The Gemini (Vertex) API now offers limited access to grounding queries and returns search phrases in the object (‘groundingMetadata’). This is exciting news, in theory, but it proves complicated in practice. We encountered two main problems:
1. Gemini is greedy for grounding
Even running Gemini 3.5 Flash at a “thinking” level (MEDIUM) or higher, it is rare to get more than 1-2 ground queries. In many cases, you won’t see any. If you’ve used AI mode a lot, you know that it almost always runs more ground queries. We were able to modify the system instructions and push Google to generate 4-6 basic queries. While these queries are “real” (Google actually ran the searches), it’s fair to say we cheated a little. AI mode and actual query fan-out might behave very differently.
2. Google is greedy for your money
In addition to token costs, basic queries currently cost (as of August 2026) $14 per 1,000 queries. This is not a cost per request. In our case, each prompt ran about five ground queries, so ground data costs something like $7 per 100 prompts. This is orders of magnitude greater than the nominal cost, and at an enterprise scale, it adds up quickly. If you use third-party tools running on your API keys, keep an eye on costs.
