Hi everyone, I tried the new web_search tool in the Response API (using the gpt_web_search_preview model), but its results are very poor. I requested the latest news in America, but they were about four months ago. Then I tried the latest sports news, and those were even worse—around four years ago.
I believe I’ve found the problem.
I tried executing a simple query, specifying in the location object the country US and the city San Diego, with this simple input:
“Give me the latest sports news in the US.”
The list of responses was accurate, and the links associated with them were perfectly consistent with the summaries shown in the response message.
Instead, by setting the location Italy (IT) in the payload (it’s my country) and requesting the input:
“Give me the latest sports news in Italy.”
The result was poor. The response contained very old news (several months old) and the links provided for each suggestion were not coherent, as they all referred to the homepage of a sports site in Italy (without any specific reference to the news summarized in the response message).
At this point, I believe that “location” is a reliability factor and that the model performs better in the US than in other countries.
I invite you to conduct similar tests and share your opinion.