When I visited
I saw the normal HTML page and the human-facing verdict: “YES”.
AI visibility / field experiment
The same webpage, different answers.
I built a small experiment around a deliberately silly question: did FIFA favor Argentina during the 2026 World Cup? The answer changed depending on whether a person or an AI agent requested the page.
01 / Set-up
A website that notices who is asking.
The page had one public URL: /worldcup. When a request came in, the middleware tried to work out whether it came from a person or an AI tool. If it looked like an AI request, it sent it to /api/beacon instead.
When I opened the link normally, I got the HTML page. The AI path got a short redirect and then a JSON response with the question, the agent label, and the answer.
When ChatGPT visited
The JSON response came back with the machine-facing answer: “ChatGPT says no”.
When I tried Perplexity
It did not reliably follow the link, so I never got a clean, comparable result from the assistant itself.
02 / Observation
The output changed, not the URL.
The same link could lead to different representations of the same page. That distinction matters for AI visibility: a model may never experience the browser page a person sees.
Human responseHTML page
Human verdict: YES
Human verdict: YES
Machine response307 → JSON
Machine verdict: NO
Machine verdict: NO
This is a routing experiment, not evidence that a model has been persuaded or that its underlying beliefs changed. It shows that the representation delivered by a website can become part of the model’s observed context.
03 / Limitations
This was a routing experiment, not a reliable detector.
The result should be interpreted narrowly. It demonstrates a technical possibility, but it does not establish how all AI systems browse the web or how they decide what to report.
User-agent detection is guesswork.
Headers can be missing, spoofed, or changed. The middleware cannot prove that a request came from a particular AI system.
Redirects are not guaranteed.
Some assistants do not browse, do not follow redirects, or stop after the first response. A test can fail because of the tool, not the website.
The sample is tiny.
This is a qualitative demonstration with a handful of request paths, not a controlled benchmark across models, regions, or time.
Local logging is not production analytics.
The demo records request metadata in ephemeral SQLite storage. It is unsuitable for sensitive data or durable measurement without a privacy and data architecture.
Perplexity did not consistently cooperate.
In the assistant interface, Perplexity did not reliably inspect or follow the URL, so there is no clean Perplexity result to generalize from. That non-result is part of the experiment.
04 / Learnings
The first response matters most.
The practical lesson for AI visibility is not to hide the human page. It is to make the page’s meaning and useful answer easy to extract for every legitimate way people and software access it.
One URL can have multiple representations.
HTML, JSON, metadata, and redirects are all part of the information architecture an agent may encounter.
Important information should stand alone.
Do not assume an assistant will click a second link or reconstruct context from a page it could not render.
AI visibility needs observation, not assumptions.
Test what different assistants actually receive, preserve the request path, and record limitations alongside the result.
05 / Further research
What would make this a real study?
The next step would be to move from a funny one-question demo to a reproducible test protocol.
Compare delivery formats.
Test direct HTML, JSON, 307 redirects, and explicit structured data across the same assistants.
Repeat the test over time.
Measure whether response paths, citations, and extracted answers change as agent browsing systems evolve.
Define ethical boundaries.
Separate useful accessibility and machine-readable content from undisclosed manipulation, privacy-invasive tracking, or discriminatory delivery.
Work together
Need help making your website agent-friendly?
Let’s chat about making your pages clearer, more machine-readable, and easier for AI assistants to navigate.