Listen: Fanout Query Analysis

When you ask an AI model like Gemini, GPT, or Nova to search the web, it does not just run your prompt as-is. Instead, it breaks down your question, explores subtopics, and verifies information by generating its own internal search queries. These are known as fanout queries, and a single user prompt can trigger several of them.

By analyzing over three hundred and sixty thousand of these background queries from real production workloads, we can see exactly how the major AI providers differ in how they search.

Google tends to generate searches that sound like natural questions. Its queries use more verbs and inquiry words like "how," "what," and "which" to find information.

OpenAI takes a different approach. Its queries focus heavily on nouns, particularly proper nouns. It searches for specific entities, and it uses twice as many numbers as the other models, likely looking for specific years and quantities.

Amazon leans into description. Its model uses significantly more adjectives, crafting queries with qualifying terms like "best," "top," or "most effective."

So, while we only see a single final answer, these models are busy behind the scenes, each using its own unique search strategy to track down the facts.