Build a 24/7 Marketing Page Generation Engine

Marketing teams aiming for scale often hit a wall with traditional content methods. Hiring freelance writers or an in-house team is expensive and slow. The process of outlining, drafting, reviewing, and publishing limits output to a handful of pages per week, making it impossible to cover thousands of long-tail topics.

To accelerate this, teams adopt AI writing assistants. These tools are useful for drafting a single article but fail at systematic, scaled generation. They produce unstructured text, not a data object that can be programmatically verified. You cannot enforce a citation-ready intro, guarantee the presence of five 50+ word FAQ answers, or automatically generate valid Article and FAQPage JSON-LD schema. The output still requires a human to format, fact-check, and publish.

Trying to connect these tools with automation platforms also fails. You cannot build a feedback loop where a failed page is sent back for regeneration with specific instructions like "The Jaccard similarity score was 0.85, rewrite to be more unique." These platforms lack the granular control to run code for data validation, check for semantic HTML, or perform atomic database and cache operations on publish. They are designed for linear, human-in-the-loop tasks.

The structural problem is that these tools are designed for content creation, not content engineering. They treat each page as a one-off document. A true at-scale system treats content as a structured data asset that must pass a rigorous, multi-stage quality gate before it can be deployed to production as a live URL.

We started by designing a four-stage process to run autonomously. The system we built is a Python application scheduled by GitHub Actions, moving page candidates through each stage without manual intervention. The goal was to engineer a system that could be trusted to publish directly to our production site.

First is the Queue Builder, which scans sources like Google PAA, Reddit, and industry forums to discover page opportunities. Each opportunity is scored based on data availability, search intent signals, and competitive density. The highest-scoring items are added to a Supabase table for the next stage. Stage two, Generate, pulls from the queue and uses the Claude API at a low temperature (0.3) for factual consistency. It applies segment-specific templates that enforce a citation-ready structure, question-based headings, and proper schema formatting.

The most critical component is Stage three, Validate. This 8-check quality gate is the system's brain. It uses pgvector in Supabase to run a trigram Jaccard similarity check to ensure uniqueness (score < 0.72). It calls the Gemini Pro API to verify data accuracy and checks for content depth with a specificity score. Pages must score 88 or higher to pass. Failed pages are sent back to the Generate stage with specific feedback appended to the prompt for up to three retries. Passing pages move to the final stage, Publish, an atomic operation that flips a database status, invalidates the Vercel ISR cache, and submits to IndexNow. The entire process from draft to live takes under 2 seconds.