Introducing LLMR: RSS for the AI Era

~8 min read AI Web Standards Innovation

The Idea

Jibberish mode but for LLMs when browsing websites.

When AI systems browse the web today, they're forced to parse HTML designed for humans. They wade through navigation bars, CSS classes, footer links, and advertising markup just to extract the actual content.

For my technical blog with 10 posts, an AI must process ~32,000 tokens to understand my content. At GPT-4 pricing, that's $0.96 per query. If 1,000 AI agents visit per month, I'm looking at nearly $1,000 in indirect crawling costs - costs ultimately passed on through API pricing.

There has to be a better way.

The Thought: LLMR Format

I built a proof-of-concept format called LLMR (LLM-Readable) that compresses an entire website into a single, structured JSON file optimized for AI consumption.

Think of it as RSS for AI systems or Jibberish mode for LLMs.

What Does It Look Like?

{
  "v": "1.0",
  "s": {
    "d": "raphaelreck.com",
    "t": "prof_tech_blog",
    "a": {
      "n": "Raphaël Reck",
      "exp": ["drupal", "laravel", "sys_arch"]
    }
  },
  "p": [
    {
      "id": "juggling-fireballs",
      "u": "/blog/2025-10-08-juggling-fireballs.html",
      "d": "2025-10-08",
      "tg": ["prod", "burn", "solo"],
      "rt": 4,
      "tc": 0,
      "emb": [0.26, 0.77, -0.21, -0.67, ...],
      "sum": "Managing heavy workload when partner experiences burnout"
    }
  ]
}

Key features:

  • Compressed keys: d = date, tg = tags, tc = technical content
  • Embeddings: Pre-computed semantic vectors for instant understanding
  • Metadata rich: Technical depth, code block count, read time
  • Single file: Entire site in one HTTP request

The Results

For my 10-post blog:

  • Traditional HTML: 32,000 tokens
  • LLMR format: 1,200 tokens
  • Reduction: 96%

If your site gets 1,000 AI queries per month:

  • Traditional cost: $960/month
  • LLMR cost: $36/month
  • Savings: $924/month (96%)

How It Works

I wrote a Python script that:

  1. Scans all HTML files in your website
  2. Extracts metadata (title, description, tags, dates)
  3. Detects technical content (code blocks, debugging stories)
  4. Generates embeddings (semantic fingerprints)
  5. Creates compressed JSON with abbreviated keys

Add it to your blog publishing workflow:

# After generating RSS/Atom feeds: subprocess.run(['python3', 'generate_llmr.py', '.'])

Done. Your site now speaks AI.

Why This Matters

We already have standards for different consumers:

  • robots.txt - for web crawlers
  • sitemap.xml - for search engines
  • RSS/Atom - for feed readers
  • manifest.json - for web apps

We need:

  • site.llmr - for AI systems

Real-World Benefits

For content creators:

  • AI understands your content better
  • Lower bandwidth costs
  • More accurate AI summaries of your work

For AI systems:

  • 95%+ token reduction
  • Instant semantic understanding
  • No HTML parsing overhead

For users:

  • Faster AI responses about web content
  • More accurate information
  • Lower API costs

The Vision: A Consortium

For this to become mainstream, we need:

RSS started with one person (Dave Winer) saying "there should be a better way." This could be the RSS of the AI era.

Try It Yourself

I've open-sourced the generator:

GitHub: github.com/djassoRaph/open_llmr

Features:

  • Zero dependencies (pure Python)
  • Works with any HTML static site
  • Integrates with existing workflows
  • Generates hash-based embeddings (production should use sentence-transformers)

Usage:

python3 generate_llmr.py /path/to/your/website

Add this to your HTML <head>:

<link rel="llm-index" type="application/json" href="/site.llmr">

The Bigger Picture

We're entering an era where AI agents will browse the web as much as humans do. They'll:

  • Research topics across hundreds of sites
  • Summarize content for users
  • Build knowledge graphs
  • Answer questions with citations

If we design for this future now - with formats like LLMR - we can:

  • Reduce global token consumption
  • Lower AI infrastructure costs
  • Enable better AI understanding
  • Maintain human-readable HTML alongside

This isn't about replacing HTML. It's about working with it with AI-optimized metadata, the same way we augmented it with RSS for feed readers 20 years ago. Or how Jibberish mode does it's own thing.

Try it out! :)

If you're:

  • A developer - try the script on your blog
  • An AI company - consider supporting this format
  • A CMS developer - build a plugin
    • (Let me know if you need help! I'll also share it!)
  • A standards body - let's formalize this

Let's build LLM-First Web Architecture together.

"The best way to predict the future is to invent it. But the smartest way to invent it is to make sure everyone can read it, including... machines."

What are your thoughts? Would you implement this on your site? Hit me up on LinkedIn or GitHub to discuss.

Thank you for reading.
No tokens were wasted in the making of this post.