Skip to main content
The model value is CHATGPT. Any behavior not listed here follows the Supported models table as-is.

Where answers are collected

ChatGPT answers are collected from the ChatGPT web interface that regular users see, not from the API. You get the same screen a real user sees when asking a question in a browser.
A prompt entered and an answer received in the ChatGPT web interface without signing in

The ChatGPT web interface used for collection. The prompt is entered as-is without signing in, and the answer shown below it is collected.

Collection conditions are as follows.
  • Collection happens without signing in. An account’s conversation history, memory, and custom settings do not affect the answer, so it is close to what any user gets when asking for the first time.
  • The prompt is entered verbatim into the input box. No instructions are added before or after it, and the question is not rewritten.
  • Only one prompt is asked in a new conversation. Context from earlier questions does not carry over.
  • The model is the one ChatGPT provides by default to signed-out users. No model is selected separately.
  • Region and location are applied by connecting from a network that matches the run request’s region / location. The same prompt can produce different answers and cited sources depending on where the connection comes from.
Screen elements map to answer object fields as follows.

Model-specific behavior

  • Fan-out is provided. The answer’s fanouts is filled with the list of search queries.
  • Location-level collection (location), which is narrower than a region, is supported. If you specify location together with a model that does not support locations, the entire request is rejected with 400 (LOCATION_NOT_SUPPORTED_FOR_MODEL).