Content Generator
How GrowDeck generates research-backed landing pages.
Last updated: May 2025
Overview
The content generator doesnβt just write β it researches first. Every page starts with a research phase that queries live data before a single word is written.
The pipeline
Phase 1: Research
Before generation, the Researcher agent:
- Scrapes the top 5 ranking pages for the target keyword
- Extracts headings, angles, and content gaps from competitors
- Fetches People Also Ask questions from Google
- Finds a relevant YouTube video to embed
- Analyzes what competitors are missing
This research bundle is passed to the generator as context.
Research takes 20β40 seconds
The research phase makes live SERP requests. This is intentional β grounded research produces significantly better content than generation from static prompts.
Phase 2: Generation (GPT-4)
The generator creates a structured page with these section types:
| Section type | Description |
|---|---|
hero | Title, subtitle, and key value proposition |
intro | Context-setting opening paragraph |
features | Feature list with descriptions |
comparison | Side-by-side comparison table |
how-it-works | Numbered step-by-step section |
code | Code example with syntax highlighting |
video | Embedded YouTube video with context |
faq | Frequently asked questions |
cta | Call-to-action section |
Phase 3: Quality scoring
The generated page is scored (0β1) on:
- Coverage depth (covers the topic adequately)
- Structural completeness (has hero, body, CTA)
- Keyword integration (target keyword appears naturally)
- Section diversity (not all text, includes tables/code)
Threshold: Pages scoring below 0.75 are regenerated automatically.
Quality threshold: 0.75
Pages scoring below 0.75 are automatically regenerated once. If the second attempt also fails quality checks, the opportunity is flagged for manual review.
Phase 4: HTML rendering
The approved structure is rendered to a full standalone HTML document with:
- Tailwind CSS (via CDN)
- Embedded styles for code blocks
- Responsive layout
- Schema markup (Article or FAQPage)
What gets generated
A typical page includes:
- 1,200β2,500 words of structured content
- At least one comparison table
- One code example (if relevant)
- One YouTube embed (if a relevant video is found)
- FAQ section with 4β6 questions
- Schema markup (JSON-LD)
Quality controls
- Research phase ensures content is grounded in real SERP data
- Quality threshold: 0.75 (auto-retry if below)
- All pages require human approval before deploying unless auto-approve PRO is enabled.
