LuminaForge.ai — the site is the case study

LuminaForge launched with a single, uncompromising constraint: the website itself must be the proof. If we claim sub-800 ms LCP, our site has to hit it. If we claim full Generative Engine Optimization coverage, ChatGPT, Perplexity, Grok, and Claude have to be citing this very page within weeks.

This is how we built it.

The constraints

The stack

Next.js 16.2.6 — App Router, React Server Components, Partial Prerendering by default, Turbopack
React 19 — for streaming RSC and useEffectEvent
Tailwind CSS v4.3 — CSS-first @theme, no tailwind.config.ts
Motion (formerly Framer Motion) — premium animation with full reduced-motion support
Velite — type-safe MDX collections for work and resources
Vercel — Edge deploy, Speed Insights, Analytics, dynamic OG images

The numbers

LCP target

<800ms

Core Web Vitals

100/100

First-load JS

<90KB

CLS

<0.02

What makes it work

1. Every page is a Server Component by default

We treat "use client" like a sin tax. The home page hero contains exactly two client components — the particle canvas and the magnetic CTA wrapper — both leaf-level. Everything else streams from the server.

2. Partial Prerendering is automatic in Next.js 16

Static shells render instantly at the edge while any dynamic content streams in. No flag required — this is the default in App Router as of Next.js 16.

3. GEO is shipped at the infrastructure layer

llms.txt and llms-full.txt live at the root, structured to the emerging llms.txt standard.
robots.ts explicitly allows GPTBot, ClaudeBot, PerplexityBot, Google-Extended, Applebot-Extended, and 10+ other AI crawlers.
Every page emits structured data via a typed <StructuredData> component.

What we are tracking next

Weekly citation footprint across ChatGPT, Perplexity, Grok, and Claude for 24 target queries.
Real-user CWV from Vercel Speed Insights.
Indexing parity between Google Search Console and Bing Webmaster Tools.

The site is live. The forge is on.