What 50 million ChatGPT prompts reveal about user intent

Generative prompts ask for an output: write the email, build the budget, summarize the notes, create the itinerary, fix the code. They accounted for 37.5 percent of the classified sample, five points above informational prompts.

Traditional search taxonomies did not need this category. A search engine could help someone find a template, instructions, or software, but the work happened after the click. ChatGPT can return the draft in the response. The user judges the output there, then asks for revisions in the same thread.

In the largest intent category, ChatGPT returns the first version of the task itself.

Informational intent still made up almost a third of ChatGPT prompts. It did not vanish. Its share was 20 percentage points lower than the traditional-search baseline used in the study.

The navigational drop was much larger. Traditional search measured 32.2 percent navigational intent. ChatGPT measured 2.1 percent. People still ask for sites and brands, but they usually enter the chat to get an answer or an artifact, not to use it as a bookmark bar.

Transactional intent moved the other way, from 0.6 percent to 6.1 percent. The raw share remains smaller than informational or commercial use. The shift tells us that users are willing to bring purchase questions into a conversation even when the final checkout happens elsewhere.

The no-intent category covered 12.1 percent of prompts. These were conversational turns such as "thanks," "please," or "make it funnier." They matter because a chat is a sequence, not a stack of independent queries. A short correction can change what the user sees next even though it carries no standalone search intent.

A prompt log contains connective language that keyword research would throw away. If we study only the first turn, we miss the rejection, constraint, or revision that reveals whether the first answer worked.

Consider a person asking ChatGPT to create a software budget for a small company. The response can research the category, compare tools, recommend a stack, and put prices into a table. Several brands may influence the decision. None is guaranteed a visit.

Referral traffic captures only part of this. Citations, mentions, recommendation language, and the prompts that produced them show the work happening before a click. Revenue measurement picks up the sale; click-based attribution sees little of this earlier stage.

At 37.5 percent, generative intent changes what content has to do. A large share of users arrive with a verb and expect the model to return a usable object. Content earns influence when the model can use it inside that object, even when the user never opens the source.