Where the Term Result Card Came From

Deliverables

• Full research dossier — Markdown

658 lines covering the taxonomy, consumer use, workplace use, platform scale, study methodology, limitations, and research audit.

• Complete fan-out query log — CSV

All 64 search queries issued during the research.

• Research source register — CSV

113 unique retained URLs, each labeled as:

• USED_CORE
• USED_CONTEXT
• CONSIDERED
• CONSIDERED_NOT_USED
• CONSIDERED_FAILED_OPEN
• CONSIDERED_FAILED_DOWNLOAD

• Complete research bundle — ZIP

Important audit limitation

The source register contains every unique URL retained in the search and browsing context. The research interface does not provide a raw export of every transient search-result card, so I cannot honestly guarantee a byte-for-byte list of every result that briefly appeared.

A small number of low-value or irrelevant result cards were screened out before their URLs were retained. The dossier lists their titles and publishers separately in Appendix C, rather than falsely claiming that their URLs were captured.

The audit provides the reproducible external process—queries, pages discovered, pages opened, access failures, inclusion decisions, methodologies, and citations—but not private token-by-token internal reasoning.

Main conclusion

Your initial mental model is substantially correct, but several of the things you identified operate at different conceptual levels.

For example:

• Search is an information-acquisition outcome.
• Research is a multi-stage workflow involving retrieval, evaluation, synthesis, and usually writing.
• Learning is an outcome involving adaptive explanation.
• Brainstorming is an interaction method.
• Writing, code, slides, images, and video are artifact formats.
• Advice and decision support are judgment-assistance use cases.
• Therapy-like conversation is a reflective, advisory, and social interaction.
• Agents are a division-of-labor model in which the system performs actions rather than merely responding.

The cleanest framework therefore has two axes.

Axis 1: What result does the user want?

Axis 2: How is the work divided?

This resolves several categorization problems. Brainstorming is not necessarily a separate subject category; it is a collaborative mode that can be applied to products, business strategy, fiction, relationships, code, or travel. Likewise, “creative work” is too broad: writing a report, generating code, designing slides, and creating an image all produce artifacts, but the underlying purposes can be completely different.

What actual usage telemetry says

1. ChatGPT: predominantly personal, advisory, and practical

OpenAI’s September 15, 2025 study classified 1.5 million consumer ChatGPT conversations using automated privacy-preserving methods. It reported:

• 49% Asking
• 40% Doing
• 11% Expressing
• Approximately 70% non-work
• Approximately 30% work-related
• About three quarters concerned practical guidance, seeking information, or writing

OpenAI specifically concluded that users were obtaining substantial value from ChatGPT as an advisor and decision-support system, rather than using it only to generate finished outputs. Researchers did not manually read the messages. (OpenAI)

The fine-grained distribution associated with that research was approximately:

The complete distribution and the source trail are in the dossier.

Interpretation

The dominant ChatGPT use case is not simply “replace Google.” It is closer to:

Help me obtain information, understand it, decide what it means, and convert it into something useful.

Traditional search often appears as one stage inside that process.

2. Google: even more heavily non-work, but with different included surfaces

Google released ATLAS v1.0 on July 23, 2026. It analyzed 15 million aggregated and deidentified interactions from the Gemini App, AI Mode, and Gemini API, spanning more than 150 countries, 140 languages, approximately 800 occupations, and 4,000 tasks.

Its central findings were:

• More than 86% of interactions occurred outside work.
• Work use appeared across 68% of occupations, representing roughly 90% of U.S. employment.
• In the typical occupation, AI was observed in only about 21% of tasks.
• Less than 10% of work interactions fully automated a task.
• Work interactions disproportionately involved ideation, strategy, information retrieval, learning, creative design, and hypothesis testing.
• Non-routine cognitive tasks represented 65% of observed work interactions, versus about 35% in Google’s economy-wide task baseline.
• Manual and technical workers were twice as likely to use multimodal features, including for diagnostics and troubleshooting. (blog.google)

This produces a useful description of current workplace adoption:

Broad across jobs, shallow within jobs, and primarily collaborative rather than fully automated.

ATLAS does not cover all Google AI usage. It excludes Google Workspace, Gemini Enterprise, Google Translate, AI Overviews, and several other major Google products. The Google numbers therefore cannot be treated as an overall market distribution. (blog.google)

3. Claude: more professionally skewed and artifact-oriented

Anthropic’s audience is more technical and professionally concentrated than the general population, which materially affects its usage mix.

