How to get a Notion export into ChatGPT

Notion offers three export formats and only one of them is worth using here. Pick the option labelled Markdown & CSV. HTML wraps every paragraph in tags that cost length and add nothing, and PDF flattens the page hierarchy that made the wiki useful in the first place.

What lands in your downloads folder is a zip. Large workspace exports do not download straight away either, Notion emails a link when the archive is ready, and a very large one can arrive in more than one part.

What is actually inside the export

Every filename carries a 32 character hex id on the end, because Notion appends the page id to keep names unique. Each markdown file also starts with the page title as a heading, so titles survive even when the filenames get messy.

The text itself is small. A wiki of several hundred pages is usually a few hundred kilobytes of words. Images and attachments are what make the zip heavy, and they are the part ChatGPT has least use for: on every plan except Enterprise, document handling is text only, except Enterprise, so embedded images are discarded during extraction.

Why the zip itself goes nowhere

Archives are not unpacked. Upload the zip and ChatGPT holds one file it cannot see inside, and the attempt still counts against your quota.

Uploading the extracted files individually is worse than it sounds. The rolling cap is 80 files every 3 hours, and a mid sized workspace exceeds that on its own. If the wiki is destined for a custom GPT the arithmetic is harsher still, because a GPT accepts 10 files per GPT for its entire lifetime, not per conversation. Projects are more generous at 25 files per project on Go and Plus and 40 files per project on Pro, Business and Enterprise, and still nowhere near a page count.

One file solves all three at once.

Flatten the pages into one file

1. Unzip the export

   Into its own folder. If Notion sent several parts, unzip them all into the same place first.

2. Delete the attachment folders

   Images and PDFs pulled out of blocks are dead weight here. The markdown links to them will dangle, which is harmless.

3. Concatenate the markdown with a path header before each page

   The header line is the whole trick. It tells ChatGPT where one page ends and the next begins, and it preserves the folder structure that carried the meaning.

4. Strip the page ids

   One find and replace over the finished file.

5. Upload the single text file and say what it is

   "This is an export of my team wiki. Each page starts with a line reading === path ===."

On macOS or Linux:

cd notion-export
find . -name "*.md" | sort | while read -r f; do
  printf '\n=== %s ===\n\n' "$f"
  cat "$f"
done > wiki.txt
sed -E 's/ [0-9a-f]{32}//g' wiki.txt > wiki-clean.txt

On Windows PowerShell:

Get-ChildItem -Recurse -Filter *.md | Sort-Object FullName | ForEach-Object {
  "`n=== $($_.FullName) ===`n"
  Get-Content $_.FullName -Raw
} | Set-Content wiki.txt
(Get-Content wiki.txt -Raw) -replace ' [0-9a-f]{32}', '' | Set-Content wiki-clean.txt

The sort matters more than it looks. Sorted paths keep child pages next to their parents, so the flattened file reads in roughly the order the wiki was structured, rather than in whatever order the filesystem happened to return.

Databases need separate handling

A CSV in the middle of a text bundle becomes prose, which is the wrong shape for it. Uploaded as its own file, a CSV is more likely to be analysed with code, and spreadsheets are exempt from the token cap that applies to text files, with a separate size ceiling of about 50 MB.

So split them by size and purpose. A small reference table with twenty rows can stay in the bundle, because you want it read alongside the pages that mention it. A task database with four thousand rows should go up as its own CSV and be asked about separately.

The per row markdown folders that sit beside each CSV are usually worth keeping in the text bundle. That is where the actual writing lives, and the CSV holds only the properties.

Where this goes wrong

The file is accepted but not fully read. A single text file can hold up to 2 million tokens, which no wiki approaches, but acceptance is not the same as reading. Past a certain length the file is searched rather than read end to end, so questions about specific pages stay sharp while questions about the whole workspace get vague. Ask about a named area and it holds up.

Internal links stop resolving. Notion writes links between pages as relative file paths. Once flattened, those paths point at nothing. Tell ChatGPT plainly that the === lines are the map and that the inline links can be ignored.

A page went missing. Ask it to list the first and last page paths it can see, and to count the === markers. If the count is short of your file count, the upload truncated or the concatenation missed a folder.

Which route to pick

If you have one question about three pages, do not export anything. Copy those pages into the message. It is faster and the answer is better, because nothing else is competing for attention.

If you will ask about the wiki repeatedly, flatten it once and keep the file in a Project, where it stays attached across conversations instead of being uploaded again each time.

And if the real task is finding things in a workspace that keeps changing, an export is the wrong tool on the day you make it. It is a snapshot. Notion's own search and AI read the live workspace, which no exported file can do.