snakereachWriting
Getting the web into a model.
Notes on the part everyone skips: what actually reaches the model, and how much of it is worth sending.
- Data · 31 July 2026We measured 16 sites. 77% of the HTML was not the page.Median page: eight characters of markup for every one of content. Here is the full table, including the site that blocked us.
- RAG · 31 July 2026Why your chatbot quotes the cookie bannerIf the answer contains a menu item, the model is not confused. It was fed a menu.
- Extraction · 24 July 2026Why LLMs read markdown better than HTMLMost of a web page is not content. Sending the rest costs you context, money and accuracy.
- Extraction · 27 July 2026Why models prefer structured contentStructure is information. Flatten a table into prose and you have destroyed the relationships in it.
- RAG · 29 July 2026RAG quality starts at extraction, not retrievalYou cannot embed your way out of bad input. The fix is three steps earlier than most people look.
- Fetching · 30 July 2026Why your scraper returns an empty pageA 200 with no content usually means you fetched the shell, not the page.