Anthropic’s June 26, 2026 Economic Index found:

• Personal use accounted for approximately 35% of Claude conversations on weekdays.
• Personal use rose to just under 50% on weekends.
• Work prompts shifted toward business correspondence, marketing content, documents, and presentations.
• Weekend activity shifted toward emotional support, medical questions, recipes, and investment advice. (Anthropic)

Anthropic classified outputs as artifacts rather than only prompt topics. It found that:

• Work conversations most frequently produced:
  • Documents and reports: 20%
  • Explanations: 9%
  • Email drafts: 7%
  • Analyses and summaries: 6%
• Personal conversations most frequently produced:
  • Explanations: 25%
  • Recommendations: 22%
• Marketing content was 80% work-related.
• Blogs and articles were 81% work-related.
• Database queries were 82% work-related.
• Plans and strategies were split between 44% work and 49% personal.
• Translation was split between 42% work and 44% personal. (Anthropic)

This shows why output-based categorization alone is insufficient. A “plan” might be a corporate strategy, workout plan, travel itinerary, or content calendar. The artifact is the same; the context and outcome differ.

Consumer and personal use cases

The evidence supports nine major consumer families.

1. Information and research

Examples include:

• Factual questions
• Product and service comparisons
• Travel research
• News and current events
• Recommendations
• Identifying something from a photograph
• Combining information from multiple sources
• Investigating an unfamiliar subject

This is the category most directly substituting for traditional search, although it frequently adds explanation and synthesis.

2. Learning and education

This includes:

• Tutoring
• Concept explanations
• Follow-up questions
• Language learning
• Homework support
• Test preparation
• Adaptive examples
• Asking for a simpler or more advanced explanation
• Learning software, technical skills, or hobbies

Learning is meaningfully different from content creation. The model is not merely generating an educational article; it is adapting the explanation through interaction.

3. Practical guidance and troubleshooting

Examples:

• Recipes and cooking
• Home repair
• Technology troubleshooting
• Appliance use
• Vehicle questions
• Forms and government procedures
• Personal administration
• “How do I do this?” instructions

Google specifically found substantial non-work use involving household tasks and difficult administrative processes such as taxes, licensing, and fines. (blog.google)

4. Planning and decision support

Examples:

• Travel itineraries
• Shopping decisions
• Financial scenarios
• Schedules
• Goal planning
• Comparing alternatives
• Preparing for an event
• Thinking through a life decision

This category is likely underappreciated when LLM use is described merely as either search or creation.

5. Personal communication

Examples:

• Text messages
• Emails
• Applications
• Complaints and support requests
• Dating-profile text
• Social posts
• Translation
• Editing for clarity or tone
• Summarizing something for another person

6. Creative and technical making

Examples:

• Stories and fiction
• Images and video concepts
• Music ideas
• Personal websites
• Hobby programming
• Games
• Crafts
• Naming and design ideas

“Programming” is professionally important but represented only about 4% of the broad ChatGPT consumer telemetry cited above. It becomes much larger in developer-specific products and technically concentrated audiences.

7. Health and self-management

Examples:

• Symptom questions
• Medical terminology
• Diet and fitness
• Exercise planning
• Sleep advice
• Preparing questions for a doctor
• Understanding test results
• Habit formation

These queries range from low-risk education to consequential medical advice, so measuring the category does not imply that all such use is appropriate or reliable.

8. Reflection, emotional support and companionship

Pew’s February 17–23, 2026 nationally representative survey of 5,119 U.S. adults found:

• 10% had used chatbots for emotional support or advice.
• 4% had used them for companionship.
• By comparison, 42% had used them to search for information.
• Among employed adults, 38% reported using them for work.
• 20% reported medical-advice use.
• 20% reported diet or fitness use. (Pew Research Center)

This means therapy-like use is real and significant, but it is nowhere near the largest general-population category.

Claims that “therapy is the number-one generative AI use case” usually derive from qualitative rankings of online discussions, not representative population surveys or conversation telemetry. Such rankings measure cultural salience and repeated discussion, not share of users or prompts.

9. Emerging personal agents

This includes systems that:

• Search multiple sources
• Manipulate files
• Organize information
• Use calendars
• Conduct shopping or travel research
• Complete administrative workflows
• Maintain longer-running projects

This remains an emerging category. Current telemetry still suggests collaboration is far more common than complete autonomous execution.

Professional and workplace use cases

Professional LLM adoption extends substantially beyond generating code, presentations, and documents.

