Kāpiti Coast District Council

A method the council can run itself, not a report that ages the moment the models change.

"How do I pay my rates?" "Can I take my dog to Raumati Beach?" "Is my property at risk of flooding?" The assistant's answer has quietly become a new front door to council services, and it is one the council cannot see and does not control.

That is a hard problem for a council specifically. The highest-volume resident questions are transactional and often time-sensitive, and some are safety-related. If an assistant sends someone to the wrong page, an out-of-date form, or a page that no longer exists, a resident cannot finish a civic task, or worse, acts on wrong information. None of it shows up in the council's own analytics, because the exchange happens inside the assistant, not on the council website. Without a way to measure it, the council had no baseline and no view of which pages the engines pointed to.

Data Story treated this as a measurement problem first, and a capability problem second. Most agencies would run an audit and hand over a report. A report goes out of date as soon as the AI models change, which leaves the council no better able to keep up. So we gave the council a method it could run itself, so it can keep improving pages without needing us back.

The mechanism is a loop, not a one-off audit. Each resident question is held constant so a later check is comparable to an earlier one, and for every answer we record not just whether the council is mentioned but which page the assistant points to and whether that page still works.

The council took this and made it their own: the content team built custom AI assistants to restructure and fact-check a page, adapted the scoring to a ten-point scale with a breakdown, and now test each page's before-and-after across several assistants. What keeps working is the plain approach: clear question-shaped headings, short sections and direct answers. That helps a resident read the page, and it helps an AI read it too.

The council is no longer dependent on an agency to keep up with AI search. It runs its own measure-and-improve loop, and the first pages through it show the method works.

The measurement also surfaced the next fix. When we checked every council page the assistants cite, most of the "broken" links turned out to be paths the AI had invented, not pages the council had lost: the same rates page guessed under a dozen plausible addresses. So the job is not chasing dead links. It is making one real page per task the clear, canonical one, which the council's page rebuilds are already doing.

For public-service content, it is no longer only about whether a resident can find the page. It is whether an AI can read it and answer correctly. The same plain edits help both. And a method the council can run itself is worth more than a one-off audit: it keeps improving pages over time, and it keeps up as the AI models change.