AFAICT, the Responses API does not offer any way to get the raw web search results as plaintext. This is possible for File Search and I’ve found it extremely useful for my applications.
I think all we can currently get from the Web Search Tool results is the generated response based on the raw web search results and any URLs (+ webpage titles) that were cited in the generated response. But that is not as helpful for me as being able to see the raw web search results.
More than that: what you don’t see is internal iterations of clicks and site reads by tools. Placed into growing chat history, until the AI calls no more tools. Then the billed one-time-fee “search” is context is not a product for you to even use in future lines of inquiry.
Also, despite disallowing results from a particular site, about that site, in the request, and stating the behavior needed in searching, 50% of the citations were directly that site. Search tool behaves little more than a trigger and delivers a googling.
I expect the behavior is tuned to overcome publisher’s potential objections: you never get more than a factual snippet, and there are citations to drive clicks.
You’d have to build your own multi-layered tool if you want an AI simply to be “informed” and not be a pattern parrot: a tool call to an AI that can use the search model and collect the provided search texts, and give them back in a format that isn’t accompanied by demands.
