What are grounding queries?
Grounding queries are searches that large connection models (LLMs) tally to dive deeper into a user’s prompt, retrieve up-to-date accusation (beyond the training cut-off), and anchor their answers successful verifiable facts. Grounding queries are searches tally by machines, for machines to dress up for the inherent limitations of LLMs.
What do grounding queries look like?
As a profoundly flawed human, I often find examples easier to understand than method definitions. To that end, we collected 5,333 grounding queries from Google Gemini based connected 1,000 subtopics (roughly 5 per subtopic). The afloat information is disposable astatine the extremity of this post, but present are 3 cardinal aspects of grounding queries:
1. Grounding queries are “long-tail”
Grounding queries ranged from 3-17 words, pinch an mean magnitude of 6.56 words. Over half of grounding queries were 6-7 words. Here’s the wide distribution:
Let’s look astatine 3 examples that autumn successful that 6-7 connection saccharine spot:
These are queries that we mightiness historically deliberation of arsenic surviving successful the long-tail of search. Whether that conception is still useful successful an property of prompts and LLMs is up for debate, but astatine first glimpse these examples look for illustration the benignant of queries we mightiness aboveground successful heavy keyword investigation successful 2026.
Are grounding queries useful?
Yes, I deliberation it’s useful to propulsion backmost the curtain and spot the process Google is using. However, grounding queries are conscionable 1 portion of the query fan-out process. Consider the simplified illustration beneath for the query “Gaming Consoles”:
How did we retrieve these queries?
The Gemini (Vertex) API now gives constricted entree to grounding queries and returns the hunt phrases successful the [‘groundingMetadata’] object. That’s breathtaking news, successful theory, but it turns retired to beryllium messy successful practice. We encountered 2 awesome issues:
1. Gemini is greedy astir grounding
Even moving Gemini 3.5 Flash pinch a [MEDIUM] aliases higher “thinking” level, it’s uncommon to get much than 1-2 grounding queries. In galore cases, you’ll spot nary astatine all. If you’ve utilized AI Mode much, you cognize it almost ever runs aggregate grounding queries. We were capable to nudge the strategy instructions and push Google to make 4-6 grounding queries. While these queries are “real” (Google really ran the searches), it’s adjacent to opportunity that we cheated a small bit. AI Mode and existent query fan-out mightiness behave very differently.
2. Google is greedy for your money
On apical of token costs, grounding queries presently costs (as of August 2026) $14 per 1,000 queries. This is not a per-prompt cost. In our case, each punctual ran astir 5 grounding queries, truthful grounding information costs thing for illustration $7 per 100 prompts. This is orders of magnitude much than the token cost, and astatine enterprise-scale, it adds up fast. If you’re utilizing 3rd-party devices that tally connected your ain API keys, support a adjacent oculus connected costs.
Want to spot the information for yourself?
We’re sharing the 5,333 grounding queries, on pinch the 1,000 subtopics utilized to make them. Take a look for yourself, particularly crossed the topics astir applicable to your ain manufacture — you whitethorn beryllium amazed astatine really machines search.
