What do people most often seek help with on individual ChatGPT accounts? The researchers report that practical guidance, information seeking and writing account for about 77% of the usage studied. For businesses, this pattern suggests considering customers’ needs to learn and solve problems alongside their interest in products.
ACADEMIC WORK REVIEWED
How People Use ChatGPT
ResearchersAaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan, Kevin Wadman
NBER · September 2025 · Working paper version examined, not peer reviewed · Working paper · Source language: English
English review: 11 November 2026 · The source date appears in the citation above.
Editorial review draft
The source has been checked; the review still carries its original editorial status. DijitalPi did not conduct a customer experiment for this paper.
LET'S READ THE RESEARCH TOGETHER
We first explain the researchers' question, method and findings. We then discuss how to interpret the results, clearly separating DijitalPi's commentary from the source.
01 / WHAT DID THE RESEARCHERS WANT TO UNDERSTAND?
What do people seek help with on ChatGPT, and what do they want the assistant to do?
The researchers wanted to understand the needs behind ChatGPT usage alongside its scale. They separated work-related and non-work messages. They examined conversation topics and whether users sought information or a completed output.
Explaining the context · DijitalPi commentary
DijitalPi commentary: A business content plan can address questions beyond the moment of purchase. This study offers a starting point for considering learning and practical guidance. The questions your own customers ask still need to be established separately.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
The researchers sampled one message from each of approximately 1.1 million conversations on individual Free, Plus and Pro accounts. The main sample covers May 2024 to June 2025. It excludes organisational plans. The research team did not read user messages; models assigned classification labels.
Understanding the method · DijitalPi commentary
The system first removed identifying information. Models then classified messages by work relevance, topic and intent, among other dimensions. They used the previous 10 messages as context. Researchers saw only the resulting labels.
The team separately compared model labels with human labels on public WildChat conversations. This check did not involve people reading the private messages in the main sample. Employment and education analyses used a secure environment that allowed only aggregate computations.
03 / RESEARCH FINDINGS
Which purposes were most common in ChatGPT messages?
Practical guidance, information seeking and writing accounted for about 77% of the usage studied. The share of non-work messages rose from 53% in June 2024 to 73% in June 2025. Estimated daily work-related messages rose from 213 million to 716 million, based on 7-day averages ending on 26 June 2024 and 26 June 2025. A falling share of work messages did not mean that their volume fell. The main findings section reports writing at about 40% of work messages for July 2025. About two-thirds of all writing requests involved modifying text supplied by the user. Across all messages, the intent classifier labelled 49% as requests for information or advice, 40% as requests to perform a task and 11% as expressions of views or feelings without seeking information or action.
Combined share of the most common topics
About 77%The combined share of practical guidance, information seeking and writing in the usage studied.
Non-work message share
73%An estimate for messages on individual plans in June 2025, not a share of users or sales.
Requests for information or advice
49%The share of all messages assigned to the Asking intent category.
DijitalPi commentary: A product page can explain price and specifications, while supporting content could explain the circumstances in which someone would need that product. The study gives a reason to consider these information needs in content planning. It does not establish which pages an AI answer will cite.
DijitalPi commentary: A substantial share of writing requests involves revising existing text. We suggest checking whether current explanations answer customer questions, rather than concentrating only on producing more pages. Frequent requests for editing do not establish that the resulting text is more accurate or effective.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
This NBER working paper has not been peer reviewed. Most authors list OpenAI in their affiliations, and the data come from OpenAI’s own platform. This is not an evaluation independent of the product company. Models classified the messages without people reading them. On a separate validation dataset, agreement with humans was higher for work relevance, more limited for topic classification and weak for interaction quality. Organisational plans and conversations from users who opted out of sharing data for model training are excluded. The study did not measure sales, brand searches or the effect of content changes on visibility. It does not provide a separate usage distribution for customers in Türkiye. The authors’ view that decision support may improve productivity is not a measured productivity gain in this study.
Our open question at DijitalPi: which questions your customers ask before choosing a product remain unanswered on your website, and how will you track the visibility of the explanations you prepare separately?
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
What information does a furniture buyer need?
An illustrative furniture retailer receives questions from customers choosing a table for a small room.
Which question should our content address?
Open the recommendation for example 01+
Explain product specifications alongside conditions of use
DijitalPi commentary: Draft a guide explaining what customers need to check against their room measurements. Use verified information about product dimensions and placement requirements. Track whether ChatGPT cites the guide separately from visits to the website. The study did not test whether this type of content improves visibility or sales.
EXAMPLE 02
What do buyers ask before choosing technical parts?
An illustrative industrial parts supplier notices customers requesting quotes before checking compatibility and operating conditions.
What should we add to the product description?
Open the recommendation for example 02+
Explain the conditions that determine compatibility
DijitalPi commentary: Collect questions the technical team regularly answers. Explain suitable operating conditions and exceptions that affect selection on the relevant page. Do not treat the study’s information-seeking share as demand for your sector. Record brand mentions in AI answers, quotation requests and completed sales separately.
EXAMPLE 03
How could an existing service description become clearer?
An illustrative training provider’s marketing manager suspects that customers struggle to understand course prerequisites.
What could we try before creating more content?
Open the recommendation for example 03+
Compare the current text with actual customer questions
DijitalPi commentary: Clarify who the course suits and what prior knowledge it requires. If AI helps simplify the draft, have a specialist check that the conditions remain accurate. Requests to edit text in the study do not prove that editing improves content quality or enrolments.
TRY IT WITH YOUR TEAM
Which customer questions should guide the content plan?
DijitalPi commentary: We suggest making a content list that maps questions received by sales and support teams to existing pages. Identify missing explanations and compare them with search data to choose pages to create or update. Assess citations in AI answers, website visits and sales separately. This proposal has not yet been implemented.
DijitalPi commentary: SEOMACHINE supports content prioritisation using search data, source checks and publication with human approval. This workflow does not guarantee citations in ChatGPT or increased sales. Explore SEOMACHINE ↗
Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan, Kevin Wadman Aaron Chatterji: Duke University, Fuqua School of Business ve OpenAI. Christopher Ong: Harvard University ve OpenAI. Zoe Hitzig: OpenAI ve Harvard Society of Fellows. Yazarların çoğunun kurum bilgisinde OpenAI yer alıyor.
The academic work belongs to the researchers named above. This page contains DijitalPi's explanatory review and original business examples; it is not a full translation of the paper.
Cookie preferences