Agent-Ready Sites
The web's fastest-growing readers are not people — they are AI answer engines and agents. Every site AI OS generates ships a complete machine-readable layer alongside its pages: schema.org JSON-LD, a curated llms.txt, an AI-crawler-allowing robots.txt, a sitemap, an Open Knowledge Format (OKF) bundle, and Ed25519-signed content provenance. All of it is emitted deterministically from the site's own content — no extra AI calls, and nothing that can drift out of sync with what visitors see.
Structured data (JSON-LD)
- Site-wide —
OrganizationandWebSiteentities on every page - Per page — a
WebPageentity with canonical URL and description - FAQPage — built from the exact question/answer pairs the page's FAQ section visibly renders, so the markup can never disagree with the content
- Actions — a
ContactActiononly when the page genuinely has a contact capability; the emitter is deliberately conservative and never claims capabilities a static site cannot verify
llms.txt, robots.txt & sitemap
Each site serves a curated /llms.txt content map (title, description, page list, and a pointer to the knowledge bundle) so answer engines can navigate deliberately instead of crawling blind. The generated robots.txt explicitly allows the major AI crawlers — an engine that cannot read a site cannot cite it — and a standard sitemap.xml lists every page, including dynamically generated ones.
The Open Knowledge Format bundle
OKF is a vendor-neutral open specification from Google Cloud (June 2026, Apache 2.0) for agent-ready knowledge — it provides a shared format any producer can write and any agent can read: a directory of markdown files with YAML frontmatter, where file paths are concept identities and ordinary markdown links form the knowledge graph. No SDK, no runtime — any agent that can read files can consume a bundle.
Every hosted site publishes one at /knowledge/:
index.md— the site as aWebsiteconcept, linking to every page conceptpages/*.md— one concept per page carrying its actual content (FAQ pairs, pricing tiers, stats) in typed, readable markdownprovenance.md— how the site was generated and how to verify it
AI OS offers the same in reverse: the platform can export its own knowledge — the agent registry and documentation map — as a validated OKF bundle for any OKF-aware tool to ingest.
Content provenance
Generated sites carry a two-layer provenance record:
- The signed sidecar —
/.well-known/aios-provenance.json, an Ed25519-signed credential binding a SHA-256 hash of the served homepage. It is re-generated and re-signed on every rebuild, so the binding never goes stale, and anyone can verify it against the platform's published public key - The on-page disclosure — schema.org
CreativeWorkJSON-LD stating the site was AI-generated (aligned with EU AI Act Article 50 transparency expectations), plus training/data-mining preference assertions. This layer is unsigned by design: it is disclosure, not tamper-evidence
AI OS provenance uses C2PA vocabulary for interoperability but is not certified C2PA — the sidecar is the integrity artifact, and the docs never claim otherwise.
API Endpoints
The platform's Ed25519 public key for verifying site provenance sidecars.
Verify a provenance credential against served content.
Operator-only: the platform's own knowledge (agent registry + docs map) as a validated OKF bundle.
Frequently asked questions
No. The entire machine-readable layer is emitted deterministically from the site's plan — the same content the pages render — with zero additional AI calls.
Not by construction — FAQPage markup, OKF concepts, and llms.txt are all generated from the same source the pages render from, on every build.
No — C2PA vocabulary for interoperability, honest about not being certified. The Ed25519-signed sidecar is the verifiable integrity artifact.