1. Research and organizational knowledge

• Web research
• Competitive intelligence
• Market research
• Internal document search
• Policy and regulatory research
• Technical research
• Source comparison
• Literature reviews
• Customer and industry monitoring

2. Writing and communication

• Email
• Reports
• Proposals
• Marketing content
• Policies
• Job descriptions
• Sales materials
• Documentation
• Executive communication
• Social content

3. Information transformation

• Summarization
• Meeting notes
• Extraction
• Classification
• Translation
• Document comparison
• Restructuring unorganized material
• Converting prose into tables or structured data
• Adapting material for different audiences

This may be one of the most pervasive professional families because it applies to nearly every information-heavy role.

4. Ideation and strategy

• Business ideas
• Product features
• Campaign concepts
• Alternative strategies
• Scenario generation
• Problem framing
• Naming
• Experiment design
• Critiquing assumptions

Your description of back-and-forth product brainstorming belongs here. It is a genuine use case distinct from looking up an answer or requesting a finished artifact.

5. Analysis and decision support

• Data interpretation
• Financial analysis
• Risk identification
• Forecasting
• Prioritization
• Root-cause analysis
• Scenario analysis
• Evaluating alternatives
• Issue spotting

6. Software and technical work

• Code generation
• Debugging
• Testing
• Refactoring
• Documentation
• Database queries
• Architecture
• Code review
• Migration assistance
• Agentic implementation

This category is extremely intensive among developers but not dominant across the total workforce.

7. Presentations, design and media

• Slide decks
• Diagrams
• Spreadsheets
• Charts
• Images
• Video and audio
• Mockups
• Visual explanations
• Document layout

8. Learning and procedural assistance

• Employee onboarding
• Training
• Understanding systems
• Explaining regulations
• Troubleshooting
• Learning unfamiliar tools
• Generating SOPs
• In-the-moment procedural guidance

9. Sales, service and customer operations

• Customer-support responses
• Support agents
• Sales preparation
• Call summaries
• Lead research
• Proposal personalization
• Objection handling
• Customer classification
• Knowledge retrieval during interactions

10. Operations and administration

• Scheduling and planning
• Forms
• Procurement
• Human-resources processes
• Compliance workflows
• Project management
• Reporting
• Standard operating procedures
• Routine internal administration

11. Physical and field work

Google’s ATLAS data found use among automotive technicians, industrial mechanics, and other trade workers for:

• Interpreting diagnostic results
• Debugging wiring
• Examining equipment
• Troubleshooting
• Generating or interpreting diagrams
• In-the-moment learning

This category will be missed by research that equates AI work exclusively with white-collar writing and coding. (blog.google)

12. Agentic execution

• Coding agents
• Research agents
• Multi-system workflows
• Tool invocation
• Autonomous testing
• File manipulation
• Support and sales agents
• Repetitive operational processes

The distinction is important: generating a procedure is assistance; carrying out that procedure is execution.

What workers actually report

Gallup surveyed 23,068 employed U.S. adults from August 5–19, 2025. It found:

• 45% used AI for work at least a few times per year.
• 23% used it at least a few times per week.
• 10% used it daily.
• Among workplace AI users:
  • 42% used it to consolidate information.
  • 41% used it to generate ideas.
  • 36% used it to learn.
  • 61% used chatbots or virtual assistants.
  • 36% used writing or editing tools.
  • 14% used coding assistants. (Gallup.com)

Pew’s workplace research found that among workers who had used chatbots on the job:

• Research or finding information: 57%
• Editing written content: 52%
• Drafting reports or documents: 47%
• Summarizing documents or meetings: 40%
• Generating new ideas: 35%
• Data analysis or programming: 27%
• Creating or editing images or video: 21% (Pew Research Center)

The practical conclusion is:

Across the workforce, LLMs are currently more broadly used for research, writing, compression, ideation, and learning than for programming.

A second useful consumer survey

Ipsos surveyed a probability-based, nationally representative sample of 2,021 U.S. adults from March 3–6, 2026. Among people who used AI:

• Information or recommendations: 80%
• Writing or editing: 59%
• Advice or learning: 55%
• Brainstorming: 53%
• Image creation: 44%
• Data analysis or programming: 37%
• Direct typed prompting: 75%
• AI-assisted web search: 48%
• File or document uploads: 41%

Among employed AI users, 51% used AI at least partly for work. Of those using AI for work, 50% relied solely on personal subscriptions or free access, while 33% had access to an employer-provided service. (Ipsos)

This is one of the few available figures that says anything useful about personally funded/free access versus employer-provided access. It still does not distinguish free logged-in, paid personal, and logged-out traffic.

