Finding #001: I tried to measure AI citations as a guest. I couldn't.

Test date: October 18, 2026 Question set: 10 buyer-intent questions (local services + software), e.g. "best plumber in Austin TX", "top email marketing tool for restaurants" Assistants tested: Perplexity, Microsoft Copilot

What happened

My first experiment was dead simple: ask 10 buyer questions as a signed-out guest, write down every cited source. It produced zero data points. Not one answer — neither assistant would talk to me at all.

• Perplexity: every query came back with "Sign up and repeat your request." No answer, no sources. I tried incognito. The third attempt earned me a full "Sign in to continue" modal. Message received.
• Microsoft Copilot: copilot.microsoft.com now drops you on a full-page "Sign into Copilot" wall. No guest mode, no back door. Even the old Bing chat entry point just redirects you there.

0 of 10 questions answered. The experiment failed before question one. I'll be honest — I stared at the screen for a minute. You don't expect your very first experiment to produce literally nothing.

Why this matters more than it looks

I've been turning this one over, because it's not a bug — it's a structural shift. The assistants are moving to logged-in experiences, and that changes the game three ways:

1. Research gets harder. Casual observation and guest-mode scraping are dead. Serious measurement now needs authenticated access or paid APIs, which raises the bar for independent work like this. Good for me, bad for everyone. 2. Answers get personal. Logged-in assistants can tailor answers to your history and location. "What gets cited" stops being one answer and starts being many — which makes citation data more valuable, not less. 3. The moat deepens. If measuring citations takes infrastructure, the people with infrastructure have an edge over everyone just eyeballing it. I'm building the infrastructure.

What I'm doing instead

I'm moving the whole program to official APIs — OpenAI, Anthropic, Perplexity's Sonar, Gemini. A few dollars per hundred questions, fully within everyone's terms of service, and more scalable than browser testing ever was. The same 10 questions go through the APIs next, and Finding #002 will have the citation data this one was supposed to produce.

Seeing different guest-access behavior on these assistants? Reproducibility is part of the method — I'll update this post.