Everyone Measures How Much People Use AI. No One Measures What Happens After.

There's a lot of data on how many people use ChatGPT. Almost none on what happens to the good answers once they're generated.

What's actually been measured

The last year produced some solid research on AI usage.

OpenAI, working with a group of economists, analyzed 1.5 million ChatGPT conversations and found that by mid-2025, work-related messages had shifted toward "doing" tasks (drafting, coding, analyzing) over just asking questions, and that non-work use had grown from about half of messages to more than 70% of them. (OpenAI / NBER, Sep 2025)

Anthropic runs a similar analysis on Claude conversations, tracking which tasks people bring to it and how that mix shifts over time. (Anthropic Economic Index)

Pew found that by early 2025, a third of US adults had used ChatGPT at least once, and the share of employed adults using it for work had jumped 20 points in a year. A follow-up study found usage had climbed further since. (Pew Research, Jun 2025)

Gallup tracks the same trend in the workplace: AI use at work rose from 40% to 45% of employees in a single quarter of 2025, more than double what it was two years earlier. (Gallup, Dec 2025)

And Stack Overflow's 2025 developer survey, with nearly 50,000 respondents, found 84% of developers now use or plan to use AI tools, up from 76% the year before. It also found something less flattering: trust in AI output accuracy dropped, from 43% to 33%, and two-thirds of developers said they're frequently frustrated by answers that are almost right but not quite. (Stack Overflow, Dec 2025)

All of this is good, rigorous work. None of it answers a much simpler question.

What nobody's measured

What do people actually do with a valuable AI answer five minutes after they get it?

Do they save it? Where? Do they ever find it again? Or does it sit in a chat history until it scrolls out of reach and they end up asking the same question a second time?

Search for research on this and you find a lot of confident-sounding claims. Vendors say people lose "hours a year" re-explaining context. Blog posts say most people "treat ChatGPT history like a disposable notepad." A widely recycled statistic says knowledge workers spend a fifth of their week searching for information, but that number is from a 2012 McKinsey report, written years before anyone had a chatbot open in a second tab. It gets cited as if it's current. It isn't.

Strip out the recycled numbers and the vendor claims, and there's no actual study of how people capture, organize, or retrieve knowledge from AI conversations specifically. Not from the AI labs, not from Pew or Gallup, not from any of the dozen or so browser extensions built to solve exactly this problem.

That's a real gap, not a rhetorical one. It would take a proper survey, and ideally some interviews, to fill it honestly. Nobody's done that yet, including us.

What exists to solve it, without data on whether it works

In the meantime, people have improvised a handful of ways to hang onto AI output:

• Native tools. ChatGPT and Claude both added memory and project features, which help within a single tool but don't do much for anyone using four or five different models a week, which the Stack Overflow numbers suggest is common.
• Copy-paste into notes apps. Notion, Obsidian, and similar tools work, but nothing about them is built for AI conversations specifically, so it's a manual habit that has to be kept up every time.
• Screenshots. Fast, unsearchable, and disconnected from the original conversation.
• Browser extensions. A growing list of them (Echoes, Superpower ChatGPT, LLMnesia, and others) each try to solve some slice of saving, searching, or exporting AI chats. None of them, as far as we could find, has published data on whether people who install them actually retrieve and reuse more of what they save.

That last point cuts both ways. It means the tools are mostly built on intuition rather than evidence, ours included. It also means whoever runs the first real study on this gets to define the terms everyone else argues about afterward.

Where this leaves things

We build Savelore, a local-first archive for saving and retrieving conversations across nine AI platforms including ChatGPT, Claude, and Gemini, so we have an obvious stake in this question. That's exactly why we're not going to publish a "state of the industry" number we don't have.

What we do think, based on what's actually documented above: AI conversations have quietly become a place where real thinking happens, at a scale that's grown fast and keeps growing. Whether that thinking survives past the browser tab it was generated in is still an open question. Somebody should go measure it properly. We're planning to be the ones who do.