How to reduce LLM token cost with cleaner Markdown

If you use ChatGPT, Claude, Cursor, or another AI tool for research, the easiest way to waste tokens is to paste a web page straight from the browser.

I tested this the boring way: same article, same prompt, three inputs.

First, I copied the page visually from Chrome and pasted it into an AI chat. Second, I used a server-side reader style tool. Third, I converted the page to Markdown in the browser with Web2MD and pasted that.

The answer quality was not just about word count. The biggest difference came from noise: navigation links, cookie banners, social widgets, repeated sidebars, hidden labels, tracking text, and weird spacing. The model had to read all of that before it got to the page content.

Cleaner Markdown reduced the amount of text I sent, but it also made the prompt easier to reason about. Headings stayed as headings. Lists stayed as lists. Code blocks stayed readable. That matters when you pay by token or when your context window is already crowded.

Where token waste comes from

Most web pages are not written for LLMs. They are written for browsers.

A browser can hide layout junk with CSS. An LLM cannot. If you paste from the rendered page, you often send text like this along with the article:

• Header links
• Footer links
• Newsletter forms
• Cookie notices
• Related article cards
• Repeated menu labels
• Image captions with poor spacing
• Buttons like "Share" and "Copy link"
• Comments or ads mixed into the main text

Even worse, some copy-paste output loses structure. A heading becomes plain text. A table becomes a pile of columns. Code loses indentation. The model spends tokens guessing what the source meant.

Here is a small example of the kind of Markdown output I want before I send a page to an LLM:

# Pricing notes from the vendor docs

The free plan includes 3 projects and 100 monthly runs.

## Limits

| Plan | Monthly runs | Seats |
| --- | ---: | ---: |
| Free | 100 | 1 |
| Team | 10,000 | 5 |

## Important detail

Overage billing starts after the monthly run limit is reached.

That is compact, structured, and easy to quote in a prompt. I can ask an AI tool to summarize it, compare it, or turn it into a checklist without first asking it to clean up the source.

Why Markdown cuts token cost

Markdown helps because it preserves meaning with fewer characters.

A heading like ## Limits is cheaper and clearer than a line surrounded by spacing, font artifacts, and copied menu text. A Markdown table gives the model row and column boundaries. A bullet list tells the model the items belong together.

The token savings vary by site. On a simple blog, the difference may be modest. On a product docs page with side navigation, footer links, and code examples, I have seen the usable content become much smaller after conversion. The more cluttered the page, the more Markdown helps.

There is also a second cost: follow-up prompts. If the first answer misses the point because the source was messy, you pay again to correct it. Clean input reduces those repair turns.

Where Web2MD fits

Web2MD is a Chrome extension that converts the current web page to clean Markdown for AI tools. The main difference is that it runs in your browser.

That sounds like a small implementation detail, but it changes what pages it can handle.

A server-side reader has to fetch the URL from its own servers. That works well for public pages. It often fails for pages behind a login, internal dashboards, private docs, paid subscriptions, or anything that depends on your browser session.

Web2MD reads the page you already have open in Chrome. If you are logged in and can view the page, Web2MD can work with that page locally in the browser. That is useful for research notes, private documentation, support portals, course pages, and paywalled articles you already have legitimate access to.

Web2MD also includes a token counter, which I found more useful than I expected. Before sending the result to ChatGPT, Claude, or Cursor, you can see roughly how much context you are about to spend. That turns token reduction from a vague concern into a quick check.

If you want a broader walkthrough, see our guide on converting web pages to Markdown for AI. If you are mostly comparing reader tools, the Jina Reader alternative notes may also help.

A practical workflow I use

My normal flow is simple:

1. Open the page in Chrome.
2. Run Web2MD.
3. Check the token count.
4. Remove sections I do not need.
5. Send the Markdown to ChatGPT, Claude, or Cursor.

The one-click send-to-AI option is convenient when I know I want the whole page. If I am dealing with a long page, I usually copy the Markdown first and trim it.

For example, I do not need a vendor's whole documentation page if my question is only about rate limits. I keep the relevant heading, table, and any warnings. Then I ask the model a narrow question.

# API rate limits

Requests are limited per workspace, not per user.

## Current limits

- Free: 60 requests per minute
- Pro: 600 requests per minute
- Enterprise: custom limit

## Retry behavior

If a request is rate limited, the API returns status code 429.
Clients should wait before retrying.

That small block is cheaper than the whole docs page, and the model has less room to get distracted.

Competitor notes: Jina Reader, Firecrawl, Turndown, and MarkDownload

Jina Reader is strong for public pages. I use tools like it when I want a quick Markdown-like view of a URL without installing anything. The limitation is access. If the page requires my browser session, a server-side fetcher usually cannot see it.

Firecrawl is better for crawling, scraping, and developer workflows. If you need an API, batch extraction, or site-wide ingestion, Firecrawl is built for that. It is not the same job as a local browser extension. It also adds API keys, billing, and a server-side processing path.

Turndown is a useful JavaScript library for converting HTML to Markdown. Developers can build great pipelines with it. But Turndown is not a ready Chrome workflow by itself. You still need extraction logic, browser integration, token counting, and a way to send the result to AI tools.

MarkDownload is a solid browser extension for saving pages as Markdown. It is useful if your goal is clipping pages into notes. Web2MD is more focused on AI use: token count, clean conversion, and one-click handoff to tools like ChatGPT, Claude, and Cursor.

So the short version is:

• Use Jina Reader for quick public URL reading.
• Use Firecrawl for API-based crawling and extraction.
• Use Turndown if you are building your own converter.
• Use MarkDownload for Markdown clipping.
• Use Web2MD when you want browser-side Markdown for AI, especially on pages that need your logged-in session.

Privacy and limits

Web2MD's privacy advantage comes from running in the browser. You are not sending a private URL to a remote reader just to find out whether it can fetch the page. For sensitive pages, that matters.

There are limits. Web2MD is Chrome-only today. If you use Safari or Firefox, that is a real constraint. The free tier includes 3 conversions per day. Pro is 9 dollars per month. Some pages with unusual JavaScript rendering or aggressive anti-copy behavior may still need manual cleanup.

I would rather be clear about that than pretend every page becomes perfect Markdown. It does not. But for the pages I tested, especially docs, articles, support portals, and logged-in content, the output was usually much cleaner than raw copy-paste.

Small habits that save tokens

The tool helps, but the habit matters too.

Do not send the entire page if you only need one section. Keep the source Markdown close to the question. Delete comments, related links, and repeated navigation if they slipped through. Preserve tables and code blocks. Add a short instruction telling the model what to ignore.

A good prompt might be:

"Use only the Markdown below. Summarize the rate limits and list any billing risks. If the source does not say something, say that it is not stated."

That kind of prompt plus clean Markdown is usually cheaper than dumping a full page and asking a vague question.

Bottom line

Reducing LLM token cost is not only about shorter prompts. It is about cleaner inputs.

Web2MD belongs in the same conversation as Jina Reader, Firecrawl, Turndown, and MarkDownload, but it solves a specific problem: turning the page already open in your browser into AI-ready Markdown, including logged-in or paywalled pages that server-side tools cannot reach.

If your AI workflow involves copying web pages into ChatGPT, Claude, or Cursor, try converting the page to Markdown first. Start with the free 3 conversions per day, watch the token counter, and see whether your prompts get smaller and easier to control.