Send Your Note to the AI as Markdown and It Never Sees the Chart

The chart in your uploaded note is invisible to the model. When you send a Markdown, Word, or plain-text file to ChatGPT, Gemini, or Claude, the API extracts text and discards the rest. Screenshots, diagrams, embedded graphs — gone. PDF is the one file format that sends both text and images. That is the whole answer, and everything below is the evidence.¹²³

This post answers the question that Your Screenshots Are Invisible to a Text-Only AI deliberately left open. That piece explains what to do with images that do reach the model — write descriptive alt text, treat it as a vault discipline. But whether the image reaches the model at all is a separate question, and for most note formats, the answer is no.

What determines whether an AI sees your images?

The file format you upload is the deciding factor. Provider APIs handle text documents and PDF documents through two different pipelines. With a text document — Markdown, Word, plain text, PowerPoint — the API parses out the text layer and passes it to the model. With a PDF, vision-capable models receive each rendered page as an image alongside the extracted text.

That difference is not a quirk of one provider. All three major APIs — OpenAI, Google Gemini, and Anthropic Claude — treat PDF as the special case that includes visual content. Every other document format is text only.¹²³

The implication: if your notes live in .md files and you upload them directly, the model has never seen your Mermaid diagrams, your embedded screenshots, or the hand-annotated image you pasted in last Thursday. The Markdown image syntax (![alt text](path/to/image.png)) is a pointer, not a picture — it references a file that the API never receives. For more on what that syntax actually does and what it means for portability, see Your Markdown Is Portable Until You Add an Image.⁴

How ChatGPT handles uploaded notes and documents

OpenAI draws a clean line between PDF and everything else. For PDF, vision-capable models extract both text and page images and pass both to the model. For every other document type — .docx, .pptx, .txt, code files — the API extracts text only. The images those files might contain never make it through.

OpenAI's File inputs guide states the rule verbatim: "Non-PDF document and text files (for example, .docx, .pptx, .txt, and code files): the API extracts text only."¹ And for PDF: "PDF files: On models with vision capabilities, such as gpt-4o and later models, the API extracts both text and page images and sends both to the model."²

The file-size limit across all types is 50 MB per file, with no stated page limit. A single API request can include more than one file.¹

Scope note: These rules describe the OpenAI API as documented in the File inputs guide, accessed 2026-11-05. Consumer-facing ChatGPT.com may apply additional pre-processing before the API call. If you are testing through the web interface rather than the API directly, behaviour may differ.

How Gemini handles uploaded notes and documents

Google Gemini's document pipeline behaves the same way — PDF gets visual understanding; everything else becomes plain text. The Gemini API documentation puts it plainly: "However, document vision only meaningfully understands PDFs. Any file-type specifics like charts, diagrams, HTML tags, Markdown formatting, etc., will be lost."³

That sentence covers the full list of non-PDF document types that Gemini accepts: TXT, Markdown, HTML, XML. Send any of them and what the model receives is stripped text. The Markdown formatting itself — fenced code blocks, heading hashes, link syntax — is also lost, not just embedded images.

Gemini's PDF limits are generous: up to 50 MB or 1,000 pages per file.⁵ No token-per-page figure is stated on the documentation page.

Scope note: This describes the Gemini API (via Google AI Studio documentation), accessed 2026-11-05. The Gemini consumer app at Gemini.google.com has its own upload interface and behaviour; that is not covered by these API-tier rules.

How Claude handles uploaded notes and documents

Claude's PDF pipeline is the most explicit of the three about what happens visually. Anthropic's documentation describes the mechanism directly: "The system converts each page of the document into an image. The text from each page is extracted and provided alongside each page's image."⁶

Each rendered page is treated both as a visual object and as a text layer — which means charts, tables, and diagrams drawn on the page are visible to the model.

The token cost is real: "Each page typically uses 1,500–3,000 tokens depending on content density."⁷ The page limit is 600 pages at the full 1M-token context window, or 100 pages at smaller context windows.⁸

Anthropic's documentation covers PDF in detail but does not state an equivalent rule for non-PDF document types. Do not infer one from this post. For the text-only rule on other formats, the clear statements come from OpenAI¹ and Google.³

Scope note: This describes Anthropic's Claude API, documentation accessed 2026-11-05. Claude.ai web may differ.

Side-by-side: what each AI receives by file type

The table below compares what the model actually receives — not what you uploaded. All figures are from each provider's own API documentation, accessed 2026-11-05. Consumer-app behaviour may differ from the API-tier rules shown here.

