900 million weekly users. 2.5 billion messages per day. ChatGPT is the 4th most visited website in the world, ahead of Wikipedia.
The dominant narrative frames AI as a productivity tool: write faster, code faster, automate. The data tells a different story.
In September 2025, OpenAI published with the NBER and Harvard economist David Deming the largest study ever conducted on real ChatGPT usage: 1.5 million conversations analyzed, anonymized, across consumer plans. The findings challenge the entire ambient discourse.
Half of all conversations are decisions
The study classifies usage into three modes: Asking (seeking answers), Doing (executing tasks), Expressing (reflecting and creating).
| Mode | Share | What it means |
|---|---|---|
| Asking | 49% | Questions, advice, decision support. The fastest-growing category. |
| Doing | 40% | Execution: writing, code, planning. Two-thirds of writing is editing, not creating. |
| Expressing | 11% | Personal reflection, exploration, play. |
Source: Chatterji et al., "How People Use ChatGPT", NBER Working Paper No. 34255, Sept. 2025.
The finding is clear: half the usage of a tool built to "produce" consists of asking questions. Users don't ask ChatGPT to do. They ask it to help them decide.
For brands, the consequence is direct. When a user asks ChatGPT which CRM to choose, which hotel to book, or which running shoe to buy, they're not looking for ten links to sort through. They expect an answer. And that answer includes, or excludes, your brand.
This is what Recommendation Visibility measures: how often a brand naturally appears when a user expresses intent through an AI system.
70% of conversations are personal
The AI-at-work narrative dominates media coverage. The data says otherwise.
Roughly 70% of messages are unrelated to work. This ratio was 53% in June 2024, and the personal share keeps growing. The three most frequent topics (practical guidance, information seeking, and writing) account for nearly 80% of conversations. Code represents only ~4%.
Researchers estimate a consumer surplus of $97 billion in the United States in 2024 (Collis & Brynjolfsson, 2025). Value created in daily life, outside workplace productivity.
For marketing, the implication is structural: the content that matters is no longer just the content that converts. It's the content that helps people make better decisions. Because that's exactly what they're asking AI to do. And when those decisions involve a purchase, a service comparison, or a brand choice, Recommendation Visibility becomes directly commercial.
A user base that looks like your customers
ChatGPT adoption is no longer a tech phenomenon. The user profile now mirrors the general population:
- Gender: users with female first names grew from 37% to 52% between January 2024 and mid-2025.
- Geography: 4x faster growth in lower-income countries than in the wealthiest nations.
- Age: nearly half of adult messages come from users under 26, but the gap is narrowing.
- France: 18.3 million users, 44% of the active workforce, 5th globally.
ChatGPT is no longer a niche tool. It's a mass-market decision infrastructure. And that infrastructure recommends brands in every conversation.
Three shifts for marketing
Content becomes an input
Your website content is no longer read directly by your audience. It's ingested by a model that synthesizes, compares, and delivers it inside an answer. ChatGPT-referred traffic to US retail sites converts at 11.4%, versus 5.3% for classic organic search (Conductor, 2026).
The content that wins is no longer the content that ranks best on Google. It's the content most citable by an AI.
Decisions replace attention
For 25 years, digital marketing optimized for attention: impressions, clicks, views. In a world where AI gives one answer, the game shifts.
49% of interactions are Asking. The seven most frequent professional activities in ChatGPT (obtaining information, interpreting it, documenting, deciding, advising, solving, creating) account for ~77% of professional use. ChatGPT functions as a decision-support system, not a content machine.
It's no longer about buying visibility. It's about earning it.
AI becomes the search interface
31% of ChatGPT queries trigger an integrated web search, with local intent as the primary driver, in 59% of cases (Nectiv, 2025). ChatGPT accounts for 87.4% of all AI-referred traffic (Conductor, 2026).
The classic model (type a query, get links, sort through them) is fading. The user asks a question, and the AI searches, filters, synthesizes, and recommends. Often without the user ever visiting the source site.
SEO isn't disappearing. It's transforming. The shift is from SEO to GEO (Generative Engine Optimization): being visible not in a results page, but in an AI answer. This is a visibility that the AI Visibility Index measures: a unified score of a brand's presence across the entire AI ecosystem.
What this changes for performance measurement
The NBER data confirms what teams facing search evolution have observed for a year: classic metrics are becoming insufficient.
If 49% of ChatGPT conversations are decision-making, and ChatGPT represents 87% of AI-referred traffic, then a brand absent from those conversations is losing an acquisition channel that no classic dashboard measures.
This is where Brand Demand Efficiency becomes critical. The ratio between paid spend and real organic demand must now factor in an additional variable: demand generated, or captured, by AI recommendations. A brand whose Recommendation Visibility is rising can reduce brand bidding spend with zero traffic loss. Because the demand is already there, carried by a channel that traditional tools don't see.
Nanga connects these signals: AI Visibility Index, branded search trends, paid spend, Brand Demand Efficiency. Not to produce another score. To produce a decision.
Sources:
- Chatterji, Cunningham, Deming, Hitzig, Ong, Shan & Wadman (2025). "How People Use ChatGPT", NBER Working Paper No. 34255.
- OpenAI (2025). "How people are using ChatGPT".
- OpenAI (February 2026). Announcement: 900 million weekly users.
- Collis & Brynjolfsson (2025). Consumer surplus estimate for generative AI in the United States.
- Conductor (2026). AEO/GEO Benchmarks Report.
- Nectiv (October 2025). ChatGPT web search trigger analysis.
- Sortlist (2026). ChatGPT France statistics.