Platform size and segmentation

Best comparable U.S. platform measurement

Pew asked U.S. adults whether they had ever used specific assistants:

The survey was conducted February 17–23, 2026. Respondents could use multiple systems, so these are overlapping reach figures, not market shares totaling 100%. (Pew Research Center)

Current first-party scale figures

OpenAI reported in February 2026:

• More than 900 million weekly active ChatGPT users
• More than 50 million consumer subscribers
• More than 9 million paying business users
• Approximately 1.6 million weekly Codex users (OpenAI)

Dividing 50 million subscribers by 900 million weekly active users gives approximately 5.6%, but that should not be described as the paid conversion rate. The numerator and denominator refer to different concepts and potentially different user populations.

Google reported:

• AI Overviews: more than 2.5 billion monthly active users
• AI Mode: more than 1 billion monthly active users
• Gemini App: approximately 950 million monthly active users (blog.google)

Those Google figures overlap heavily and include different product surfaces. AI Overviews may be encountered as part of ordinary Search, whereas ChatGPT and the Gemini App require more explicit assistant use.

Anthropic has not publicly disclosed an official active-user figure that is directly comparable with ChatGPT WAU or Gemini App MAU. Third-party traffic estimates exist, but they do not provide a sufficiently clean answer to this question.

Free, logged-out and paid segmentation

The requested breakdown does not currently exist publicly in a reliable, comparable form.

What is available:

• OpenAI’s Signals consumer dataset covers Free, Go, Plus, and Pro messages and supports analysis by plan type.
• Public documentation does not disclose what percentage of the underlying user population belongs to each plan.
• Logged-out ChatGPT use is not separately quantified.
• Google does not provide a public free-versus-paid usage distribution across AI Mode, AI Overviews, and the Gemini App.
• Anthropic does not publish a comparable free-versus-Pro-versus-Team population breakdown.
• Employer-provided access can be partially estimated through surveys such as Ipsos, but that is not equivalent to account-plan telemetry.

Therefore, a rigorous platform table should leave those cells as unknown, rather than inventing shares from web traffic or subscriber totals.

What is changing

The evidence suggests five directional changes.

1. From answer retrieval toward judgment support

Users increasingly ask systems to compare, diagnose, plan, recommend, and evaluate—not simply retrieve a fact. OpenAI’s Asking category already exceeds Doing. (OpenAI)

2. From single prompts toward iterative workflows

The user asks a question, evaluates the response, supplies more constraints, requests alternatives, and converts the result into an artifact. This makes “conversation” itself part of the functionality.

3. From generic chat toward embedded use

AI is increasingly encountered within:

• Search
• Office software
• Coding environments
• Customer-service systems
• Enterprise knowledge systems
• Mobile operating systems
• Creative tools

This complicates measurement because users may use an LLM without conceptualizing the action as “using a chatbot.”

4. From text-only use toward multimodal assistance

Images, files, screenshots, voice, diagrams, video, and structured data expand LLM use into troubleshooting, physical work, document analysis, and creative production.

5. From assistance toward execution—but not yet predominantly

Agentic coding, research, file operation, and workflow execution are growing rapidly. However, Google’s current large-scale work telemetry still finds complete task automation in fewer than 10% of work interactions. (blog.google)

The most defensible high-level segmentation

For your own taxonomy or research database, I would use these top-level categories:

1. Information acquisition and research
2. Learning and understanding
3. Advice, planning and decision support
4. Ideation, exploration and critique
5. Writing, communication and content transformation
6. Technical, analytical and creative artifact production
7. Procedural assistance and troubleshooting
8. Workflow execution and automation
9. Social interaction, reflection and emotional support
10. Entertainment, simulation and play

Then assign separate fields for:

• Personal, professional, coursework, or mixed context
• Ask, collaborate, transform, delegate, execute, or express
• Text, code, document, spreadsheet, slide, image, audio, video, structured data, or action output
• Single-turn, iterative, or long-running agentic workflow
• General assistant, search surface, API, embedded product, or specialized application
• Free personal, paid personal, employer-provided, institutional, API, or unknown access

That structure should remain useful even as the relative shares change every six months.

Research quality warning

The numbers in this field answer different questions and should not be merged into one pie chart:

• Conversation telemetry measures messages, conversations, or interactions.
• Surveys measure people who have ever done something.
• Workplace surveys may measure workers, AI users, or workers who have used chatbots at work.
• Platform figures measure weekly active users, monthly active users, subscribers, accounts, or feature exposure.
• Product datasets include different surfaces and different user populations.
• Categories frequently overlap in surveys but are mutually exclusive in automated prompt classifications.

The closest thing to an actual prompt-share distribution is OpenAI’s consumer telemetry. The closest current cross-surface workplace telemetry is Google ATLAS. The strongest population-level U.S. comparison is Pew. The best output-oriented view is Anthropic’s artifact classification. None of those alone describes the whole market.