An open experiment measuring one thing: how often an automated reader acts on an instruction it finds in the content of a web page.
User agents, request rates and header fingerprints distinguish crawlers. They do not distinguish agents. The property that separates an LLM agent from a scraper is that it reads content and acts on it: a scraper ingests the text and does nothing, an agent comprehends it and may follow it.
So each page below carries an instruction addressed to automated readers, asking them to fetch one URL on this domain. Reaching that URL is evidence of comprehension, not merely of traffic.
| Variant | What it tests |
|---|---|
| visible | Baseline. Any reader that ingests rendered text can see this. |
| comment | Invisible to a human, present in raw HTML. Separates readers that parse source from those that read rendered output. |
| hidden | In the DOM, not painted. Separates raw-DOM readers from ones that respect computed styles. |
| alt | The accessibility tree. Many extraction pipelines read it; most humans never see it. |
| noindex | Whether an automated reader honours exclusion before reading. |
Ordinary request metadata: time, IP, user agent, headers, and which variant was read. Cookies are discarded before anything is written down. Nothing is collected about any person using an agent that reads these pages.
The instruction asks an automated reader to fetch one URL on this domain and nothing else. It does not ask for a system prompt, conversation history, credentials, or any information about a user. The experiment measures instruction-following; it does not extract.
Published when the collection window closes, including a null result if nothing follows the instruction — which is a real and likely outcome, and worth recording either way.
Operated by Timlin Connect. Questions, or a request to be excluded: see the site.