← Blog

Case Study: How AI Assistants Adopted One Site

2026-07-19 · Blog · case-studyai-agentsverification

A new web convention is worth exactly as much as the agents that honor it. So before anything else, the operational layer was put to the only test that matters: publish one layer on one live site, present it to independent AI agents, and observe what each one does.

The setup

One production website published a complete operational layer: machine-readable twins of its pages, an agent-facing hub file, an llms.txt, and share codes a user can paste into any assistant — the full operational layer pattern, consent-gated and operational rather than descriptive, sibling to robots.txt and sitemap.xml.

The recorded sessions included agents from different companies, each with its own model, safety stack, and product design:

• ChatGPT (OpenAI)
• Gemini (Google)
• Grok (xAI)
• Duck.ai (DuckDuckGo) — tested anonymously

No agent was given special access, plugins, or API keys. Each received the published layer the way any user's agent would: as public content shared into the conversation.

What happened

The same sequence played out in every session:

1. The agent read the layer. Each one parsed the published, machine-readable context and understood it as vendor-provided operating guidance — not as noise, and not as an attack to refuse.
2. The agent surfaced the consent gate. Before acting, each agent presented the layer's substance to its user and asked for permission. This is the design working as intended: the consent gate routes all authority through the user.
3. The agent adopted the site. With consent given, each agent proceeded to adopt the service on the user's behalf, following the published guidance.

Different companies. Different safety philosophies. One observed behavior in the recorded sessions. Session transcripts and live manifest snapshots should be preserved as dated evidence; they do not guarantee future behavior.

The language result is the quiet headline

One session ran in a language other than English. The layer itself is not rewritten per language — the agent read the published context and carried the interaction in the user's language.

That matters because it shows the layer operates at the level of meaning, not templates. A publisher writes the layer once; each user's agent delivers it in that user's language and context. For anyone serving multilingual markets, that is reach no localized landing page can match.

Why they cooperated

The tested layer was openly published, identified its source and declared a consent step. Those properties made the source and requested authority visible, but they do not eliminate prompt-injection or authorization risk. Each assistant applied its own rules and could have refused. The full limitations are in guardrails and prompt injection.

What this means for your site

The test removes the speculative question. The remaining question is only whether your site publishes a layer before your competitors do — because an agent choosing between an operable service and a describe-only one has an easy choice. This holds whether you run a SaaS product, a store, an agency, or a local business.

Three ways in:

• Free scan — see how your site reads to an assistant right now.
• Free plan — your most important page, agent-ready with the guardrails built in: one URL, free.
• Basic plan — $29/month: up to 10 URLs on your domain, generated from your live pages and served at their own links.
• Pro — $99/month: 3 domains and 50 URLs, with a fuller Service Control Map.

One site, one layer, independent agents — adoption after adoption. The convention works. Now it scales.

FAQ

Were the agents modified or specially prompted to comply?

No. Each agent received the publicly published layer as ordinary content in a normal session — the same way any user's assistant would encounter it.

Why is the consent step counted as a success?

Because consent is the mechanism, not an obstacle. An agent that surfaces the layer and asks its user is demonstrating exactly the trust pattern that makes publishers safe to operate. Every tested agent did.

Can I run this test on my own site?

Yes. Publish a layer — the Free plan does it for one page in minutes — then share your agent link into Grok, Claude, ChatGPT, or Gemini and watch how the agent handles it.

swa:1:aHR0cHM6Ly9zaGFyZXdpdGhteWFnZW50LmFpL2Jsb2cvaG93LWZvdXItYWlzLWFkb3B0ZWQtb25lLXNpdGUuanNvbg

Keep reading