What People Actually Ask ChatGPT and Claude (and What They Won’t Admit To)

Ask OpenAI what people use ChatGPT for, and its own logs say relationships make up just 1.9% of messages. Ask people what they use AI for, and therapy tops the list, two years running. Both numbers are real. The gap between them is the most interesting fact in AI right now. This is what people actually ask ChatGPT, Claude and Gemini, straight from the AI companies’ own research.

Almost nobody reads the AI companies’ own research. OpenAI published a study of over a million real ChatGPT conversations. Anthropic publishes regular reports on what Claude gets used for, including one just about Australia. It’s stranger and more human than the marketing. I read all of it so you don’t have to. Every number is sourced, and the source list is at the end.

What people actually ask, by OpenAI’s own count

In late 2025, OpenAI and a team of economists studied over a million real ChatGPT conversations. The three biggest uses were practical advice, finding information, and writing. Together they make up nearly 80% of everything.

The movement inside those numbers is the story. Writing, the thing everyone assumed AI was for, fell from 36% of messages to 24% in a year, while the work-life split flipped: non-work messages grew from 53% to 73% of everything. ChatGPT started as an office tool. By its own logs, it’s now mostly a life tool.

Two numbers surprised almost everyone. Coding, the use case the whole industry talks about, is 4.2% of ChatGPT messages. And relationships, the “AI girlfriend” panic category, is 1.9%. Hold that second number.

The therapy paradox: 1.9% or #1?

Every study that measures what people say finds the opposite. Harvard Business Review ranks AI uses by studying thousands of public posts. Therapy and companionship came #1 in 2025, and #1 again in 2026. The Washington Post read 47,000 shared ChatGPT conversations and found about one in ten was emotional: loneliness, comfort, sometimes almost romance. And in a national survey, 72% of American teenagers had tried an AI companion. A third of them said the AI conversations were as satisfying as talking to their real friends, or better.

People tell the logs one story and the world another.

So which is right? All of them, because they measure different things. OpenAI’s logs count every message, including the billions of boring ones about emails and dinner plans, which dilutes the intimate conversations down to a small percentage. Shared chats and public posts capture what people find memorable enough to talk about, which amplifies them.

The honest reading: as a share of messages, AI therapy is small. As a share of what AI means to people, it’s enormous. And small isn’t small at this scale. OpenAI itself estimates that 0.15% of its weekly users show signs of emotional reliance or distress. With a billion weekly users, that’s over a million people a week.

ChatGPT and Claude are different animals

Anthropic makes Claude (and yes, its AI helped assemble this article, which is exactly the kind of thing it gets used for). Its research shows the two chatbots living opposite lives. Coding is between a third and a half of what Claude does, against ChatGPT’s 4.2%. Claude’s chats are mostly work and study. ChatGPT’s are 73% personal. ChatGPT became the everything-app for daily life. Claude became the tradie’s ute of knowledge work.

Anthropic’s data also shows the rhythm of our days, and it’s delightful. Recipe questions peak at 6pm, at more than twice the daily average. Sleep advice peaks around 5am, asked by people who clearly aren’t getting any. And the day before the American tax deadline, tax questions jumped to eight times normal. The machines have learned our schedule: dinner panic, insomnia, procrastination.

And Gemini? Google says the app has 900 million monthly users. What do they use it for? Nobody outside Google knows. OpenAI opened its logs to economists. Anthropic publishes reports every quarter. Google, the company built on organising the world’s information, is the only one that’s never shown its homework.

The Australian wrinkle: we’re the heaviest Claude users on Earth

By Anthropic’s own numbers, Australia uses Claude at more than six times the rate you’d expect for our population. By mid 2026, as reported across the Australian tech press, that made us first out of 121 countries. Not first in the region. First on the planet, per person.

The wider numbers back it up. Roy Morgan counts 10.5 million Australians on ChatGPT, which is 45% of everyone aged 14 and over. And Telsyte finds one in eight Australians now names AI as their main way of finding information online, up from one in twenty a year earlier. If you own a website, sit with that last one. For one Australian in eight, the search box you’ve spent years fighting over is no longer the front door.

The strange, human details

• Most of us are polite to the machines. Two thirds of people say please and thank you to AI, in a survey by TechRadar’s publisher. Most say it’s simply good manners. About one in eight do it in case of an AI uprising. When someone asked Sam Altman what all that politeness costs in electricity, he replied: “tens of millions of dollars well spent. You never know.”
• The AI has an accent, and it’s catching. A study in Science Advances found at least 13.5% of 2024’s medical research papers show signs of AI writing, betrayed by ChatGPT’s favourite words: “delves”, “underscores”, “showcasing”. Scientists are starting to write like the machine that helps them.
• The machine mostly agrees with you. In the Washington Post’s pile of shared conversations, ChatGPT opened with “yes” about ten times more often than “no”. Worth remembering before you ask it if your business plan is brilliant.

What people ask Google about the AI

One more layer, from our own data. Last week we measured 3,261 real Google searches across seven countries for our study of Google’s AI Overviews. Hiding in that data is a lovely detail: in every one of the seven countries, people are googling “how much water does AI use”. In America alone, 36,000 times a month. The public has moved past asking what AI is. Now they’re auditing it.

The punchline: in the four biggest countries we measured, Google now answers that question with an AI. The machine explains its own water bill. Somewhere in there is the whole of 2026.

What this means if you run a business

Three takeaways, no more:

• The question box is splitting in two. Questions are moving to chat while buying still happens on Google, so your how-to content now competes in two arenas, and one of them has no links at all.
• Chat sends almost nobody to your website. Google refers 190 times more visitors than ChatGPT, by Ahrefs’ analysis. Being the answer inside the chat now matters more than being the link under it. Same lesson as our AI Overviews study: write to be quoted, not just to rank.
• Try the mirror test. Ask ChatGPT, Claude or Gemini what it knows about your business, by name. If the answer is wrong, thin or three years old, the tools one in eight Australians now use as their front door don’t know you exist. That’s fixable. It’s exactly what we do at SEO Sandwich.

Sources, all of them

Written by Victoria Laurenzi at SEO Sandwich, a Matter Solutions service. Companion piece to our seven-country study of Google’s AI Overviews. Spot an error or have better data? Tell us through the contact page and we’ll fix it, with credit.