Grounding Classifier
Our model that predicts whether a query "deserves grounding" — whether an AI system will run a live web search to answer it.
When a user types a prompt into an AI, the system has to make a split-second decision. Will it answer using its frozen memory, or will it run a live web search to find fresh information? This process is called grounding, and knowing when it happens is crucial. If an AI doesn't search the web for a specific query, your online content has zero chance of influencing its answer.
To solve this mystery, we built the Grounding Classifier. We started by sending ten thousand prompts to Google’s Gemini model with search grounding enabled. By recording exactly when Google chose to search the web and when it didn't, we were able to train a replica of Google's internal "query deserves grounding" classifier.
This model matches Google's default threshold for dynamic retrieval. It gives us a highly accurate tool to predict whether a prompt will trigger a live web search.
Understanding this distinction is vital. Grounded and ungrounded answers to the exact same question can look completely different. One is anchored to the live, shifting web, while the other is locked in the past. By predicting which path a query will take, you can target the search-enabled prompts where your content can actually make an impact.
Grounding Classifier is our model that predicts whether a prompt will trigger grounding — that is, whether an AI system will run a live web search to answer it or reply from its trained-in memory alone. Knowing this in advance tells you which queries your content can actually influence.
We built it using the same technology behind AI Rank: we prompted Google's Gemini 2.5 Pro with search grounding enabled across 10,000 prompts, recorded whether Google grounded each one, and trained a replica of Google's internal "query deserves grounding" classifier. This matches Google's documented default dynamic retrieval threshold of 0.3.
The result is a fine-tuned binary classifier we use across our machine-learning pipelines, demonstrated publicly in our QDG tool. It matters because grounded and ungrounded answers to the same question can differ drastically — one is anchored to fresh web sources, the other to the model's frozen parameters.