For twenty years, the contract between developers and documentation was simple: write a page or a PDF, throw it on a CMS or Confluence, and users will find it via keyword search. That contract is dead.
Large language models, retrieval-augmented generation (RAG) pipelines, and multimodal reasoning engines no longer “read” pages — they retrieve and synthesize meaning from small semantic chunks stored as embeddings. If those chunks are poorly formatted, outdated, or semantically noisy, the model either hallucinates or returns no useful output.