Files ending with .md often appear when people work with AI, and a reasonable question arises: why not just use Word or PDF? The popular answer is "AI understands Markdown better." This claim is half true, and the other half needs to be clarified with data.
What is Proven: Markdown is Efficient and Maintains Structure
Markdown is plain text with simple syntax. Hashes for headings, asterisks for bullet points, vertical bars for tables. The MarkItDown documentation, an open-source tool from Microsoft, describes Markdown as "extremely close to plain text" that still manages to represent important document structures, adding that its conventions are "highly token-efficient." The tool itself was created to convert PDF, Word, PowerPoint, Excel, and HTML into Markdown before being fed to language models.
The difference can be measured. Cloudflare, in its official post about the Markdown for Agents feature (February 12, 2026), exemplifies a single blog page that consumes 16,180 tokens in HTML and 3,150 tokens after being converted to Markdown, a reduction of about 80 percent. The simple heading "## About Us" is said to be only about 3 tokens, while its HTML counterpart is 12 to 15 tokens. This figure is for HTML, not for Word or PDF files, so it should not be read as a difference between Word and Markdown. However, the logic is the same: the less markup and decorative elements, the more contextual space for content.
What is Not Proven: Markdown is Always More "Understandable"
This is where the popular claim needs correction. The paper by He et al. (Microsoft and MIT, November 2024) tested plain text prompt formats, Markdown, YAML, and JSON on several GPT models. The conclusion is clear: there is no universally optimal format, even within the same lineage of models. In one MMLU testing task, the highest accuracy for GPT-3.5 turbo was actually with JSON (59.7) and the lowest with Markdown (50.0). Larger models, such as GPT-4, are more resistant to format changes.
A note on limitations: the paper tested OpenAI models from 2023, not the latest models, and tested prompt formats, not a comparison of Word versus Markdown files. So it does not prove that Markdown is inferior, only that the assertion "Markdown is definitely the best" lacks a strong foundation.
So When Does .md Make Sense
Three justifiable reasons:
- Context-efficient. Clean text without markup provides more space for the document's content.
- Structure remains readable. Headings, bullet points, and tables remain as text, while scanned PDF or Word files with many text boxes often lose reading order when extracted.
- Easy for humans to review and edit. You can open the file in any text editor and see exactly what the AI received.
And three things that should not be taken at face value: that AI's "native language" is Markdown so it must be smarter, that automatic format conversion improves answer quality, and that one format fits all models.
Practical Advice
If your document is long and full of structure, first convert it to Markdown and then quickly check the results: are the tables intact and are the headings in the correct order? If the answers feel odd, try another format and compare for yourself, as the results on your model are the most relevant evidence. If you prefer to see the explanation directly, there is a clip at taalenta.id/video/acc-kenapa-file-markdown-md-lebih-dipahami-ai-daripada-word-atau-pdf.
Sources
- Microsoft, MarkItDown (README GitHub, section "Why Markdown?"): https://github.com/microsoft/markitdown
- Cloudflare, "Introducing Markdown for Agents", February 12, 2026: https://blog.cloudflare.com/markdown-for-agents/
- He et al., "Does Prompt Formatting Have Any Impact on LLM Performance?", arXiv 2411.10541, November 2024: https://arxiv.org/abs/2411.10541