Guide
How do I stop ChatGPT from misreading the numbers in my pitch deck PDF?
Give ChatGPT the numbers as real text, not as pixels: export a text-based PDF, put key figures in text or tables instead of chart images, and keep each number next to its period, currency and qualifier. The most reliable fix is to share a fidelity-checked, AI-readable version of the deck at a permanent link — for example with The Papers — so ChatGPT, Claude and investors’ own agents read the same checked figures instead of re-extracting them from the PDF.
Why ChatGPT gets deck numbers wrong
A pitch deck is designed for eyes. Numbers live in charts, callouts, footnotes and images, and their meaning depends on layout: which period a bar belongs to, whether “$2.4M” is ARR or revenue, whether a figure is a forecast.
OpenAI’s own help center says ChatGPT may not reliably extract exact values from image-based tables, scanned files or complex visual layouts, and recommends uploading a spreadsheet or text-based file when exact values matter. It also notes that reading charts and images inside PDFs (visual retrieval) is not available everywhere — in other contexts only the text layer is read, so a number that exists only inside a chart can be missed.
Fixes you can make in the PDF today
- Export from the source app (Keynote, PowerPoint, Google Slides) so text stays text. Avoid “print to image”, scans and flattened exports.
- Write every key metric as text at least once: “ARR $2.4M, December 2025, unaudited” — not only as a bar or a big styled number.
- Keep the qualifier next to the number: period, currency, actual vs forecast, gross vs net.
- Give charts a data label or a small table with the underlying values.
- Use one term per metric across the deck (do not switch between revenue, sales and GMV for the same figure).
- For financials, attach the model or a CSV/XLSX with clear column names — OpenAI recommends text or spreadsheet files when precision matters.
- When you ask ChatGPT, ask about a specific slide or section rather than the whole file at once.
Why this is still not enough when investors use AI
You control your own prompts, but not your readers’. Each investor uploads your PDF to their own assistant, which extracts the numbers again — differently each time — and an old version keeps circulating after you update the round.
What fixes this is a canonical version: one text-based representation of the deck at a link that never changes, which every AI reads instead of guessing from the layout.
How The Papers solves it
- Upload the deck (PDF or PowerPoint). The Papers builds one canonical AI Version as Markdown and JSON.
- Every figure is checked against the source slide and linked to the page it came from; periods, currencies and qualifiers stay attached.
- Values the source cannot settle on its own are sent to you to confirm, instead of being guessed.
- You get a permanent link. Send it next to the PDF; when you update the deck, the same link shows the new version and keeps the history.
- Readers can paste the link into ChatGPT, Claude, Gemini or Perplexity, and agents can read the Markdown, JSON and llms.txt directly.
FAQ
- Does ChatGPT read charts in PDFs?
- According to OpenAI’s help center, visual retrieval from PDFs is available in some products (stated for Enterprise); in other contexts only the text in the PDF is read. A figure that exists only inside a chart image can therefore be missed. Put important values in text or a table.
- Is a text-based PDF enough?
- It helps a lot, but each reader’s AI still re-interprets the layout. A fidelity-checked AI-readable version removes that step: the figures, their periods and their qualifiers are already explicit.
- Do investors need an account to read the AI Version?
- No. A public or unlisted AI Version opens from its link; private documents are limited to members of your Data Room.