Articles written for searchers and the AI that answers them

A separate request scores each article from 0 to 10 on eight criteria: it answers the question up front, it's specific to this business, every fact matches the site, nothing is invented, the links are real, an AI assistant could quote it, it reads well, and it covers new ground. The grader also returns concrete fixes that quote the problem text, and a draft under 8 is revised once automatically.

On the clinic article, the grader caught an unsourced claim that "research has found" virtual care works as well as in-person care, and a service the clinic doesn't offer. One revision took the score from 8.8 to 9.5.

People increasingly ask ChatGPT, Perplexity or Google's AI Overviews instead of clicking through results, and those answers quote pages. So the answer comes in the first sentence, headings are phrased the way people ask, every section opens with a sentence that stands on its own, and facts name the business and town instead of saying "we". On the gym site, the article's FAQ also becomes FAQPage structured data.

On the physical therapy site, every sentence about what a treatment does, who it helps or what the risks are gets copied word for word into a review list. One command prints that list as a numbered message for the clinic's DPT, and the tool won't publish live until a clinician's name is given.

It started as a hosted Next.js admin with Postgres, a login and sign-off links. For two sites and one editor, that was more to maintain than it was worth, so it became a command-line tool. Each article is a markdown file you can read and edit in any editor, kept in git.

Publishing converts the file to Sanity's Portable Text, refuses a URL that another article already uses, and keeps any photo added in the CMS. A signed webhook then refreshes only the pages that changed.