Methodology

Here's exactly how the sausage gets made. If you think I'm doing it wrong, the findings posts include enough detail to check my work.

The basic unit: one experiment

An experiment is a fixed set of questions, asked the same way, across one or more AI assistants, with every cited source written down. A question set might be "best plumber in Austin TX" style local queries or "best CRM for a small real estate team" style software queries. The questions don't change between runs. That's the whole discipline.

What I record

• The exact question, word for word
• Which assistant, and the date
• Every domain cited, in order
• Whether the sources are businesses, directories, publishers, or forums
• Whether I got an answer at all — some assistants now hide behind sign-in walls

What I test

I change one variable at a time. A page's opening paragraph. FAQ structure. Review counts. Schema markup. Brand mentions on third-party sites. Then I re-run the questions and compare citation rates. One variable. That's the rule — break it and you're just guessing.

What I don't do

• I don't scrape assistants in ways that break their terms. The scaled testing runs through official APIs.
• I don't sell placement. Findings are findings — nobody can pay to look good here.
• I don't claim permanent truths. Models change monthly. Every finding carries a test date, and I re-run the important ones.

Reproducibility

Every finding post lists its questions and test date so you can sanity-check me. If you re-run one of my experiments and get a different result, I genuinely want to know — that's how this gets better.