Your Next.js Site Is Serving 26 KB of Noise to LLMs. Here's the Fix.

Kacper Siniło — Tue, 03 Mar 2026 20:51:52 +0000

LLMs are crawling your Next.js site right now.

They're downloading your full HTML page — RSC payloads, hydration scripts, font preloads, inline styles, the works — just to pull out a product title, a price, and a description. 26 KB parsed. 101 bytes kept.

That's not a failure of the LLM. It's a failure of the server.

The Problem No One Talks About

HTTP solved this decades ago. It's called content negotiation. The client sends an Accept header telling the server what format it wants. The server responds accordingly.

Browser:   Accept: text/html
LLM agent: Accept: text/markdown
API client: Accept: application/json

Backend developers do this routinely. Express, Django, Rails — they all support it out of the box.

Next.js doesn't.

So you're stuck with two bad options:

Separate endpoints like /api/products/123.md — duplicates your routing, drifts out of sync, and forces clients to know about a non-standard URL scheme.
Markdown-only pages — breaks the browser experience for humans.

Content Negotiation, The Way HTTP Intended

I built next-md-negotiate to close this gap.

Same URL. Same route. Different response based on what the client actually wants.

Browser   → GET /products/42  Accept: text/html     → Normal Next.js page
LLM agent → GET /products/42  Accept: text/markdown → Clean Markdown

No new URLs. No duplicate routing. The client just sets a header.

How It Works Under the Hood

The library hooks into Next.js at the rewrite layer (or middleware, your choice). When a request comes in with Accept: text/markdown:

                     Accept: text/markdown?
                            │
                  ┌─────────┴─────────┐
                  │ yes               │ no
                  ▼                   ▼
          Route matches?        Normal Next.js
                  │              page renders
           ┌──────┴──────┐
           │ yes         │ no
           ▼             ▼
    Rewrite to        Pass through
   /md-api/...
           │
           ▼
  Catch-all handler
  runs your function
           │
           ▼
 200 text/markdown

It checks for text/markdown, application/markdown, or text/x-markdown in the Accept header. If a configured route matches, the request is internally rewritten to a catch-all API handler that calls your Markdown function.

Your browser users never see any of this. They get the same HTML they always did.

Setup in 3 Steps

1. Install

npm install next-md-negotiate

Or scaffold everything automatically:

npx next-md-negotiate init

2. Define Your Markdown Versions

// md.config.ts
import { createMdVersion } from 'next-md-negotiate';

export const mdConfig = [
  createMdVersion('/products/[productId]', async ({ productId }) => {
    const product = await db.products.find(productId);
    return `# ${product.name}\n\nPrice: $${product.price}\n\n${product.description}`;
  }),

  createMdVersion('/blog/[slug]', async ({ slug }) => {
    const post = await db.posts.find(slug);
    return `# ${post.title}\n\n${post.content}`;
  }),
];

Parameters are type-safe — { productId } is inferred directly from the [productId] in the pattern.

3. Wire Up Rewrites

// next.config.ts
import { createRewritesFromConfig } from 'next-md-negotiate';
import { mdConfig } from './md.config';

export default {
  async rewrites() {
    return {
      beforeFiles: createRewritesFromConfig(mdConfig),
    };
  },
};

And create the catch-all handler:

// app/md-api/[...path]/route.ts  (App Router)
import { createMdHandler } from 'next-md-negotiate';
import { mdConfig } from '@/md.config';

export const GET = createMdHandler(mdConfig);

That's it. Three files, zero config duplication.

Why This Matters

257x smaller payloads. That's the ratio between a typical Next.js HTML response and the equivalent Markdown for a simple product page.

For LLMs, this means:

Fewer tokens consumed — you're not burning context window on script tags and hydration data
Better extraction accuracy — no parsing HTML soup to find the three fields that matter
Faster responses — less data over the wire, less processing on the model side

For you, it means:

Single source of truth — one URL, one routing layer, multiple representations
No drift — your Markdown definitions live next to your page definitions
Standard HTTP — any client that sets an Accept header gets the right format

Middleware Alternative

If you already have a middleware.ts handling auth, i18n, or redirects, you can integrate content negotiation there instead of using rewrites:

// middleware.ts
import { createNegotiatorFromConfig } from 'next-md-negotiate';
import { mdConfig } from './md.config';

const md = createNegotiatorFromConfig(mdConfig);

export function middleware(request: Request) {
  const mdResponse = md(request);
  if (mdResponse) return mdResponse;

  // ...your other middleware logic
}

Same config, same single source of truth. Just a different integration point.

Try It

# Normal HTML
curl http://localhost:3000/products/42

# Markdown for LLMs
curl -H "Accept: text/markdown" http://localhost:3000/products/42

Two commands to see the difference.

The repo is MIT-licensed and works with both App Router and Pages Router.

GitHub: github.com/kasin-it/next-md-negotiate

npm install next-md-negotiate
npx next-md-negotiate init

If you're building anything that LLMs or AI agents interact with, HTTP already has the answer. We just need Next.js to speak the language.

Forem: Kacper Siniło