OpenAI-Harvard Study: 1.1M ChatGPT Chats Reveal a Quiet Shift Toward Search

ChatGPT study: Executive snapshot

An increasing share of ChatGPT conversations resembles search-type queries. The points below summarize the latest usage mix and scale based on the OpenAI-Harvard working paper and OpenAI platform metrics.

• Information-seeking messages account for 24% of usage, up from 14% a year earlier; Writing fell from 36% to 24%, and Practical Guidance is ~29%. Together, these three categories are ~77% of usage [S1].
• Intent mix: 49% Asking, 40% Doing, 11% Expressing. Asking messages are rated higher quality than Doing/Expressing via automated evaluation and user feedback [S1].
• Personal use increased from 53% (Jun 2024) to 73% (Jun 2025). At work, Writing is the top task (~40% of work messages). Education/tutoring spans ~10% of all messages [S1].
• Programming is 4.2% of messages; relationship/personal reflection is 1.9% [S1].
• OpenAI reports 700+ million weekly active users by July 2025, sending ~2.5 billion messages per day (~18 billion per week) [S2].

One-line implication: information-seeking at this scale indicates measurable intent deflection from traditional search to LLM chat, warranting content built for conversational answers and citation.

Method and source notes for the ChatGPT study

The primary source is an NBER working paper (not yet peer-reviewed) produced by OpenAI and Harvard, analyzing a privacy-preserving sample of ~1.1 million consumer ChatGPT conversations from May 2024 through June 2025. The authors classify messages into topics (e.g., Practical Guidance, Seeking Information, Writing) and intents (Asking, Doing, Expressing) using automated classifiers, supplemented by platform-level user signals (e.g., thumbs-up/down) for quality estimates. No human reviewed user messages; the study excludes enterprise products, API traffic, and other LLMs [S1].

OpenAI’s platform scale figures (weekly active users and message volumes) are reported separately in an OpenAI post summarizing how people use ChatGPT (see research). These metrics are company-reported, not independently audited [S2].

Key caveats: findings generalize to consumer ChatGPT within the study window; automated classification introduces error; privacy constraints limit qualitative checks; and results do not directly measure displacement of web search behavior (no linkage to search logs) [S1][S2].

Findings: how people use ChatGPT for information, writing, and guidance

Information-seeking is now a core use case. The study estimates “Seeking Information” at 24% of messages, up from 14% a year prior, and notes it functions as a close substitute for web search for many tasks. Writing fell from 36% to 24%, while Practical Guidance is steady at ~29%. Combined, these three topics comprise ~77% of total usage, indicating concentration in pragmatic tasks rather than recreation or coding [S1].

Intent composition shows 49% Asking, 40% Doing, 11% Expressing. Asking-type prompts are consistently rated higher quality than Doing/Expressing by the automated evaluator and user feedback, suggesting ChatGPT responds more reliably when users request facts, definitions, or explanations than when they push execution tasks or emotional expression [S1].

Personal vs. work use shifted materially toward personal contexts: non-work messages rose from 53% in Jun 2024 to 73% in Jun 2025. Within work, Writing accounts for about 40% of work-related messages, while Education/tutoring spans ~10% of all messages across contexts. Programming content remains a small share at 4.2%, and relationship/reflection topics are 1.9% [S1]. OpenAI reports platform scale of 700+ million weekly active users and ~2.5 billion daily messages by July 2025, underscoring the potential reach of these usage patterns [S2].

Demographic and geographic adoption broadened: the share of users with typically feminine names rose from 37% (Jan 2024) to 52% (Jul 2025), and growth rates in the lowest-income countries exceeded those of the highest-income countries by more than fourfold during the period studied [S1].

Implications for SEO, SEM, and content strategy from ChatGPT information-seeking growth

Likely

• Expect incremental deflection of how-to, definitional, and quick-fact queries from traditional search to ChatGPT. Content that answers discrete questions with clear sourcing and concise summaries is more likely to be quoted or paraphrased in chat responses [S1][S2].
• Given that “Asking” prompts get higher quality responses, prioritize content formats that map to common questions: FAQs, comparisons, step explanations, and troubleshooting trees with unambiguous language and cited evidence [S1].
• With 73% of usage being personal, plan for consumer-oriented information needs around daily tasks and learning, not only B2B use cases [S1].

Tentative

• Writing remains a major work task in ChatGPT; documentation, policy pages, product specs, and briefs that contain precise terminology and data can help both human users and models generate accurate outputs that point back to your expertise [S1].
• Education representing ~10% of all messages suggests demand for structured learning content. Modular lessons, glossaries, and worked examples may travel well in conversational summaries [S1].

Speculative

• Rapid uptake in lower-income countries points to emerging markets for lightweight, text-based assistance. Localized, low-bandwidth pages with clear schema could improve inclusion in LLM-driven answers in these regions [S1].
• If platform-level advertising or affiliate surfaces expand within assistant interfaces, some upper-funnel SEM budgets could shift toward LLM environments, but evidence on performance and inventory quality is not yet available [S2].

Limitations, contradictions, and unknowns about ChatGPT usage data

• Scope bias: The study covers consumer ChatGPT only; it excludes enterprise, Teams, and API usage, as well as other assistants (e.g., rival LLMs). Findings may not reflect workplace or developer behavior at large [S1].
• Method constraints: Categories and intents are assigned by automated classifiers. Misclassification risk remains, and topic granularity may blur adjacent tasks (e.g., guidance vs. information) [S1].
• Privacy-preserving design: No human viewed content, reducing the ability to audit edge cases or nuanced intents (e.g., health vs. general advice) [S1].
• External validity: The paper characterizes information-seeking as a close substitute for search, but it does not observe web search logs, click-outs, or substitution at the user level. Actual displacement of search volume is unquantified [S1].
• Platform scale: Weekly active users and message counts are reported by OpenAI and are not independently verified. Cross-platform duplication and multi-account behavior are unknown [S2].
• Geography and demographics: Name-based gender inference is an imperfect proxy, and country-level growth comparisons may be sensitive to normalization choices and access constraints [S1].
• Time window: May 2024-June 2025 results may shift as product features (e.g., browsing, GPT-4 variants, voice) evolve. Longitudinal stability beyond the window is untested [S1][S2].

Sources

• [S1] NBER working paper (OpenAI + Harvard), “How People Use ChatGPT,” May 2024-Jun 2025 sample. Method: privacy-preserving analysis of ~1.1M consumer ChatGPT conversations; automated classification for topic/intent; user feedback signals; not peer-reviewed.
• [S2] OpenAI blog, “How people are using ChatGPT.” See OpenAI’s platform metrics in its research. Platform figures cited: 700M+ weekly active users by July 2025; ~2.5B messages/day. Caveat: company-reported, not audited.