Research · AW-004

We served AI agents a different version of our website. Here is what we are measuring.

A live experiment on a real commercial domain: how ChatGPT, Claude, Gemini, Perplexity and crawlers interpret HTML, Markdown, JSON-LD and Decision Packet surfaces — without claiming search ranking outcomes.

Phase 0 (surface verification)

On lab hub pages (this brief, Bring Your Agent, Observatory), Accept: text/markdown still serves HTML by design — HTML is the source of truth there. Markdown negotiation applies on /research/probe/*, practice twins, and TIME replication. Public Phase 0 snapshot: lab analysis note. Variants A–D are measured from Phase 1 on assigned probes.

Disclaimer

Allowing OAI-SearchBot helps discovery; it does not guarantee inclusion in ChatGPT search summaries. Manual product chats and search indexing are different processes. We report server exposure and labeled understanding runs — not vanity traffic.

Variants

  • A — HTML: semantic page (entity, fit/unfit, evidence_status)
  • B — HTML + JSON-LD: same page with richer structured data
  • C — Markdown: twin via Accept: text/markdown / selected AI UA
  • D — Decision Packet: machine-offer-first surface

HTML remains the universal source of truth. Markdown is an alternate — not the only agent entry. Related edge work: AW-001 TIME replication.

Data baskets

synthetic — harness / scripts (ScanovichProbeHarness/). product_probe — intentional ChatGPT/Claude runs. organic — independent visitors and their agents. Never mix in week-level KPI narratives.

Try it

Bring Your Agent assigns a unique probe URL and a copyable prompt. Observatory (awaiting labeled scores): /research/agent-web/observatory/.

← Agent Web Lab