How I Built a Fully Automated SEO Blog for $0/Month (Cloudflare + Gemini API)

I run Nextrader — an AI-Powered Trading Bot Platform for Deriv trading. To drive organic traffic, I built a fully automated blog that publishes fresh, SEO-optimized articles twice a day without me touching a thing.

Here's exactly how it works — the tech stack, architecture decisions, and lessons learned.

The Goal

• Publish two trading-focused blog posts per day, fully automated.
• Each post includes a unique topic, SEO title, meta description, full article content, and an AI-generated cover image.
• Zero recurring cost — free tier everything.
• No CMS, no WordPress, no maintenance.

The Stack

The "Two-Call" Gemini Trick

The biggest challenge with Gemini 2.5 Flash is thinking tokens. When you ask for JSON and a full article in one prompt, the thinking tokens corrupt the JSON output.

The fix: split it into two calls.

1. Call 1 — Metadata only (small JSON): Returns the title, slug, meta description, tags, and an image prompt.
2. Call 2 — Article content (plain text): Writes the full 600–800 word article as plain text. No JSON wrapping.

This gave me 100% reliable JSON parsing with zero corruption.

Unique Topics Forever

The early version used getDayOfYear() — meaning topics would reset every January 1st and eventually repeat. I fixed this by switching to days since the Unix epoch. Today is day 20,546. Unique topics until the heat death of the universe.

AI Image Generation

Each post gets a unique cover image from Gemini's image API, with:

• 8 rotating color themes — deep blue, emerald, crimson, purple, teal, cyan, amber, rose.
• Topic-specific scene composition described by the metadata call.
• Varied layouts — close-ups, overhead desk shots, phone screens, laptops.
• WASM compression using @jsquash/jpeg — reduces 1.1 MB PNGs down to ~50 KB JPEGs.

SEO Built In

Every post automatically gets:

• Article JSON-LD structured data.
• Open Graph tags.
• Auto-generated sitemap.xml and robots.txt.
• Clean slug URLs with no /post/ prefix.

Lessons Learned

• Two-call LLM pattern: Never mix structured JSON and long-form text in a single prompt with thinking models.
• WASM in Cloudflare Workers: Pass the .wasm binary manually via init(), or it fails silently at runtime.
• Wrangler deploys wipe dashboard secrets: Keep all variables in wrangler.toml if the repo is private.
• Route order matters in Hono: /:slug catches everything, so always register it last.
• PNG vs. JPEG from Gemini: Detect magic bytes (0x89 0x50 = PNG, 0xFF 0xD8 = JPEG) to handle both formats.

The live blog auto-publishes at 08:00 and 20:00 UTC every day. No servers. No bills. Just Cloudflare Workers and Gemini doing the work.

Check out the live blog: blog.nextrader.live