Markdown exposes the token economics hidden inside enterprise AI workflows
Christopher S. Penn’s test found the same document cost 28,693 tokens in Markdown and 52,163 tokens as Word DOCX. In enterprise AI, document format is no longer a publishing preference; it is an operational cost and retrieval-efficiency variable.
Editor’s Note by Dageno: Dageno isolates the “Fat Documents” section of Almost Timely and turns it into a GEO and AI operations brief about why clean, text-based formats reduce token cost, improve processing speed and make knowledge easier to retrieve.
Understand the article in 60 seconds
- Penn tested a Word document and equivalent Markdown file with Claude Code connected to a local AI server using Qwen3.6-35B-3AB on a MacBook.
- The Markdown version summarized in 28,693 tokens with 20,480 cached tokens and 71.4% efficiency. The DOCX version used 52,163 tokens, 26,624 cached tokens and 51% efficiency.
- The Word file was 81% heavier and 28% slower/less efficient in the test. At enterprise scale, hundreds of employees dropping Office documents into AI can turn file format into a large token budget issue.
- The recommended direction is not “never use Office,” but convert operational knowledge into open, text-based formats: Word to Markdown, Excel to YAML, and PowerPoint to small images where visual context matters.
Document format is now part of AI operating cost
The original article’s “Fat Documents” section reframes file format as a token-economics problem. In a token-billed environment, the wrapper around the content can materially affect cost, speed and efficiency.
Markdown is lighter because it carries semantic text with minimal formatting overhead. DOCX contains more structural and formatting complexity, which can become extra tokens when an AI tool parses the file.
The test matters because it used the same content in two formats
Penn’s test compared a Word document with its equivalent Markdown version under the same naive instruction: summarize this document. The backend model was Qwen3.6-35B-3AB running locally through his setup.
That design matters because the performance difference is not about the content changing. It is about the format changing how much the model had to process.
Cached tokens do not always mean cheaper tokens
The Markdown run used 28,693 tokens, including 20,480 cached tokens, and reached 71.4% efficiency. The Word run used 52,163 tokens, including 26,624 cached tokens, and reached 51% efficiency.
In many pay-as-you-go AI services, cached tokens can be cheaper than new tokens. But subscription and quota plans may still count cached tokens against usage. That means inefficient formats can burn quota even when compute reuse exists.
The enterprise impact scales through everyday behavior
One employee summarizing one Word file is a small cost. Hundreds or thousands of employees uploading DOCX, PowerPoint and Excel files all day can create a persistent 81% overhead pattern.
This is why document governance belongs in AI operations. Format standards, knowledge-base hygiene and conversion workflows can reduce spend without changing the model.
Format choice also changes GEO and internal retrieval quality
For GEO and AI visibility teams, cleaner formats are not only cheaper. They are easier to chunk, embed, retrieve and cite. Markdown headings, lists and plain-text structure make passages more self-contained.
When internal content becomes the source for RAG systems, the same principle applies: clean source documents produce cleaner chunks and more reliable answers.
What the test reveals about AI-ready documentation
The following points are rewritten as operational paragraphs rather than a repetitive table, so each source fact carries its own strategic meaning.
The Markdown version consumed 28,693 tokens and reached 71.4% efficiency, making it faster and lighter for the same summarization task.
The Word DOCX version consumed 52,163 tokens and reached 51% efficiency, showing how format overhead becomes model work.
Penn calculated the DOCX file as 81% heavier in token usage. In a token-billed system, that difference becomes budget, not aesthetics.
The Word version was also 28% slower/less efficient in the test, which matters for agent workflows and batch processing.
The local setup let Penn see token behavior in real time. Enterprises often miss this because hosted tools hide the processing details.
Word can be converted to Markdown, Excel to YAML and PowerPoint to small images where the slide’s visual structure is the asset.
Cached tokens may be cheaper in some API pricing, but not every plan discounts them. Token quotas can still be consumed by inefficient documents.
A potential 45% enterprise-level token reduction is large enough to justify documentation standards and employee enablement.
Clean documents improve both cost control and retrieval precision
The GEO lesson is that AI systems do not only care about what you say. They also depend on how that information enters the retrieval pipeline. A clean Markdown page is easier to split into meaningful chunks than a bloated office file.
This affects public content and internal knowledge bases. If sales decks, product specs, support articles and research notes are fed into RAG systems, the source format influences answer quality, latency and cost.
The operational move is to create a “document token budget” discipline: default formats, conversion rules, repository standards, and periodic audits of which assets create the most token overhead.
Teams that need to act on this signal
Set default document formats for AI workflows and identify where Office files create avoidable token overhead.
Publish AI-facing knowledge in Markdown or clean HTML so important passages are easy to chunk and retrieve.
Use lightweight structured formats for glossaries, evidence libraries, FAQs and comparison assets that may feed public or internal AI systems.
Create automated conversion pipelines from DOCX, XLSX and PPTX into AI-friendly representations.
Measure token usage by format before negotiating AI contracts or approving enterprise quotas.
Track token consumption, processing latency and answer quality before and after format changes.
Signals to monitor after the brief
Average tokens required to process equivalent content by format.
Share of reused tokens and whether the vendor discounts them or counts them against quota.
Time required for summarization, extraction or RAG ingestion by file type.
Whether generated chunks remain self-contained and semantically clear.
Token cost for recurring tasks such as report summaries, proposal drafts or knowledge-base ingestion.
Percentage of frequently used knowledge assets available in AI-friendly formats.
Execution priorities that turn the article into operating work
Audit the heaviest recurring AI workflows
Find which document types consume the most tokens in daily work.
Define format defaults
Use Markdown for text documents, YAML for structured tabular data and compressed images for visual slide context.
Build employee conversion paths
Make it easy for teams to convert files before uploading them to AI tools.
Measure before and after
Show token count, speed and output quality improvements to justify governance.
Apply the same logic to public GEO assets
Use clean structure so public content is cheaper to parse and easier to retrieve.
Turn AI visibility signals into execution.
Dageno AI is a complete execution platform from data monitoring → strategy → content generation → result attribution. It helps teams monitor AI answer surfaces, find strategy gaps, generate content actions, and connect visibility gains to business impact.
Alexa for Shopping