Here’s a study of what people use Gemini for. It covers the consumer app, the AI Mode thing in Search, and the free tier of the API.[1] Free-tier and paying consumers are in there, corporate seats are probably a small minority, if at all.[2]
So you’re getting a consumer look at usage. Other companies’ AI products are missing too, obviously – no Anthropic, no OpenAI.
The study probably also misses most regulated work. For example, a hospital that’s using AI in compliance with regulations like HIPAA is probably doing it somewhere this study can’t see. Same for regulated banks, spies and soldiers, and Europe – whatever regulated work you want. Of course, there’s shadow AI, which often accounts for a shit-ton of enterprise use at this phase of any technology.
More in the methodology section below.
But these are still great insights into how people use AI!
What does it mean?
First, weighted heavily toward consumer use, the biggest category is socializing, relaxing and leisure at 26%. That’s bigger than education, and it’s stuff like people asking about TV, games, books, and hanging out. The study is probably pulling in a lot of Google searches that end up in Gemini chat when,[3] like, people ask what Tom Holland’s shoe size is in US measurements.
After that, what you can take away is that people are using AI for learning and answering questions, handling paperwork and writing basic stuff[4], and otherwise taking care of bullshit and bureaucracy.
As Claude put it: “The over-representation is concentrated in high-friction bureaucracy.”
This is great! Who wants to do that crap, or at least, spend a lot of time and bottleneck on it.
It’s also good to focus in on what people who are using Gemini for work are doing. Again from Claude: “within the 13.5% of conversations that are work, 65% is the non-routine, judgment-y stuff.” That works out to about 9% of everything.
More from Kyle Orland at Ars:
Their initial review of the data finds that, while AI sees some significant use across a wide variety of occupations, that use “remains shallow and overwhelmingly collaborative in nature, with end-to-end task automation limited in scope.” … For many occupations (29%), not a single relevant work task achieved this “non-negligible” Gemini usage threshold, suggesting those jobs have been minimally impacted by the AI revolution so far. For another 30 percent of all occupations, less than one-quarter of tracked tasks saw significant related Gemini usage, suggesting humans were still the ones responsible for the vast majority of the component parts of those jobs.
The rest of this post was written summarized and by Claude.
Uses
Here are the uses the study found:
What people actually ask about
Everything below is share of conversations, not share of people. One person with forty questions about their tax residency outweighs forty people asking one thing each about dinner. Keep that in mind when the numbers look weird.
| Rank | What it is | Share of all conversations |
|---|---|---|
| 1 | Socializing, relaxing, leisure – TV, games, reading, hanging out | 26.4% |
| 2 | Education – classes, homework, learning something | 20.7% |
| 3 | Work, all of it | 13.5% |
| 4 | Household activities – managing, budgeting, chores, repairs, garden | 11.0% |
| 5 | Personal care, including health questions about yourself | 9.0% |
| 6 | Everything under 2%, pooled | 6.6% |
| 7 | Consumer purchases – comparison shopping, researching a buy | 5.2% |
| 8 | Professional services – medical, legal, financial | 4.8% |
| 9 | Traveling | 2.7% |
That row 6 is where the government and bureaucracy stuff hides. It gets twenty times more conversation than it gets human hours, which sounds enormous until you notice it’s still under 2% of everything said.
Inside that 13.5% of work, what people are asking for:
| What they want | Share of work conversations |
|---|---|
| A draft or a template they’ll still have to fix | ~40% |
| Look this up, explain this, teach me this | ~30% |
| Think this through with me | ~16% |
| Just do the whole thing | ~10% |
| Check what I already wrote | ~4% |
So roughly 1.4% of everything anyone says to Gemini is an attempt to hand over a work task completely. The split by kind of task is where it gets interesting: routine cognitive work hits 27% automation intent, non-routine cognitive drops to 6.5%, and manual work is 82% just asking questions.
Which jobs. Computer and Mathematical is about 29% of US work conversations, Business and Financial about 22%, Office and Admin Support 12%, arts and media 9.5%. Financial analysts, market researchers, software developers, and sysadmins are the most over-represented individual jobs.
Blue collar isn’t absent, it’s differently shaped. Over ten thousand conversations from auto techs testing components, rewiring, inspecting parts for wear. Thousands from industrial mechanics reading error messages. Those folks send photos at more than twice the rate of everyone else.
And the bit nobody quotes: usage skews toward the least expert cognitive tasks. Rewriting a press release into another language. Drafting a product spec. Not the hard parts.
How they measured it
14.65 million conversations, two weeks in April 2026, across the Gemini app, the AI Mode thing in Search, and free-tier API. Queries under ten tokens dropped.
The unit is the conversation. Not the person. There’s no step anywhere in this where they build a profile of what Coté does with Gemini and then add up a million of those. Conversations get stripped of personal info, summarized, embedded, clustered against similar conversations, and then mapped onto taxonomies that already existed: BLS occupation codes, O*NET tasks, the American Time Use Survey. Then they count.
