When I first started using the Grounding option, I was under the impression that search happens during the CoT process and that the model is able to react (during CoT) to the results of the search.
Today, however, I was inspecting some thinking chains, and realized that the model often says something like: “I will search for X… Let’s assume I found out Y…”.
Could someone from the team clarify when exactly does grounding take effect? Is it simply at the end of thinking, all in a single go, with the real output following it? Is the model able to “reflect” on the search results at all?