Why missing images costs more than you think

A model that never received your chart cannot reason about it — but the accuracy penalty is larger than most people expect. Research published in Yoon et al., ChartMuseum: Testing Visual Reasoning Capabilities of Large Vision-Language Models (arXiv 2505.13444, 2025) measures the gap directly on chart-reasoning benchmarks.⁹

On the ChartMuseum benchmark — which tests charts that require visual reasoning rather than text extraction — models relying solely on extracted text scored 15.2% accuracy, versus 61.3% for models given the actual images.¹⁰ The authors are precise about why: "The 46% performance gap in ChartMuseum reflects information that is inherently visual and cannot be captured in the easily-accessible text form."⁹

The gap appears even on the more established ChartQA benchmark, where text-only Claude-3.7-Sonnet achieves 74.1%, rising to 87.4% when given the page images.¹¹ And the effect is not model-specific: "all models experience a 35%-55% performance drop from text-reasoning-heavy question performance" when shifted to visually complex chart questions.¹²

The practical translation: upload a Markdown file containing a performance graph and ask the model to interpret it. What you get back is the model reasoning from the numbers in your prose, if any exist — not the shape of the line, not the axis scale, not the annotation you drew. For a note with no inline text transcription of its visual content, you get a hallucinated interpretation or an admission of ignorance.

For situations where you do not want to export to PDF but still need an AI to handle your visual content, the adjacent approach — sending the image directly — has its own trade-offs. Send Your Notes to the AI as an Image and It Guesses the Numbers covers that path.

When you should export to PDF before sending to an AI

Export to PDF whenever the visual content in your notes is load-bearing — when the chart, diagram, or screenshot is the thing the model needs to reason about, not a decoration beside prose that already describes it.

Specific cases where PDF export changes the answer:

• Charts and graphs. A line chart showing a trend that you have not transcribed numerically. A bar chart where the relative heights matter. Any visualization where the shape conveys the finding.
• Annotated screenshots. Images with callouts, arrows, or highlights that are part of the information.
• Diagrams and flowcharts. System architecture sketches, decision trees, Mermaid diagrams rendered to images.
• Scanned documents with text as image. OCR-requiring content embedded as a rasterized image inside a .docx.
• Multi-column or visually structured layouts. PDF's per-page render preserves layout in a way the extracted text layer does not.

The step is portable across all three providers. Any Markdown-based note tool that lets you export your note as a PDF — MNMNOTE included — gives you this path. The export step is yours to control.

When plain text is the right choice

Plain text — Markdown, Word, TXT — is the right format to send whenever the images in your notes do not matter for the task at hand. If the note is primarily prose reasoning, structured headers, and bullet points, the text-only extraction is complete. Sending the PDF version would use more tokens (1,500–3,000 per page on Claude ⁷) without adding any useful information.

Plain text also removes the page-limit concern. A 10,000-word research note sent as a .txt file is a single text payload. The same note as a PDF might span 20–30 pages and cost proportionally more.

The rule of thumb: if you would describe your note as "mostly prose," send the text file. If you would say "there's a chart in here that matters," export the PDF.

Frequently Asked Questions

Does ChatGPT see images in my Word document? No. The OpenAI API extracts text only from non-PDF document types including .docx, .pptx, .txt, and code files.¹ Any images embedded in the Word document are not extracted. If the image has alt text, that text may pass through — but the pixels do not.

Does Claude see charts in uploaded PDFs? Yes. Claude converts each PDF page to an image and extracts text alongside it.⁶ Charts and diagrams visible on the page are part of what the model receives.

Why can't the AI read my graph? If you uploaded a Markdown or plain-text file, the model received text only. The image the Markdown syntax points to never reached the API. Export the note to PDF to change the outcome.

Does Gemini understand Markdown files? Gemini extracts plain text from Markdown. The documentation states that "file-type specifics like charts, diagrams, HTML tags, Markdown formatting, etc., will be lost."³ Formatting is stripped along with any embedded image references.

How many tokens does a PDF page use with Claude? "Each page typically uses 1,500–3,000 tokens depending on content density."⁷ At 200,000 tokens, a dense PDF could use most of a standard Claude context window in roughly 65–130 pages.

What file types can I upload to ChatGPT? OpenAI supports PDF — which provides text and images — and non-PDF documents including .docx, .pptx, .txt, and code files, which are text only. The per-file limit is 50 MB; a single request can include multiple files.¹

How do I make the AI see my screenshot in my notes? Export the note as a PDF. That is the one portable step that sends images to any of the three providers' APIs at the API level. The Markdown image pointer (![alt](path)) is a path reference — the file at that path is not uploaded with the note.

The file format you choose is the first decision in the pipeline — and unlike prompt-writing or model selection, it is entirely in your hands before you press send.

Notes stored as open Markdown files travel light and stay readable anywhere. When an AI session needs the charts too, one export step changes what arrives — and what gets answered. mnmnote.com

References