How people use ChatGPT: NBER study reveals usage patterns

I wouldn't personally take the percentages shown as facts, but you can take a general picture of the most popular usecases. Obviously the market is mature enough that most software developers are running away from GPT models to better dev focused models.

Public-facing LLMs will likely become the “idea machines” best suited for creativity, conversation and exploration. In these cases, non-deterministic outputs aren’t a flaw, they’re a feature (classic transformer architecture). I think brainstorming, tutoring, roleplay, or writing feel more valuable when the answers aren’t identical every time. On the other side, smaller or specialized language models (SLMs), often connected to deterministic computational engines, will dominate in areas where precision matters i.e. programming, math, compliance, data analysis, or industry-specific use cases. Here, correctness can’t be optional. The future will probably be hybrid - LLMs handling natural language fluency, with deterministic systems running in the background to ensure accuracy where it’s non-negotiable. #SLM #LLM #Deterministic #NonDeterministic

🤖 The State of the ChatBot 🤖 🥇With 46.59B visits, ChatGPT accounts for more than 83% of total traffic among the top 10 chatbots. 🥈The second most-used chatbot, DeepSeek at 2.74B visits, has barely 6% of ChatGPT's traffic. While traffic is concentrated, the list includes a mix of U.S. 🇺🇸, Chinese 🇨🇳, and European 🇪🇺 players.

v interesting (c/o Shelly Palmer) "The National Bureau of Economic Research working paper reveals that 70% of ChatGPT usage is personal, not professional ... “Seeking Information” messages now account for 24% of conversations, up from 14% a year earlier. The researchers explicitly state this “appears to be a very close substitute for web search,” which is Google’s nightmare scenario realized. https://lnkd.in/gv3t5Fqi

Two ChatGPT’s walk into a bar. One says turn left. The other says make a U turn. I went in circles. I always bounce between the app and the web but this time the thread kept looping. Then I noticed the web was set to ChatGPT 5 and the app was on ChatGPT 4. Two models in one conversation is a recipe for confusion for ChatGPT and for me. It’s like talking to Dr Jekyll and Mr Hyde! Here is What works • Pick one model for the full thread • If you switch models, start a fresh chat and paste a short recap • If answers repeat or slow down, reset and restart Bartender’s rule. No mixed models. One pour at a time.

Your customers learned to expect more from digital conversations the moment they first chatted with ChatGPT. Now they open your widget expecting intelligence, speed, and familiarity - and within 3 seconds, they've already decided if you're worth their time. That's why we rebuilt our widget from the ground up. Not a revolution, but a thoughtful response to what customers actually want: ⚡ 70% faster load times (because every second counts) 📱 Finally, a chat that actually works on mobile 🎨 Full customization without touching code 🗣️ Text-to-voice switching built right in The bar for "good enough" has shifted. Widget 2.0 helps you clear it. Read how we're rethinking customer conversations → https://hubs.la/Q03JWvSj0

In July 2025, 18 billion messages were being sent each week by 700 million commercial ChatGPT users*. In less impressive terms, a mean of 3.6 messages sent per user per day. Now think about how that distribution might look like. *Commercial ChatGPT user behavior is honestly what we care about more.

Once upon a time, ChatGPT recorded 200 million 'monthly' users. Fast forward a year later, over 100 million 'weekly' users. 📌. Key Take Away: AI search are really closing up on search engine use.

Al's Hidden Hand: How ChatGPT-5 Quietly Erodes Independent Thinking Rising concerns over ChatGPT-5 are sparking debate among Al researchers, who warn that the latest model's seamless operation could accelerate "agency decay" - the gradual loss of independent thinking skills. According to Forbes, the unified architecture of ChatGPT-5 removes a crucial moment of user decision-making that earlier models preserved. Instead of prompting individuals to choose... https://lnkd.in/dNMteYm8

This chart is a breakdown of overall ChatGPT usage. Lots of interesting info here. I was happy to see Asking About The Model as a category. It’s not huge but it’s still good to know I’m not the only one that asks ChatGPT how it works and what it can do 😂. Does anything else stand out to you? Additional information available here - https://lnkd.in/eDeUp9Xp