matthieupesesse.com: a personal media platform that publishes, distributes and optimizes itself

matthieupesesse.com is not a blog with a CMS. It is an autonomous media platform: it writes, optimizes, distributes and monitors its own content — every day, without supervision. The goal was to build authority in the AI space without spending hours per week on content operations.

The platform stack

The public site is a Next.js 15 application running server-side rendered on a VPS (PM2 process: matthieupesesse-next, port 3055). nginx serves static assets (sitemap.xml, robots.txt, llms.txt, llms-full.txt) directly without hitting the Node process. Three separate PM2 processes run alongside it: the API backend (matthieupesesse-api), the asset integrity watcher (matthieupesesse-assets) and the autoblog engine (matthieupesesse-autoblog). All four are deployed from the same git repository and updated via a single GitHub push.

The autoblog engine: LLM pipeline with three-tier fallback

Every morning, the autoblog engine queries the Google Trends API for ten balanced topic categories (Anthropic, NVIDIA, Google AI, Apple, ElevenLabs, Suno, Tesla/SpaceX, Hugging Face, Intelligence Artificielle, AI général) and selects one topic for the day. It then runs a three-stage LLM pipeline: research and outline (Claude Sonnet 4), article generation (Claude Sonnet 4 primary → NVIDIA NIM meta/llama-3.3-70b secondary → MiniMax M3 tertiary), and editorial review against a rubric. If a provider is unavailable, the next tier activates automatically — publication never fails due to a provider outage.

Distribution: every channel, from minute one

On publication, IndexNow submits the URL to Google and Bing for immediate crawling. A newsletter digest is sent weekly. The article is cross-posted to LinkedIn and formatted for social. All structured data (Article, BreadcrumbList, WebPage in JSON-LD) is injected server-side with correct canonical URLs, hreflang and og: tags — verified monthly against a GEO/SEO audit checklist.

Self-improving loop: articles that underperform get rewritten

Every six days, a monitoring script checks each article's view count via the analytics API. Articles below 30 views trigger an automatic rewrite cycle: a new angle, a refreshed title and updated internal links. This means the content library improves autonomously over time — no editorial team required.

Measured results

GEO audit score before overhaul: 66/100. After consolidating structured data, canonical signals, hreflang, llms.txt and AI crawler permissions: 90+. GSC indexed URLs: 172 duplicate issues resolved in one deploy (dated slugs canonicalized to localized versions). Time spent on content operations per week: approximately zero. The platform has published every day since January 2026 with no missed publications.

Token cost vs. Belgian human equivalent

Running this platform with a traditional team would require: a content writer for 30 articles/month (€50–70/article freelance = €1,500–2,100/month), an SEO manager (€2,000–3,000/month), a developer for platform maintenance (€1,000/month part-time at €80/hour for ~12h), a newsletter and social distribution manager (€1,500–2,000/month). Conservative total: €6,000–8,100/month. The full AI stack: Claude Sonnet 4 primary generation at ~10,000 tokens/article × 30 = 300,000 tokens/month ≈ €2.50; NVIDIA NIM llama-3.3-70b fallback ≈ €1/month; MiniMax M3 tertiary ≈ €0.50/month; distribution via IndexNow, newsletter and social cross-posting adds zero marginal cost per article; VPS compute ≈ €3/month allocated. Total: under €15/month. At the conservative team estimate of €6,000/month, the cost ratio is 400:1. The 200+ articles published since January 2026 represent a permanent indexed asset library; an equivalent body of content commissioned from a Belgian content agency would cost upward of €120,000 to produce.