Teardown No. 91 · SEO

24 Pages of Keyword Stuffing: How Michael Kerkhoff's overnight SEO monitor caught and fixed the mess by 10 AM

Michael Kerkhoff runs a solo AI agency in Berlin called Context Studios. He built a content machine using OpenClaw, 23 cron jobs, and a Mac Mini that publishes 150+ blog posts in 4 languages. His overnight SEO audit caught keyword stuffing on 24 pages at 4 AM; the auto-healer fixed it by 10 AM while he ate breakfast. The system costs roughly €300 per month and took about 3 months to build.

Michael Kerkhoff · michaelker_kiVerified Nov 6, 2026Markdown copy

The problem

He describes himself as a solo founder with no team. Content marketing works, he writes, but the time cost broke his day: “I was spending 4 hours a day writing, editing, translating, posting.” He tried ChatGPT, Jasper, and Copy.ai but found the results “Generic. Every output sounded like every other company using the same tools.”

The blueprint (as he describes it)

  1. Agent runner and cron on a Mac Mini. OpenClaw schedules 23 cron jobs that run overnight audits, morning proposals, midday visuals, and afternoon engagement rounds.
  2. MCP server with 143 tools. Tools cover blog CRUD, CMS pages, social posting via Typefully, SEO checks, image generation, and the Cortex memory system. Transport is HTTP on Vercel with API-key auth.
  3. Content pipeline. Research topic → generate keywords → draft in 4 languages (DE, EN, FR, IT) linked by translationGroupId → generate hero image (blocking step) → publish → verify with curl -I → post to social → submit URL to Google Search Console.
  4. SEO and quality gates. A 4 AM cron runs NLP entity salience checks, keyword density scans, and GEO compliance. A 5:45 AM auto-healer fixes Tier 1 issues and proposes Tier 3 to him.
  5. Dead letter queue. Failed MCP calls are caught, logged, and retried with exponential backoff. Every write operation has a verify step; every update saves the original to a backup directory first.
  6. Cortex memory. Stores decisions and context across sessions in sensory, episodic, semantic, procedural, and prospective stores on Convex. A daily decay cron applies forgetting curves.
  7. Human review. He picks from 3 AI-proposed topics each morning. He reviews every blog post before publish, roughly 10-15 minutes per post. The agent cannot publish without his review, cannot send emails to leads, and cannot spend money without approval.

Stated results

  • 150+ blog posts published in 4 languages.
  • 200+ CMS landing pages.
  • 258 pages audited daily.
  • NLP average entity salience moved from 0.164 to 0.177.
  • Keyword stuffing detected on 24 pages at 4 AM, fixed by 10 AM. He writes, “I found out over breakfast.”
  • Monthly running cost: roughly €300 (Claude API ~€200, Vercel ~€40, Convex ~€25, image generation ~€10, other ~€25).
  • Development time: roughly 3 months.
  • Comparison cost he cites: 2 content writers at €8,000+/month or a content agency at €3,000-10,000/month.

Caveats

  • Self-reported figures; no independent verification.
  • Context Studios is his company; the site lists workshop pricing, so this post doubles as marketing.
  • DEV byline is “Michael K”; full name confirmed on contextstudios.ai/brand-facts.
  • Visual/carousel generation he says works “maybe 70% of the time - the other 30% need manual fixes.”
  • Keyword stuffing incident required him to “rewrite 40+ pages manually.”
  • LinkedIn engagement failures went unnoticed for “three weeks” until he built the dead letter queue.
  • A month of volume optimization (10-20 posts/day) “Traffic didn’t scale. Engagement dropped. Quality tanked.”
  • X: @michaelker_ki (verified 2026-11-06), company X: @_contextstudios (verified 2026-11-06), LinkedIn: /in/michael-kerkhoff (verified 2026-11-06).

FAQ

What stays manual in this system?

Topic selection from 3 AI-proposed options, quality review at roughly 10-15 minutes per post, strategic decisions, client work, and fixing the 30% of infographics that come out broken.

How does the overnight SEO audit work?

A cron job at 4 AM runs NLP entity salience, keyword density, schema, and hreflang checks across 258 pages. A follow-up auto-healer at 5:45 AM fixes Tier 1 issues and flags Tier 3 issues for his review.

What were the major failures?

Aggressive keyword targets produced pages that “read like SEO spam from 2010” and required rewriting 40+ pages manually. Silent MCP failures caused LinkedIn engagement to drop to zero for weeks. A month of optimizing for volume (10-20 posts/day) hurt quality and traffic.

How much does it cost to run?

He reports roughly €300 per month: Claude API ~€200, Vercel ~€40, Convex ~€25, image generation ~€10, other ~€25. OpenClaw is open source and runs locally at no cost.

Sources

Want agents like these working your distribution? Get It Built. Get new teardowns weekly: subscribe (double opt-in).

Get one teardown a week.

Same format, a new operator every issue. Double opt-in.