What an optimized artifact is
An optimized artifact is a purpose-built representation of a single page, produced by pure extraction and restructuring of your existing page. It is composed of three parts:Answer-first markdown
Your page’s content reformatted so the most direct answer appears near the top — the page title as an H1, a lead answer block, then each of your headings followed by its nearest paragraph.
JSON-LD structured data
Deterministic
Article schema.org markup, plus a FAQPage block when your page’s own headings are phrased as questions answered by the following paragraph.Clean-HTML mirror
A stripped-down HTML rendering of the same extracted content, stored alongside the markdown.
How artifacts are generated
Bluesky generates artifacts offline, in the background, from your existing published pages. There is no large language model on the request path — when a verified AI agent arrives at your site and the edge middleware fetches an artifact, it is retrieving a pre-built document from Bluesky’s control plane, not triggering a live generation step.Because generation is offline and pre-cached, artifact delivery adds no model latency to the request path.
Your published page is read
Bluesky fetches the live URL. Scripts, styles, navigation, headers, footers, asides, and forms are dropped so only the semantic content remains.
Content is extracted and restructured
The page title, meta description, section headings, and substantial paragraphs are pulled out and composed into answer-first markdown, an
Article/FAQPage JSON-LD block, and a clean-HTML mirror — all sourced from the page itself.Parity is enforced
A grounding guardrail compares the generated tokens against the source. If the markdown introduces any content word the source didn’t contain, the artifact is rejected and never served.
Strict parity: no fabrication
The single most important property of an artifact is that it never asserts anything your page didn’t. Two gates enforce this before an artifact is ever marked servable:Grounding guardrail
Grounding guardrail
The generator tokenizes both the source text (title, description, headings, paragraphs, and full main text) and the generated markdown, then checks for any content token present in the artifact but absent from the source. For pure extraction this set is empty. A non-empty result marks the artifact as an error — it is not stored as fresh and is not served. This one check defends against both hallucination and cloaking.
Minimum content
Minimum content
If the restructured markdown has too few content tokens, the page has no usable content and the artifact is marked as an error rather than served. Thin pages simply pass through to the human page.
Only artifacts with
status: fresh are ever served. Anything marked as an error or stale results in a transparent passthrough to your normal page.Citability score
Each fresh artifact is assigned a citability score between 0 and 1 (shown as a percentage in Agent Analytics). It is a heuristic signal of how citable the artifact is, computed from the extracted structure — not a live model call. The factors it rewards:| Factor | Contribution |
|---|---|
| Page has a title | +0.20 |
| Meta description present (answer-first lead) | +0.15 |
| Section headings present (chunkable sections) | +0.20 |
| Last-modified date present (freshness signal) | +0.10 |
FAQPage JSON-LD generated | +0.15 |
| Sufficient content density in the markdown | +0.20 |
Which URLs get served the artifact
The edge middleware only serves an optimized artifact for URLs that have a fresh artifact on file. If a verified AI agent requests a URL that hasn’t been generated (or whose artifact is stale/error), the artifact endpoint returns a 404 and the middleware passes the request through transparently — the agent receives your normal page, no error is returned, and the visitor experience is unaffected.Passthrough is silent and safe. A missing artifact never degrades the experience for any visitor — human or agent.
How the middleware fetches an artifact
When a verified answer-engine request is eligible and the serve mode isdefault, the middleware computes a SHA-256 hash of the full request URL and uses it to look up the artifact from Bluesky’s edge API:
Authorization: Bearer <edgeToken>). If the API returns a fresh artifact, the middleware responds with the markdown body and these headers:
| Header | Value |
|---|---|
content-type | text/markdown; charset=utf-8 |
x-aen-served | variant |
cache-control | no-store |
Serving only happens in
default mode for verified, eligible agents. In shadow mode the middleware logs what it would serve but always returns the human page. See Serve modes for the full vocabulary (off / shadow / canary:<pct> / default).Regenerating an artifact
After you update a page’s content, regenerate its artifact so the served markdown and its citability score reflect your latest content.Open Agent Analytics
In the Bluesky app, go to Agent Analytics and select the domain you want to inspect.
Related
Serve modes
off / shadow / canary / default — how you control when the variant is served.
Agent Analytics
See what was served versus cited, broken down by purpose and trust.