibility in an answer. Its real job is telling actual agents where your API, MCP and docs live. His line: he doesn't know anyone strict about measurement who thinks llms.txt is a lever.