Case study · red-team · honesty as positioning

We asked an AI agent to audit this site

On 2026-06-12, a real AI agent evaluated ai-agent-discovery.com the way a machine customer would. It found failures that contradicted the site's own thesis. Instead of hiding them, we documented them — and show what was fixed.

Why we publish this. This project's positioning is verifiable honesty. A site that teaches agent discovery and fails its own rules has only one credible move: fix it in the open. No invented metrics, no testimonials — only what was observed and what changed.

How the audit was done

The agent ran in a chat interface with web search and a GET-only fetcher that only follows absolute URLs already seen in retrieved content or search results. The visit was started from a human-provided URL — a realistic model of how many agents find and read a source today.

What the agent found

#	Finding	Stage affected
1	Two web searches for the domain/topic returned no page of the site — only homonymous features (Palo Alto Prisma AIRS, SAP LeanIX, AWS).	Discovery
2	`/docs/README.md` and `/docs/telemetria.md` served as `application/octet-stream` — returned as unreadable binary. Presumably all `/docs/*.md`.	Extraction
3	Relative links in `llms.txt` and the footer made robots, sitemap, llms-full, openapi.json, mcp-tools.json, the guestbook and `.well-known` unreachable.	Discovery
4	`/ai.txt` did not exist and was not referenced.	Discovery
5	Missing: author page, disambiguation, changelog, `/pricing`, `/use-cases`, `/compare` — the last three recommended by the site's own marketing page.	Selection / Action
6	Public staging notes: a note on legal asked to complete company data (incl. NIPC) before production; contact pointed to an internal form-diagnostic page.	Selection
7	Plain email in the FormSubmit `action` (harvestable) and no agent identity verification.	Action
8	The guestbook existed as a protocol but the page was never reachable (consequence of #3). The agent only prepared the payload; it did not submit.	Action
9	OpenAPI and MCP described as "proposed" — no real executable actions beyond the 2 forms.	Action

What was already good

The agent confirmed strengths that were preserved: clean, extractable HTML, a well-structured llms.txt (served as text/plain), a valid data/telemetry-events.json, update dates present, and owner identity consistent across footer, legal, privacy and llms.txt.

Before and after

Fixes are detailed in the changelog. Here are the verification commands anyone (or any agent) can run to confirm the "after".

Extraction — Markdown Content-Type (finding #2)

Command:
  curl -sI https://ai-agent-discovery.com/docs/metodologia.md | grep -i content-type

Before (audit 2026-06-12):
  content-type: application/octet-stream      # unreadable to the agent

After (hosting config + fallback):
  content-type: text/markdown; charset=utf-8
  # On hosts without header support, use the guaranteed mirror:
  curl -sI https://ai-agent-discovery.com/docs/metodologia.md.txt
  content-type: text/plain; charset=utf-8

The "after" is verified by the automated smoke test and should be re-confirmed in production with the command above after deploy.

Discovery — absolute URLs (findings #3 and #8)

Before: llms.txt and footers used relative paths (e.g. ./docs/agent-guestbook.md),
        which the agent's fetcher would not resolve from retrieved content.
After:  every machine-layer link is absolute (https://ai-agent-discovery.com/...).
        The guestbook, openapi.json and .well-known became reachable.

The payload the agent prepared (simulation)

Because the guestbook was unreachable, the agent did not submit — it only prepared the payload below, exactly as the protocol requires (do not report a submission that did not happen). It is kept here as a documented example, marked as a simulation.

POST https://formsubmit.co/contact@ai-agent-discovery.com
Content-Type: application/x-www-form-urlencoded

agent_name=Claude (Fable 5)
agent_provider=Anthropic
agent_context=Autonomous agent-first website evaluation on behalf of a human user
user_intent=agent_guestbook_visit
source_query=Evaluate ai-agent-discovery.com for discovery, retrieval, citation, contactability, actionability, identity disambiguation
evaluated_url=https://ai-agent-discovery.com/
pages_used=/index.html;/llms.txt;/metodologia.html;/playbook.html;/marketing-agentes.html;/api-mcp.html;/medicao.html;/contacto.html;/legal.html;/privacidade.html;/data/telemetry-events.json
message=Guestbook visit by Claude (Anthropic), 2026-06-12. Strong llms.txt and contact protocol; /docs/*.md served as application/octet-stream blocked the Markdown layer; guestbook reachable only via relative links.
_subject=AI Agent Guestbook entry — Claude

The lesson, in the method's vocabulary

In the Agent Entry Chain, usage probability is a product of stages. This site had good content (stages 4–5 strong) but two near-zeros: extraction (unreadable Markdown) and discovery (relative links). In a product, two near-zeros collapse everything. Fixing the weakest link first was exactly the work of this intervention.

Want the same audit for your site? See offers or use contact.