LLM Search Volume Prediction
Using a language model to estimate a query's monthly search volume — directionally useful for sizing topics, but not precise.
Can we use large language models to predict how often people search for something online? The short answer is no, not precisely. While language models cannot give you exact monthly search volumes, they do have a decent sense of general scale.
To find out just how close they can get, we tested Google’s Gemini against real search data. The correlation was weak to moderate. The AI was much better at ranking which topics were big versus small than it was at predicting precise numbers.
When we grouped the search volumes into five distinct buckets, only about thirty-five percent of the AI’s predictions landed in the exact right category. However, nearly seventy percent fell into either the correct bucket or the one right next to it. Interestingly, the model was much more accurate with average, middle-of-the-road search volumes than it was with the extreme highs or lows.
The practical takeaway here is clear. You should treat language model search volumes as general opportunity tiers for brainstorming, not as precise statistics. They are useful signals for high-level planning, but they are not a replacement for measured, real-world data. Always verify the numbers against your own analytics.
LLM search volume prediction is using a large language model to estimate how much a query is searched each month. The short answer on accuracy: no, it isn't precise — but it does have a general sense of scale.
We tested Google's Gemini against real Search Console impressions. Correlation was weak-to-moderate (Pearson ~0.41, Spearman ~0.57), meaning the model ranks big-vs-small topics better than it predicts exact numbers. Grouped into five volume buckets, only about 35% of predictions landed in the exact right bucket, though roughly 69% were in the right bucket or an adjacent one, with middle buckets more accurate than the extremes.
The practical takeaway is to treat LLM volumes as opportunity tiers for ideation, not precise figures, and always verify against your own analytics. It sits alongside query classification as an LLM-assisted planning signal rather than a replacement for measured data.