The problem

Our domain is about 10 years old, but until this fall it had almost no real content. Google showed it about 3 times a day, nearly always to people already searching for our name. ChatGPT and Claude did not mention us at all.

We wanted two things: B2B leads from Google and Bing, and to be a source AI answers cite. We did not want a content mill.

What we built

A weekly loop that treats the website like a process with a scorecard. Each run reads what happened, decides one thing to do, does it and logs the reason.

It runs on a schedule with no one at the keyboard. The rules it follows are written down in the repository it works from, so every run starts with the same instructions and the same history.

How the loop works

  1. Read the history first. The article log, the keyword backlog and the AI visibility log, so it never repeats a topic or undoes past work.
  2. Look for real material. It reads the past week of the owner's sent email and calendar for technical work worth teaching from, and skips anything about clients, deals or personal matters.
  3. Read the scorecard. Search Console shows which queries the site already appears for, and at what position. Anything with impressions at position 8 or worse is a candidate.
  4. Check demand in both channels. Google search volume, plus how often the same question is asked of AI assistants.
  5. Check the competition. The top 10 Google results, and the answer ChatGPT and Claude give to the same question, including which sources they cite.
  6. Pick one action. A new article, or a refresh of a page that is close to page one. Refreshes are capped at one in every three runs so the site keeps growing.
  7. Write in the owner's voice, from a style guide built from his own sent mail, with a direct answer up top that an AI can quote.
  8. Check before shipping. No banned punctuation, no client names, word count in range, and a build audit that fails if any page points at a file that does not exist.
  9. Deploy, then prove it. Every file on the live site has to return a 200, the sitemap is resubmitted to Google and Bing, and the log gets a line explaining what was done and why.

Once a month it also runs a fixed set of eight buyer questions through ChatGPT and Claude and records whether we are mentioned and who is cited instead. That list does not change, so the months compare.

What we got wrong

We shipped broken images. The scheduled run built the site from a fresh copy of the repository, and some image files had never been committed because of an ignore rule. Everything looked fine locally and was broken in production.

The fix was two checks, not a reminder. The build now fails if any page or stylesheet references a file that is missing, and every deploy ends by requesting every file on the live site and stopping if any of them fails.

What changed

Google impressions went from about 3 a day to roughly 150 a day after the content was indexed in late October. Clicks are still small, because most pages sit on pages two to four while Google tests them. Four guides have already reached positions 6 to 10.

The first AI check found Shiptronix mentioned in 0 of 8 answers. That same check told us why: the answers cite directories and pages that state a clear focus, so we added both to the plan. We judge this on a 3 to 6 month horizon, not a few weeks.

What we would do again

Our take: the loop is the product. A system that reads its own scorecard every week finds the pages that are almost winning. One that posts on a schedule never does.

Make the checks fail loudly. Every mistake we caught became an automatic check, so it cannot happen twice.

Measure AI answers the same way you measure Google. Ask the same questions every month and write down who gets cited. That list is your roadmap.