Notes on the data behind search and AI answers: how answer engines retrieve, what makes them cite a page, and how to measure any of it. Written for the people doing the work.
There are two ways to end up in an AI answer, and they need different work on different timescales. Which one is failing, what the evidence says actually moves each, and what you are being sold that does nothing.
We put five questions about measurement to the published documentation of seven AI visibility tools. Three collect answers in ways that are not comparable, one uses a visibility score that shares nothing but its name, and the gap between the lightest and heaviest sampling is a hundred to one.
Run the same brand through two trackers and you get two answers. Neither vendor is lying. There is no single number underneath for them to be right about, and once you see why, you know what to ask before you buy.
A model that has never heard of you will not go looking for you. That is a different failure from ranking eleventh, and almost nobody measures it.
Most claims about what AI engines reward have never been tested against a control, and untestable claims are exactly the kind that spread. A worked example of how to read one properly, and four questions that separate a method from a story.
Answer engines do not remember your brand, they search for it. Understanding the queries they actually run turns AI visibility from a mystery into a ranking problem you already know how to work on.