Phase 5 — SEO/GEO blog engine

[AirKyzzZ]

Roadmap placeholder, pre-brainstorm. Phase 5 builds the organic-discovery moat: a long-tail blog tuned for student-job intent queries in FR (then EN/ES/DE).

Likely scope:

• MDX-based blog under /{locale}/blog
• Editorial templates respecting the brand charter (Instrument Serif headings, cream/ink palette)
• Programmatic SEO pages for high-intent queries (e.g. 'comment décrocher un stage chez {company}', 'lettre de motivation {role}')
• GEO-tuned content for AI search engines (Perplexity, Claude, ChatGPT search citing the site)
• Sitemap + RSS + structured data
• Optional: AI-assisted content generation pipeline gated by editorial review

Workflow: brainstorm → spec → plan → implement.

Hard deadline: June 2026 v1 launch — lowest priority, can slip past launch if pre-launch traffic comes from other channels.

[AirKyzzZ]

Brainstorm done 2026-06-05 (no code yet, awaiting confirmation)

Ran the brainstorm. Decisions so far:

• Storage: MDX in-repo (web/content/blog/), git-versioned, static-rendered.
• Content: editorial long-form, answer-first, GEO-citation-shaped, FR-first like /outils.
• Cadence: fully-automatic 1/day, hands-off AI generation, no human review gate.
• Scope this session: engine + 3-5 seed posts.

Key reconciliation: MDX compiles at build time, so the daily pipeline can't just write a DB row, it has to commit a new .mdx via the GitHub API and ping a Vercel Deploy Hook to rebuild. That folds the generation pipeline into the build (more than "engine + seed posts" alone).

Proposed architecture, 5 parts:

1. Rendering chassis /{locale}/blog + /blog/{slug}, MDX + frontmatter, editorial template inheriting the /outils visual language.
2. SEO/GEO: BlogPosting + FAQPage JSON-LD, canonical/OG metadata, FR-index / non-FR noindex, sitemap.ts extension (already has the hook) + RSS.
3. Funnel: in-content links to the matching /outils tool + end-of-post signup CTA, PostHog blog_* events.
4. Auto-pipeline: Vercel Cron daily, CRON_SECRET-guarded route, next topic from an in-repo queue, OpenRouter Sonnet generation with a charter-encoded prompt, strict validators that reject and retry on any violation (em-dash, "piston", banned jargon, too-short, missing structure, missing internal link; skip the day rather than ship junk), GitHub-commit the .mdx, Deploy Hook. Idempotent.
5. Seed: run the pipeline to draft 3-5 posts, polish to charter, commit. Seeds the blog and deep-tests the generator before it runs unattended.

Honest risk: unreviewed AI content committed to main and deployed daily is a real brand/SEO risk. Mitigation is the strict validator gate. Safety valve if a gate is ever wanted: open a PR instead of committing to main.

Awaiting confirmation on three points before the spec:

• a. fold the auto-pipeline into this session (the scope bump)
• b. fully-auto commit-to-main + strict validators vs a PR gate
• c. Sonnet for generation

New manual-ops secrets needed: GitHub fine-grained token (contents:write on this repo) + a Vercel Deploy Hook URL. OPENROUTER_API_KEY and CRON_SECRET already exist in prod.