Content Autopilot — the €10,000 app Nutrimedia built for itself for $39

Nutrimedia runs a B2B media for marketing, R&D and regulatory decision-makers in the food-ingredients industry. Its standard: crossing marketing and science with verifiable primary sources — this readership spots approximation instantly.

The problem is classic for a small structure: sustaining a credible publishing rhythm takes an entire newsroom — daily monitoring, editorial arbitration, sourced writing, SEO optimisation, LinkedIn repurposing, performance tracking. Doing it all by hand does not scale; delegating it all to a generic AI produces inaccurate, voiceless content.

The project's thesis: it is not the writing that should be automated, it is the entire newsroom — keeping the human where they are irreplaceable: decision and validation.

Built in pair-programming with Claude Code: the human sets the course, the AI agent explores, implements, tests and deploys — every building block validated on real data the day it is written. Five weeks, 52 commits, from the foundation (auth, multi-source monitoring across RSS + PubMed + Perplexity, writing calibrated on the existing corpus, WordPress drafts, AI visuals) to market intelligence (DataForSEO volumes, printable audit reports, an action plan derived from the numbers).

Three architecture principles. Real data first: demand comes from Search Console, volumes from DataForSEO, audience from GA4, AI citability from real Perplexity probes — the model interprets, it never invents a market figure. AI under guardrails: every source-anchored article goes through a dedicated fact-checking pass, unverified links are downgraded by code, internal linking points to a closed catalogue of real URLs. The human decides: nothing ships to WordPress or LinkedIn without validation, and every automatic optimisation is preceded by a restorable snapshot.

The system then runs as a loop: a declining article becomes a "refresh" topic with the real queries to consolidate; a position 5–15 with demand becomes a quick win; a buyer question on which AIs do not cite nutrimedia.info becomes an article topic in one click; each audit's conclusions feed the next strategic analysis.

• Private web application (Next.js, Supabase, Claude API, Vercel) — 7 modules: audit, monitoring, strategy, editorial pipeline, LinkedIn, performance, costs
• Eight scheduled tasks — from Monday-morning monitoring to automatic drafting of the best topic and the weekly digest
• Full SEO editor: real RankMath score, humanisation, internal linking, fact-checking
• Generative-AI citability tracker (GEO), probed on real buyer questions
• Cost tracking: every model call logged (tokens, action, cost)
• ~15,000 lines of TypeScript, 88 automated tests

• Google impressions: 33,372 → 39,425 over compared 90-day windows (+18%).
• Citability by generative AIs: 6% → 13% (Perplexity audits across 16 buyer questions).
• Production: 83 qualified topics, 25 generated articles, 12 published — $0.61 generation cost per article, fact-checking included.
• Total project API cost: $39.16 for 153 model calls, all logged — less than a subscription to a single SEO tool, for an application whose equivalent development, billed by a contractor, would run around €10,000. Human time per article: review and decision.

The site is young: these numbers are an honest starting point, not a proclaimed victory. That is precisely the Autopilot's role — every monthly audit re-measures the same indicators, automatically.