ChatGPTが回答を生成する際、内部ではどのような検索が行われているのか。umoren.aiの無料ツールを使い、実際のクエリファンアウトと検索プロセスを可視化します。
Let's Take a Look at the Search Process Actually Performed by AI
Generative AI responses are often said to be a black box. Users can only see the final answer, and it is almost impossible to check what kind of searches are being conducted behind the scenes.
However, in reality, ChatGPT executes multiple search queries in stages before generating a response. This internal search process is referred to as "Query Fan-out (クエリファンアウト)."
This time, umoren.ai has released a free tool that visualizes the search queries actually executed internally by ChatGPT.
In this article, we will introduce how to use the tool, the observed search structure, and how AI collects information.
What is the ChatGPT Query Fan-out Visualization Tool?
With this tool, you can simply input the prompt you want to check, and see the actual search queries (Query Fan-out) executed internally.
Importantly, this data is based on actual search logs, not on guesses or simulations.
How to Use the Tool
The operation is very simple.
1. Input the prompt you want to check
For example, you can input a prompt like the following.
“I am looking for a coffee maker for living alone. Ideally, it should brew automatically from beans and be easy to clean. Please recommend 3 to 5 models along with their features (extraction method, size, maintenance, price).”
2. Click on "Get ChatGPT Fan-out Queries"
Then, the search queries that ChatGPT executed internally will be displayed in a list.
ChatGPT's Query Fan-out Structure
Observing with the tool, it becomes clear that ChatGPT's searches are not conducted in a single round but in multiple rounds. In this example, three rounds of searches were conducted.
Round 1: Broad Exploration
In the first search, information is explored over a fairly wide range.
For example, the following query:
Here, the search is broadly scanning product categories, products available in the market, review articles, etc., to first pick up potential candidates for comparison. This can be seen as the initial exploration phase to find a group of candidates.
Round 2: Targeted Search and Site Deep Dive
What is interesting in the next stage is that the nature of the search changes significantly. The query includes specific brand names and product names such as Delonghi, siroca, and Panasonic.
This means that after finding candidates in the initial broad exploration, ChatGPT is now digging deeper into those candidates individually.
Moreover, it is important to note that this deep dive is not limited to mere targeted searches. Looking at the actual queries, we see that there are site: searches such as site:delonghi.com, site:siroca.co.jp, site:panasonic.jp, indicating that it is explicitly specifying certain sites to gather information.
From this, we can understand that ChatGPT's Query Fan-out is not merely expanding related terms. It first explores a wide range of candidates and then, based on which sites to look at for the promising candidates, gathers more accurate information.
Although rounds 2 and 3 are displayed separately in the image, it is more natural to view both as part of the same deep dive phase. In other words, overall, it can be summarized as:
Broad exploration -> Candidate extraction -> Deep dive through targeted searches and site specification
This constitutes a two-stage research process.
How is the Final Answer Generated?
The tool allows you to check not only the results of the query fan-out but also the sources of information that ChatGPT ultimately referenced and the generated answer.
The pages displayed here are the actual sources that ChatGPT referenced when generating the answer.
These pages are not simply selected from the initial search results but are chosen after multiple rounds of searches.
In other words, ChatGPT conducts a step-by-step research process where it first searches broadly for candidates, then narrows down the targets through targeted searches such as product names, and finally checks official sites directly to gather information.
After that, it integrates the content from the selected multiple pages to generate the final answer.
What can be understood from this is that ChatGPT's answers are not created by summarizing a single page, but rather are generated as a result of integrating multiple search rounds and multiple sources of information.
What This Discovery Means
This search structure provides very important insights when considering AI search strategies.
To be cited by AI, it is not enough to simply rank high in search results.
You need to be selected as a target for deep dives in the AI search process.
What emerges from this observation are the following stages:
- First, be recognized as a candidate in a broad search
- Next, become a target for targeted searches such as product names or service names
- Finally, become a page referenced as official information
In other words, to be cited by AI, you need to pass through these multiple stages.
Let's Actually See the Query Fan-out
By using the tool we released this time, you can check what kind of searches are triggered by the prompts you usually input into ChatGPT.
You can observe what queries AI issues, which sites it references, and how it integrates information to create answers.
You can now see the AI's research process, which was previously invisible, as actual data.
Please take a moment to check the query fan-out with your own prompts.
Get Found by AI Search Engines
Our LLMO experts will maximize your AI search visibility