About 「Grounding with Google Search」--where is the limit?

Google AI Studio
,
1

hi there.
This long time I am using the extremely excellent Gemini 2.5 Pro Preview 03-25 for my academic research about sociology, law and political science.
As you know, academic research always needs to cite various reports and articles, so it is inevitable that Grounding with Google Search will be heavily used.

But in the process of using it, I have found so many problems:

  1. 「Grounding with Google Search」 function will not always comply with my prompt as effective, it will sometimes find, sometimes not: not mostly, even if I have already constrained in the prompt for the need to find the content of the text, as well as the scope of the search and the way to find.
  2. The 「Grounding with Google Search」 feature tends to work only once or twice at the beginning of a conversation, and not the rest of the conversation. Basically, after 30k tokens, this function will be disabled.
  3. There seems to be a hidden number of times a conversation can use this feature, often after 5 times, and no matter how low the tokens are, it will definitely not work.

So I’m curious as to why this feature fails so many times and seems to wander around like a ghost. The only explanation I can offer is the following:

  1. 「Grounding with Google Search」 has a daily limit. 2.
  2. The way 「Grounding with Google Search」 is implemented requires reading the above, and once the number of tokens processed in the above is too many, the limit will be reached and then it will be invalidated.

I don’t know if my guess is correct, but I think this would be better:

  1. 「Grounding with Google Search」 should more clearly describe when it will take effect.
  2. 「Grounding with Google Search」 should be labeled if it has a limit.
  3. provide a button or something to let users decide when to use 「Grounding with Google Search」.

Extra: I hope to reveal more information about this feature, including the technical principles of implementation.

2 Likes
3

what a mystery, but i love this function

3 Likes
5

It really is mysterious. I would also love some more guidance on how to get it to reliably call grounding everytime a prompt is run

2 Likes
9

hi clintin,
i am glad to talk with you about this, and to be honest, all of my words are here :joy:

2 Likes
10

AWWWWWWWW MAN HERE WE GO AGAIN
When this issue crosses over with my earlier reflection that the 2.5 pro preview is often lazy and doesn’t do Thought, the quality of the whole model drops drastically.
I’m curious, could it be that my intense work is making the model tired?
But I have used various methods! Including telling him I’d pay a tip incentive - which I actually did, I turned on ADVANCED!

1 Like
11

me too, man :innocent:
wish developer can be more generous

1 Like
12

Most of the time I just see “synthesizing/generating a search result” and grounding with google search not used at all - be happy the model even cares to use it haha

2 Likes
14

image
image735×217 13.3 KB

I’m 100% sure this is the functionality I need when using CoT and Grounding with Google Search, specifically, the latter would like to be refined to the point of choosing to activate or deactivate it for each conversation

1 Like
16

I have seen this happen too. If it helps, I have noticed that the flash model tends to call the grounding search tool more consistently than the pro model. Wondering if you see the same?

1 Like
18

YES !!
The Flash model lives up to its newest release, both with the addition of the manual think button and the invocation of the Grounding with Google Search feature, which is much better than the 2.5 pro. This is something I got from testing both on Vertex AI.
All I can do now is hope that Gemini 3.0 comes soon!