# Markdown vs PDF vs JSON: which format AI reads best

> PDF, Markdown or JSON for ChatGPT, Claude and AI agents? What each format preserves, where each fails, and why the best answer is offering more than one.

AI reads Markdown best for prose and structure, JSON best for exact values that need fields like period, currency and qualifier, and PDF worst because it stores layout rather than meaning. For important documents, keep the PDF for people and publish Markdown and JSON versions of the same fidelity-checked content for AI.

## The short comparison

|  | PDF | Markdown | JSON |
|---|---|---|---|
| Made for | Printing and viewing | Reading and writing text | Programs |
| Headings and order | Inferred from layout | Explicit (#, ##) | Explicit if the schema has them |
| Tables | Positioned text, fragile | Simple pipe tables | Rows and fields |
| Charts | Shapes; values often lost | Only if written as a table | Series as data |
| Units, period, qualifiers | Scattered in layout | In text next to values | Separate fields per value |
| Good for quoting | Weak | Strong | Weak for prose |
| Good for exact values | Weak | Good | Strongest |
| Human-friendly | Strongest | Good | Poor |

## PDF: great for people, weakest for AI

A PDF looks identical everywhere, which is why business documents are shared as PDFs. But an AI has to reconstruct structure from character positions. OpenAI’s help center notes that ChatGPT may not reliably extract exact values from image-based tables, scanned files or complex layouts, and recommends spreadsheets or text files when exact values matter. Charts inside PDFs are read only where visual retrieval is available.

## Markdown: the natural format for language models

Markdown is plain text with light structure: headings, lists, tables, links. Models read it directly with no reconstruction, it is compact, and quotes taken from it are accurate. Its weakness is that a value is still just text: “$4.2M” in a table cell depends on the header above it for its period and on a nearby sentence for “unaudited”.

Front matter (a YAML block at the top) helps: it can state the canonical URL, version and status, so an agent knows what it is reading and whether it is current.

## JSON: exact values with their meaning attached

JSON is for when an agent has to use a number rather than quote a sentence: compare periods, compute a ratio, fill a form. A good schema stores each metric with its own fields, so nothing depends on layout.

- value and unit/currency
- period (e.g. FY2025, Q3 2026)
- classification (actual, forecast, budget)
- qualifiers (unaudited, pro forma, “up to”)
- source reference (page, slide or cell)

## Why the answer is “both, from one source”

Different readers want different formats: a person opens the page, a chat assistant prefers Markdown, an agent doing analysis wants JSON. The risk is drift — three files maintained separately end up disagreeing. The format question matters less than having one fidelity-checked source from which every format is generated, at addresses that do not change.

That is how The Papers works: each document becomes one canonical AI Version with a web page, Markdown (/{room}/{doc}.md, with front matter for canonical URL, version, status and AI Readiness) and JSON (/{room}/{doc}.json, with blocks, metrics with period, currency and classification, qualifiers, claims and page provenance). All are checked against the source before publishing, and per-document analytics show which format each AI reader used.

## FAQ

### Is plain .txt as good as Markdown?

It is readable, but you lose headings and tables, which help models find and quote the right section. Markdown costs nothing extra.

### Should I convert all my PDFs to Markdown?

Converting is a start, but check the numbers: converters can drop headers or signs without warning. For documents where figures matter, verify the output against the source.

### Will ChatGPT read a JSON file I upload?

Yes, as text. JSON is most useful for agents and tools that process values programmatically; for conversational questions Markdown is usually easier.

### What about Word or Excel files?

They carry more structure than PDF — Word has real headings and tables, Excel has cell values and formulas. They are good sources for building Markdown and JSON.

Publish one fidelity-checked source and serve it as a page, Markdown and JSON. → https://thepapers.link/

## Sources

- [OpenAI Help Center — Data analysis with ChatGPT](https://help.openai.com/en/articles/8437071-data-analysis-with-chatgpt)
- [OpenAI Help Center — Visual retrieval with PDFs FAQ](https://help.openai.com/en/articles/10416312-visual-retrieval-with-pdfs-faq)

## Related

- [What is a canonical AI version of a document?](https://thepapers.link/blog/what-is-a-canonical-ai-version.md)
- [One source of truth for AI agents: why a link beats attachments](https://thepapers.link/blog/one-source-of-truth-for-ai-agents.md)
- [How to give your support bot and AI agents one source of truth for product documentation](https://thepapers.link/blog/product-docs-source-of-truth-for-support-bots.md)
- [How do I stop ChatGPT from misreading the numbers in my pitch deck PDF?](https://thepapers.link/blog/how-do-i-stop-chatgpt-from-misreading.md)

Updated 2026-10-27. Canonical: https://thepapers.link/blog/markdown-vs-pdf-vs-json-for-ai
