The formal technical specification governing how enterprise websites structure knowledge, eliminate hallucinations, and secure preferred source status in large language models.
All participating digital entities MUST expose an /llms.txt plaintext manifest at the web root and under /.well-known/llms.txt. Payloads must maintain token efficiency (<8,000 tokens) and exclude styling, client JavaScript, and navigational boilerplate.
Generative search rerankers compute Information Gain IG = H(D) - H(D|T). To qualify for primary citation attribution, published passages MUST maintain a quantitative density score >= 0.75, containing verifiable empirical metrics, time-stamped benchmarks, and non-generic terminology.
Atomic passages intended for RAG vector retrieval MUST be semantically self-contained. The opening sentence of any extracted passage MUST NOT employ ambiguous third-person pronouns ('it', 'they', 'this system') without antecedent definition.
To defend against AI model hallucinations and competitor displacement, domains MUST provide cryptographically verifiable authority certificates signed via SHA-256, corroborating entity ownership across models.
AI web crawlers (GPTBot, ClaudeBot, PerplexityBot) enforce aggressive timeout thresholds (<= 2,500ms total crawl window). Edge reverse proxies MUST serve cached schema graphs and machine protocols with TTFB <= 50ms globally.