Recording what AI assistants say about your business

Why a screenshot is not a measurement

Someone in the business asks an assistant about the company, reads the answer, and forwards a screenshot. The answer is wrong about a fee or flattering about a competitor, and for a week it is the most discussed document in the office. Then it is forgotten, because nobody can say whether the answer is the same today.

An assistant composes each answer at the moment it is asked. It draws on what the model already holds and, for anything current, on pages it fetches while it answers. Change the pages and the answer changes. Ask on a different day, in a different phrasing, and the answer changes again. A single answer therefore cannot be compared with anything. A measurement needs the same questions, the same assistants, the same day each week, the full text kept, and the sources recorded beside it. Once those five things are fixed, each week's answers become a row in a series, and the series is what shows whether anything moved.

Writing the question set

The questions are the ones customers ask, in the words customers use. They ask what the service costs, whether the business is licensed, accredited or registered, which provider is closest to a suburb, whether a particular thing can be done at all, and how long it takes. A set in the low twenties covers one business line without the weekly read turning into noise. Include the questions where a competitor is the honest answer, because those are the ones that show what the assistant knows about the market and where the business sits in it.

The set stays fixed for months. A question changes only with a dated note, so that a shift in the record can be traced to the question and not mistaken for a shift in the answer. New questions join as a new row, with their own start date.

What to keep from each answer

Keep the full text, the list of sources the assistant cited, whether the business was named at all, and each claim it made about the business. Then check every claim against the business's own published pages and mark it as matching, differing or unverifiable. A claim that differs is the start of a finding. A claim that is unverifiable is often the more interesting case, because it usually means the business has never published the fact anywhere an assistant can read it.

What the record changes about the engineering

A wrong fee in an answer traces back through the cited sources. Often the business's own site is absent from that list and a directory or a review site took its place. Often the site is present and the fee is still wrong, because the fee schedule sits in a PDF, or in a tab that only appears after a script runs, or on a page that disagrees with another page on the same site. Each of those is a page problem, and page problems are engineering work: publish the fact as HTML text, mark it up with structured data, serve it to the agents that fetch it, and keep every copy of it in agreement. The pages on the sources assistants cite and machine-readable delivery cover the mechanisms in detail.

The fix is never a rewrite of what the business says about itself. It is the delivery of what the business already says, in a form an assistant can read and trust. That line is what keeps the weekly record an engineering instrument. The managed form of this work is AI Visibility Management, and the process for turning each row into a ticket is described in turning findings into engineering tickets.