Claude builds and runs 10 data-driven niche websites with weekly newsletters, with ChatGPT as the reviewer

Launching and maintaining a content site with a newsletter normally takes days of research, building, domain and analytics setup, and ongoing editing. The author wanted to see how much of that Claude could do with almost no human involvement, across many small sites at once.

1. ChatGPT produced a list of 50 possible site topics. The first site (a western US water-conditions tracker) was built with close human feedback and became the template. 2. Claude was then told to build the next nine sites with parallel agents. Running all nine at once froze the computer, so it was cut back to three at a time. 3. For each site: take the next unused topic; search for usable information, especially free government data, while ChatGPT does the same research in parallel through Claude in Chrome; compare both outputs and plan the data sources and their update times, the site structure and any evergreen articles. 4. Claude builds the site, finds available domains and presents them; the human picks one; Claude buys it and launches the site (Cloudflare's agent-friendly tools made this possible). 5. Claude creates Google Analytics 4 and Search Console properties. 6. ChatGPT reviews the site for data errors, tone and audience fit, and layout problems; Claude applies the fixes it agrees with. The author calls this review step critical because Claude often left inconsistent numbers across pages. 7. Claude configures a Buttondown newsletter and sets recurring jobs to pull fresh data, update the site and send the newsletter the day after new data is released. 8. Improvement loop: a weekly email report shows traffic and search appearance per site. At first the author picked improvements from the search queries; then Claude proposed its own list each week for review. Feedback on bad ideas was saved to an instruction file for the reviewer. After three to four weeks the suggestions were consistently good, so the whole loop was automated. 9. Human steps that remained: logos, creating the Buttondown accounts and approving domains. The author planned to add a weekly ChatGPT scheduled task that audits all sites by email, and to do backlink outreach by hand because he does not want AI sending outreach emails.

Ten sites are live and update and improve themselves weekly. The author reports a trickle of traffic, with visits and search impressions rising on most sites, and says he does not know whether it will become a real business. No traffic numbers, subscriber counts, costs or revenue are given. He found live errors while writing, including a homepage section that would never populate because Claude reused a pattern from a shared engine.

Self-reported by one operator; no traffic figures, subscriber numbers, costs or revenue are shown. Prompts, project instructions and code are not shared, so the setup cannot be followed exactly. Claude repeatedly bent site quality to fit a shared engine despite instructions, and made audience mistakes (for example proposing a consumer propane calculator for a trade audience). Backlinks, which the author says the sites need most, are not automated. The author runs physical-product e-commerce brands; these sites are an experiment, not his main business. Written about two months into the project.

Claude with browser access (Claude in Chrome), a ChatGPT account, a Cloudflare account for domains and hosting, Google Analytics and Search Console, a newsletter service, and topics with free public data.

Start with one site built hands-on to create the template, then scale. Run only a few build agents at a time.

• The Automated Operator (Alex Willen), Substack ↗ Firsthand write-up, published August 13, 2026

Source published: August 13, 2026. Last reviewed here: November 3, 2026. Spot a mistake? Tell us.