ChatGPT Exports Changed

ChatGPT Exports Changed — Here’s What Broke (and How We Fixed It)

We recently noticed something odd in Export Reader — some ChatGPT conversations were showing up as blank, even though the data was clearly there. At first glance everything looked fine: titles were correct, summaries existed, the archive loaded. But when you opened a chat, there was nothing inside.

It turns out OpenAI quietly changed the structure of their export files. No announcement, no versioning — just a subtle shift that broke anything expecting the old format.

What Changed

The differences weren’t obvious at first, but once we dug into the exports, a few key changes stood out:

• New file naming — instead of a single conversations.json, exports can now include multiple files like conversations-019.json
• Deeper nesting — message text is no longer in predictable locations and can be buried several layers deep
• Extra metadata — internal system instructions and structured data are now mixed into the payload
• Non-visible content — things like tool calls and internal context can appear alongside real conversation text

None of these changes are a problem on their own — but together, they break parsers that expect a consistent structure.

What Broke in Practice

Because of these changes, a few things started going wrong:

• Blank conversations — chats would load but show no visible messages
• Incorrect summaries — internal instructions were being picked up instead of real conversation content
• Missing message text — valid content was present but not being extracted correctly
• Weird characters — internal formatting markers leaking into visible text

If you’ve experienced any of these recently, this is why.

How We Fixed It

To make Export Reader work reliably with both old and new exports, we made a few key improvements:

• Flexible file handling — support for both conversations.json and conversations-###.json
• Deep text extraction — instead of relying on fixed paths, we now scan the entire message structure for readable content
• Cleaner rendering — internal markers and system metadata are filtered out before display
• Better fallbacks — if content appears in unexpected places, it’s still picked up and rendered correctly

The goal is simple: if there’s readable text anywhere in the export, you should see it.

Why This Matters

AI platforms are evolving quickly, and their export formats are changing with them. Small structural changes like this can quietly break tools that rely on consistency.

This update makes Export Reader more resilient going forward — not just for this change, but for future ones too.

Conclusion

If you recently uploaded a ChatGPT export and something looked off, it should now be resolved. Conversations that previously appeared blank should render correctly, and summaries should better reflect the actual content.

If anything still doesn’t look right, feel free to get in touch — these formats are still evolving, and we’ll keep adapting as they do.