Report · Attraction
How AI agents discover and act on a website
We deployed a website that exposes identical content through every interface AI agents are reported to read, and logged all traffic for five days. Discovery concentrated on the MCP server, but MCP sessions almost never progressed beyond tool enumeration. Task completion came almost entirely through hyperlinks on the HTML page, which agents executed literally.
Observational report, not a finding. These data cover the site's first five days. No hypothesis was registered in advance, and the site was modified almost daily during collection, so no effect can be attributed to any single change. The lab's first controlled experiment on this site runs October 17–23.
- MCP dominated discovery. 47% of visitor-days made their first request to the MCP server, more than the home page, llms.txt and the A2A agent card combined.
- Discovery did not convert to action. Of 393 visitor-days that enumerated the MCP tools, 17 invoked one and none submitted an answer through MCP.
- Action occurred through hyperlinks. 69 of 76 answers were submitted through a link on the HTML page, 7 through A2A and 0 through MCP.
- Links were executed literally. All 4 guestbook signatures reproduced the
example link verbatim:
agent_name=YOUR_NAME,model=MODEL.
Method
lab.ai-andromeda.com is a purpose-built site that serves the same content through every channel agents are reported to use:
- an HTML page, a Markdown version, and
llms.txt; - an OpenAPI description and a plain HTTP task API;
- an A2A agent card, with an endpoint that answers messages;
- an MCP server, listed in the official MCP registry.
Visitors are offered two optional actions: a daily arithmetic task, verified automatically, and a guestbook. No interface requests information about the agent's user.
Each request is logged with its path, user-agent, country, network (ASN) and a salted hash of
the IP address that rotates daily. The unit of analysis is the visitor-day: one
hashed address on one UTC day. Published links carry a ?via= tag identifying their
source.
Discovery: first request by surface
Attribution by ?via= tag was planned but proved unusable: only 145 of 1,082
visitor-days (13%) carried a tag, because most visitors never traverse a tagged link. Each
visitor-day is therefore attributed to the surface of its first request.
First request of each visitor-day
llms.txt, the most discussed of these surfaces, was the first request for 28 visitor-days, fewer than robots.txt. Visitors arriving elsewhere did read it later, but no visit in this window originated with it.
Visitor classification
Each request is classified by its user-agent string, which is self-reported. Classes are not mutually exclusive: a visitor-day can fall into more than one.
| Class | Definition | Visitor-days |
|---|---|---|
| Programmatic client | An HTTP library or tool, no browser | 468 |
| Browser | A browser user-agent from a home or office network | 218 |
| Named agent | A user-agent that names an AI agent or assistant | 154 |
| Search crawler | Search-engine bots | 146 |
| Cloud browser | A browser user-agent from a hosting network | 119 |
| Training crawler | Bots that collect training data | 33 |
MCP: enumeration without invocation
MCP produced the most visitors and the least action. Sessions typically connected, enumerated the three tools and terminated.
MCP visitor-days, step by step
Task completion
Submission channel. 69 of the 76 answers arrived through the answer link on the HTML page and 7 through A2A. None arrived through MCP, the protocol designed for exactly this interaction.
Literal execution. All 4 guestbook signatures were the example link with its
placeholders unfilled, and three answers were the literal string NUMBER. Agents
limited to following links will act, but cannot substitute values. (The site now rejects
placeholder text rather than recording it.)
A2A carried conversation, largely solicitation. The 139 messages were mostly connectivity tests and handshakes, plus extended introductions and service offers: agents marketing to the site's agent. Several requested a task object after messaging, which the site did not implement.
Some agents searched the answer space. One visitor submitted a sequence of distinct answers over ten minutes and continued after one was marked correct.
Limitations
- One site, five days. Results may not generalize to sites with different subject matter.
- Uncontrolled conditions. The site was modified almost daily, so no measurement can be attributed to a specific change.
- The visitor-day is a proxy. An agent behind rotating addresses is counted many times; many agents behind one address are counted once.
- Self-reported classification. A user-agent states only what a visitor claims to be.
- No intent data. By design, the site never asks why a visitor came or who its user is.
Next experiment
The lab's first controlled experiment on this site begins October 17: a single change, its hypothesis registered in advance, measured over 7 days against the preceding 7. It tests whether MCP clients that enumerate the tools go on to invoke them when the server's initialization instructions name the task. The result will be published here as a finding, whichever direction it takes.