Author: Metehan Yeşilyurt
Profile: https://www.linkedin.com/in/metehanyesilyurt/
ChatGPT retrieval system LEAK alert PART 2💥
How ChatGPT ranks and scores chunks from fetched pages? 🤔
It was one of the great debates in AI visibility. Now we have proof. Yes, ChatGPT ranks chunks. But how?
The ChatGPT SSE dump gives us the core retrieval mechanics.
We started with one shopping prompt ("What are the best wireless noise cancelling headphones under $300 right now? Compare the Sony WH-1000XM5, Bose QuietComfort Headphones and Sennheiser Momentum 4."), got 5 product cards back, and pulled the web retrieval results that ran behind them.
Here is one raw record from a SoundGuys headphone comparison:
1. Chunks from fetched pages
When ChatGPT fetches search candidates, it does not treat documents as uniform blocks. It segments fetched pages into discrete text passages and ranks individual chunks with neural relevance scores. THIS CHANGES WITH A/B testing variants, still working on it.
2. One score per passage
The record stores snippet::parts (3 passages from the page) and snippet::scores (3 scores, one for each passage). The battery life passage scored 0.9817, the "what to buy instead" passage 0.9937, and the Sennheiser and Monoprice alternatives passage 0.9984.
3. Passages are copied raw
Each passage keeps its markdown heading and its inline citation anchors like 【29†Our standard battery test】. Nothing is rewritten. The chunk is scored as it is. (markdown/text)
I didn't add any key takeaways, leaving it to you. So let's go with takeaway for GEO and AEO teams, drop yours below 👇 … more
How ChatGPT ranks and scores chunks from fetched pages? 🤔
It was one of the great debates in AI visibility. Now we have proof. Yes, ChatGPT ranks chunks. But how?
The ChatGPT SSE dump gives us the core retrieval mechanics.
We started with one shopping prompt ("What are the best wireless noise cancelling headphones under $300 right now? Compare the Sony WH-1000XM5, Bose QuietComfort Headphones and Sennheiser Momentum 4."), got 5 product cards back, and pulled the web retrieval results that ran behind them.
Here is one raw record from a SoundGuys headphone comparison:
1. Chunks from fetched pages
When ChatGPT fetches search candidates, it does not treat documents as uniform blocks. It segments fetched pages into discrete text passages and ranks individual chunks with neural relevance scores. THIS CHANGES WITH A/B testing variants, still working on it.
2. One score per passage
The record stores snippet::parts (3 passages from the page) and snippet::scores (3 scores, one for each passage). The battery life passage scored 0.9817, the "what to buy instead" passage 0.9937, and the Sennheiser and Monoprice alternatives passage 0.9984.
3. Passages are copied raw
Each passage keeps its markdown heading and its inline citation anchors like 【29†Our standard battery test】. Nothing is rewritten. The chunk is scored as it is. (markdown/text)
I didn't add any key takeaways, leaving it to you. So let's go with takeaway for GEO and AEO teams, drop yours below 👇 … more
Images:
