The method, in the open
An AI visibility evaluation you cannot argue with is an assertion, not a measurement. These pages are how it works, so you can check it, apply it yourself, and tell us where we are wrong.
The four measures
They fail independently, they have different fixes, and the order you address them in decides whether any of the work shows up.
Retrieval: can AI systems reach you at all? The cheapest failure to fix and the most frequently missed.
Comprehension: being described wrongly is a worse outcome than being unknown.
Corroboration: the largest block, and the one you cannot write yourself.
Selection: being named when it counts, and who is named instead.
Why all four have to be read together, and the most expensive misdiagnosis in the category.
Why the order matters
Each measure depends on the one before it. If an AI crawler cannot reach your pages, nothing written on them can be read, so rewriting your descriptions first produces no visible change. If the pages are read but describe the business vaguely, there is nothing clear for outside sources to confirm. And if nothing outside your own site confirms what you say, an assistant has little reason to choose you over a business it can verify.
In practice that means working from the bottom up. First confirm crawlers can fetch your pages and that no robots rule or firewall setting turns them away. Then make the homepage say plainly what the business is, where it operates and who it serves. Then build the outside record, such as directory listings, reviews and press, that repeats those facts in the same words. Selection is the result of the first three, which is why it is measured last and never worked on directly.
A common case: a firm adds service pages and answers to customer questions, sees no change in AI answers, and concludes that content does not work. The evaluation shows its pages return an error to AI crawlers because of a security rule. One setting changes, no new writing is needed, and the work already done starts to count.
What makes a measurement trustworthy
The two-run test: run any tool twice an hour apart. If the number moves, it was never measuring your business.
Not measured is not zero: absent, unreachable and measured-zero are three different findings.
How to compare AI visibility tools honestly: six questions, and they apply to us too.
Using it well
How often to re-evaluate, and why more frequently is not better.
What an AI visibility evaluation cannot tell you: the limits, stated plainly.