Feature deep-dive
The 11-stage drafting pipeline
Last updated
Mentionwell isn't a one-shot generator. Every article passes through 11 distinct stages — each with its own model strategy, prompt, and cost ledger entry. You can watch the pipeline run live in the activity feed; failed stages retry with a different model before surfacing in the kanban.
The 11 stages
- Research — live SERP + competitor-page web research with a transparent source list on every draft.
- Synthesize — extract key insights, AEO signals, top questions, and the unique angle.
- Outline — search-intent-aware section plan; H2s become canonical TOC entries.
- Draft — section-by-section writer with grounded citations.
- Reflection — editorial critic graded on accuracy, AEO readiness, factual grounding, brand fit.
- Rewrite — automatic rewrite when the reflection critic scores below 0.8 confidence.
- Metadata — meta title, description, excerpt, tags, category — written for CTR + search-intent match.
- FAQ — frequently-asked-questions block + JSON-LD FAQPage schema.
- Embed — vector embedding for dedupe + internal-link discovery on the next draft.
- Images — featured + inline images at 4K, generated against per-site visual identity.
- Render — final HTML + Markdown + JSON-LD ready for headless consumption.
Per-stage routing
Each stage routes to its own model. Planning + outline use a fast cost-efficient model; drafting uses a frontier model; the reflection critic uses a different frontier model so “writer” and “critic” aren’t the same instance grading itself. Image stages route to FAL or OpenAI image depending on the per-site style brief.
Per-stage cost ledger
Every stage emits an entry in the cost console: tokens in, tokens out, USD spent, model used. You can see exactly where the budget went on every article.