They can’t do it any other way. The privacy design replaces internal IDs with, in their words, “mathematically unlinked” UUIDs specifically so nothing joins back to product logs. Users show up only as thresholds – ten unique people to keep a cluster, twenty-five to count a task, fifty to count an occupation. Nobody at Google reads a conversation. The original text gets thrown away.
Which also means no per-person normalization at all. Heavy users are simply weighted more, because they generate more rows.
Accuracy falls off a cliff as you get specific: 71.6% at the broad occupation group, 42.5% at the job title, 22.6% at the individual O*NET task. Human raters called 86% of even the granular labels “reasonable,” which is a nice save but not the same thing. This is why the paper keeps retreating to broad categories and to presence-versus-absence rather than counts. Task saturation is binary: twenty-five people touched this task or they didn’t. A task with 100,000 conversations and one with 26 both count as one.
They know what they’re throwing away. The distribution has a skewness above 16 and the mean is five times the median. They dropped magnitude on purpose, to survive their own classifier.
What isn’t in here
Paid API text isn’t logged at all, so it can’t be classified. Vertex and Google Cloud are out. Enterprise app subscriptions are out. Workspace, NotebookLM, AI Overviews, Translate, Maps, and their agentic coding tools are all out.
That last one matters more than it sounds. Nobody automates a task end-to-end by typing into a chat window. You do it by writing a pipeline against an API and running it ten thousand times, which produces zero conversational turns. The study’s headline finding is that automation is rare, and the study is structurally blind to where automation happens. Anthropic, measuring a much more API-heavy crowd, reports 43-45% automation against Google’s sub-10%. Pick your denominator, pick your answer.
Europe is worse. European API content isn’t logged, so European countries get dropped from the API work analysis entirely, and the paper admits coding is therefore understated for the whole OECD. The compliance regime already ate a hole in the dataset.
And no competitors. Gemini is somewhere around 27% of global AI chatbot web traffic and about 19% in the US, running heavily mobile. Whatever the professional end of this market is doing, most of it happened somewhere this study can’t see.
Compared to Anthropic’s Study
Recently, Anthropic run a similar thing over their own data, the Economic Index, on the same O*NET task taxonomy. The findings are noticeably less reassuring. Where Google puts end-to-end automation at under 10% of judgment work, Anthropic puts it at around 45% overall, and rising: augmentation has slipped from 57% to 52% across their reports. Task coverage is wider too: about 49% of occupations see Claude used for a quarter or more of their tasks, against Google’s 30%.
Which is mostly a story about pipes. Google is measuring a consumer app, Search, and a free API tier. Anthropic is measuring a customer base that skews enterprise and API-heavy, which is where automation actually happens, because nobody automates a task by typing into a chat window.
Pick your denominator, pick your headline.
Where they line up is more interesting. Anthropic’s Cowork numbers have business process and operations at 33.4% and content at 16.4%, half of all usage, and describe it as the connective work that surrounds every role rather than the core of any one role. Google gets to the same place from the other end: drafting and looking things up is 70% of work conversations, and AI touches 21% of tasks in the median job. Two studies, opposite populations, same answer. It’s doing the glue, not the job.
Anyhow: check out the Google study.
It excludes paid API usage entirely, along with Vertex, Google Cloud, enterprise app subscriptions, Workspace, NotebookLM, AI Overviews, Translate, Maps, and their agentic coding tools. ↩︎
🤖: Where exactly this line falls is fuzzy. The study drops Gemini Apps enterprise subscription data, along with Workspace, Vertex, Google Cloud, and paid API. But it never says where “enterprise” starts. So an SMB construction firm that put everyone on Workspace is out, while the same firm expensing a handful of individual subscriptions, or just letting people use the free tier on their phones between jobs, is in and gets counted as work. Which is how a consumer product ends up with 13.5% of its conversations classified as work: the auto mechanics photographing wiring harnesses and the industrial mechanics reading error codes off a machine are not sitting on procured enterprise seats. ↩︎
As I understand it, the study excludes AI Overviews, that AI-generated summary box that now shows up at the top of a normal Google search whether you wanted it or not. What it does count is AI Mode, the conversational tab you have to opt into – either by clicking through from an Overview or just picking the tab, same as you would Images or News. ↩︎
Like me using AI to help write this piece…though, I used a Claude Max, paid plan. Of interest there, I spent a little over 90 minutes on this post from asking Claude to summarize the original 100 page PDF, discussing the survey and having it generate the list of usage, going back and forth in writing it, etc. I spent about 30 minutes wrangling the text and posting, which is awhile other annoyance. I probably would have spent half a day (3 or 4 hours) doing all of that myself. I’m of course suffering first hand reading, which means I took a narrow view of what I wanted to see in the report, and also introduced the chance of Claude hallucinating and being steered by me. ¯(ツ)/¯ ↩︎
Leave a Reply