Case studies — AI TOOL · SEO AGENT
Autonomous SEO that writes, reflects, and publishes.
7-stage agentic pipeline. GEO engine tracking 10 AI search engines. Long-form content at scale.
AT A GLANCE
Research → Outline → Draft → Reflect → Refine → Image → Publish
AI search engines tracked, ChatGPT, Perplexity, Claude, Gemini, and more
Isolated brand profiles, keyword targets, tone guidelines per client
The pipeline.
Every article produced by MentionWell passes through a seven-stage autonomous pipeline, no human bottleneck, no copy-paste, no thin content.
Serper pulls live SERP data for target keywords. Firecrawl scrapes top-ranking pages for structure and coverage signals. An outline agent synthesizes a content brief with heading hierarchy and target word counts per section.
A draft model (via OpenRouter) writes the full article against the brief. A second model pass, the Reflect stage, reviews the draft for factual gaps, thin sections, and missing search intent coverage. It produces a structured critique that feeds directly into Stage 05.
The refine agent rewrites flagged sections. FAL generates a unique header image matched to the article topic and brand palette. A CMS publishing agent delivers the final article, with SEO metadata, slug, alt text, and structured data, directly to Webflow or client API.
Proprietary module that queries 10 AI search engines, ChatGPT, Perplexity, Claude, Gemini, Copilot, and more, with branded queries on a schedule. Measures citation frequency, sentiment, and share-of-voice across AI platforms over time.
Each client runs in a fully isolated profile: brand voice guidelines, keyword targets, competitor watchlists, tone instructions, publishing credentials, and CMS config. One platform, zero cross-contamination.
Real-time article status tracker across all 7 stages. Keyword ranking history. GEO citation score over time per AI engine. Scheduled publish queue with override controls.
The reflect step.
The single most impactful stage in the pipeline. After the first draft is written, a second model is invoked with the draft, the original brief, and a critique rubric. It returns a structured JSON object identifying: → Sections with factual claims that need verification → Headings where search intent is underserved → Word count deficits by section vs. SERP average → Missing semantic keywords identified from competitor coverage → Tone deviations from brand guidelines
OpenRouter routing.
Pipeline stages run on different models via OpenRouter, heavier models for Draft and Reflect, faster models for Outline and metadata generation. Routing rules are per-client and configurable without code changes.
Trigger.dev — Job queue + retry logic Supabase — Client profiles + article state OpenRouter — Multi-model routing FAL — Image generation
Technology.
On the bench.
- Next.js
- OpenRouter
- FAL
- Firecrawl
- Serper
- Supabase
- Trigger.dev
- Webflow API + Custom