Unlike search, which updates in real-time, Large Language Models (LLMs) like Gemini are heavily biased by their training data, which can be a year or more out of date. To bridge that gap, LLMs run their own searches, sometimes called “grounding queries.”
Recently, Google has given us more access to these queries, so we dug into over 5,000 examples based on 1,000 subtopics across a range of industries.
What are grounding queries?
Grounding queries are searches that large language models (LLMs) run to dive deeper into a user’s prompt, retrieve up-to-date information (beyond the training cut-off), and anchor their answers in verifiable facts. Grounding queries are searches run by machines, for machines to make up for the inherent limitations of LLMs.
What do grounding queries look like?
As a deeply flawed human, I often find examples easier to understand than technical definitions. To that end, we collected 5,333 grounding queries from Google Gemini based on 1,000 subtopics (roughly five per subtopic). The full data is available at the end of this post, but here are three key aspects of grounding queries:
1. Grounding queries are “long-tail”
Grounding queries ranged from 3-17 words, with an average length of 6.56 words. Over half of grounding queries were 6-7 words. Here’s the overall distribution:
Let’s look at three examples that fall in that 6-7 word sweet spot:
standard types of auto insurance coverage
must have baking tools equipment beginner
extended stay hotel amenities vs traditional hotels
These are queries that we might historically think of as living in the long-tail of search. Whether that concept is still useful in an age of prompts and LLMs is up for debate, but at first glance these examples look like the kind of queries we might surface in deep keyword research in 2026.
2. Grounding isn’t based on human searches
So, grounding queries are driven by human search data, right? Well, no. Let’s look at three longer examples from our dataset (starting with the longest, at 17 words):
"best indie games of 2025" OR "biggest indie game hits of 2025" OR "popular indie games 2026"
laptop gpu vs desktop gpu performance difference 2025 2026 rtx 40 50 series
semiconductor fabrication front end back end steps lithography deposition etching doping
The first one is clearly using search operators, and the last two are a bunch of relevant terms crammed into a string. I’m willing to bet that the monthly search volume for these phrases is safely in the single digits (and that digit might be 0). While grounding queries might be informed by search data, the grounding process is ultimately machine-driven.
As SEOs, it’s comfortable to assume that grounding works much like the keyword research we know and love, but the data doesn’t care about our comfort. Grounding is Google running searches for itself, to surface content it might otherwise miss.
3. Gemini recognizes its own training cut-off
You may have noticed specific mentions of 2025 and 2026 in those examples. A full third (33%) of grounding queries we collected mentioned a year, and in some cases multiple years in sequence (ex. “2025 2026”). Note that Google reports the training cutoff of Gemini 3.5 as March 2026, but acknowledges that some data may be as old as January 2025 (in line with earlier Gemini models). One likely goal of grounding is to overcome the limitations of training data and collect the most up-to-date results.
Are grounding queries useful?
Yes, I think it’s useful to pull back the curtain and see the process Google is using. However, grounding queries are just one part of the query fan-out process. Consider the simplified example below for the query “Gaming Consoles”:
Query fan-out has to consider the searcher’s intent and any specific needs for that intent. Someone searching for “gaming consoles” may be interested in making a purchase (commercial intent), looking for a specific brand or product, and likely wants the most current information. Grounding queries are generated to cover those intents and needs.
Grounding queries give us insight into the fan-out process, but they’re a machine’s interpretation of the searcher’s intent. We still need to add the human perspective.
How did we retrieve these queries?
The Gemini (Vertex) API now gives limited access to grounding queries and returns the search phrases in the [‘groundingMetadata’] object. That’s exciting news, in theory, but it turns out to be messy in practice. We encountered two major issues:
1. Gemini is greedy about grounding
Even running Gemini 3.5 Flash with a [MEDIUM] or higher “thinking” level, it’s rare to get more than 1-2 grounding queries. In many cases, you’ll see none at all. If you’ve used AI Mode much, you know it almost always runs multiple grounding queries. We were able to nudge the system instructions and push Google to generate 4-6 grounding queries. While these queries are “real” (Google actually ran the searches), it’s fair to say that we cheated a little bit. AI Mode and actual query fan-out might behave very differently.
2. Google is greedy for your money
On top of token costs, grounding queries currently cost (as of August 2026) $14 per 1,000 queries. This is not a per-prompt cost. In our case, each prompt ran about five grounding queries, so grounding data cost something like $7 per 100 prompts. This is orders of magnitude more than the token cost, and at enterprise-scale, it adds up fast. If you’re using 3rd-party tools that run on your own API keys, keep a close eye on costs.
Want to see the data for yourself?
We’re sharing the 5,333 grounding queries, along with the 1,000 subtopics used to generate them. Take a look for yourself, especially across the topics most relevant to your own industry — you may be surprised at how machines search.
