Every PDF Converter Says It's 'For AI.' Almost None Tell You If It Fits.
A PDF to Markdown converter built for AI should tell you how many tokens the result is. It should also say if the result fits your model, and show how it would split into chunks. Many converters sell themselves as "PDF to Markdown for AI", yet most stop once the .md file is made. Of the 15 converters named in the dated update below, two show a token count and none does all three.
The workflow they leave out
Picture the job itself. You have a 90-page vendor contract, an annual report, or a research paper with dense appendices. You want to ask a model questions about it. So you find a converter, drop the PDF in, get a tidy Markdown file, copy the whole thing, and paste it into Claude or ChatGPT.
Then the chat rejects it, or cuts it off without saying so, or shows a length error with a token count you can't check. So you do the only thing the interface allows. You scroll to roughly the middle, select the top half and delete it. Or you paste the second half in a follow-up message and hope the model joins the halves. You have cut a 90-page document at a random page, maybe mid-table, maybe mid-clause. Nothing in your workflow told you where the safe cut points were, or how big the document was to begin with.
Every tool in that chain treated "produce Markdown" as the finish line. The chat window assumed you would bring text that already fit. No tool asked whether the document would fit, or how to break it up if it didn't.
"For AI" is a claim about the result
Markdown is a reasonable format to give an LLM. It is compact, it marks structure with a few plain symbols, and headings stay headings instead of dissolving into a wall of text. Getting to Markdown is still only the first step. Whether the document then works with a model depends on its size in tokens and on how it is split.
A model doesn't read files. It reads tokens, up to a hard limit called the context window. It also reads them with uneven attention. Facts in the middle of a long input are harder for a model to recall than facts near the start or end. Liu et al. (2023) measured this effect in the paper "Lost in the Middle: How Language Models Use Long Contexts".
That leaves three questions for every converted file:
- How many tokens is this? Pages, words and kilobytes don't tell you. Tokens depend on the tokenizer your model uses.
- Does that number fit the model I chose? A document that fits one model's window can be too big for a smaller one. It can also be far too big for an embedding model's limit.
- If it doesn't fit, where do I cut? And if I'm building retrieval, how do I split it so each piece stays coherent and carries enough context to be found later?
What closes the gap
MarkPrep's AI-Ready Pack answers the questions about tokens, fit and splitting once the Markdown is ready.
A live token count with a model selector. When the conversion finishes, you see the token count, and it changes when you switch models, because tokenizers differ. For OpenAI models, MarkPrep uses the o200k_base encoding. For Claude and Gemini it uses each vendor's published characters-per-token figure. Neither vendor publishes a tokenizer that runs offline in the browser, and OpenAI hasn't confirmed the encoding for its newest models. So all counts are estimates, labelled "approx".
A context-window check. Pick your target model and MarkPrep shows whether the document fits its context window. If the count comes back larger than the model's window, you know before you paste. You can then solve the problem on purpose instead of by deleting half the file.
A chunk preview. When a document doesn't fit, or when you feed an embedding model with its own token limit, it has to be split. MarkPrep shows the split in a heading-aware chunk preview that you can adjust. The default is 800 tokens per chunk with 100 tokens of overlap, adjustable from 100 to 2,000. It prefers to break at headings rather than mid-sentence. You see each chunk's boundaries before anything is indexed.
Run the same 90-page contract through that workflow. Instead of a bare file, you get a token count, a fits or doesn't-fit answer for the model you picked, and a chunk preview. The preview splits the contract at its own section headings, such as "Indemnification", "Limitation of Liability" and "Termination". Each chunk carries its heading path, so you know where it came from. You can then switch to a model with a larger window, or chunk the contract and retrieve the relevant sections.
Which converters show a token count?
Updated 13 Oct 2026. On 12 Oct 2026 we checked 15 document-to-Markdown converters, using each one's own website, documentation or README.
- Show an estimated token count: markdown.new, and Cloudflare's Workers AI Markdown Conversion API (a developer API).
- Count tokens in code to size chunks, with no count on screen: Docling, through its HybridChunker.
- List token allowances per plan, with no documented count on screen: BlazeDocs.
- No token count found: Firecrawl, LlamaParse, macmdviewer.com, markdowntools.com (it shows a word count), Marker, MarkItDown, markitdown.online, markitdown.tech, Pandoc, pdf2md.net and word2md.com.
None of the 15 combines a count for each model, a check against the context window and a preview of the chunks.
markitdown.tech is one converter that markets itself "for AI". It runs Microsoft's MarkItDown and produces clean Markdown from a range of formats. Some files convert in your browser and others go to its server. It can also split the result by heading level into a ZIP. Its home and pricing pages mention no token count and no context-window check, and its split has no token sizes. This doesn't fault MarkItDown's conversion. The gap is in what converters labelled "for AI" leave out after it.
What "for AI" should mean
MarkPrep's aim is to show people what their model will see before they paste. The AI-Ready Pack also exports the chunks as JSONL, with heading-path and page metadata for retrieval pipelines.
Convert a file and watch the estimated token count appear next to the Markdown. To compare models for a file, drop it into the token counter. To try different splits, use the chunk preview.
Try it on your own file
Convert a document and watch the token counter. It's free, needs no account and uploads nothing.