Last Tuesday, a mid-market CEO I was onboarding asked me a question I've heard dozens of times over the last year.
“My IT team set us up on Copilot through the Microsoft licenses we already had. The seats are basically free. Why are you pushing us toward Claude?”
It's a fair question. I've given a fair answer for a year. I talk about model quality, about tool use, about the managed-agent architecture. I pull up benchmark comparisons. I show him Claude Code running across his own repos. By the end of the call he says something like, “okay, that's interesting,” and I leave the meeting feeling like I've done maybe 60% of the work of convincing him.
The next day, Epoch AI published a chart that did the remaining 40% in a single image.
The chart
Epoch AI Research is one of the most credible independent AI research organizations working today. On April 22 they posted a clean piece of data about US consumer AI usage. It's the kind of data point that reframes a whole conversation.
Their summary, in one sentence:
80% of US adults who report using Claude in the previous week live in households earning $100,000 or more a year, compared to 37% of Meta AI users. Other major providers cluster in a relatively narrow band, with 56–64% of users in $100,000+ households.
Two things jump out.
The first is that Claude's user base is strangely concentrated at the top of the income distribution. Four in five US Claude users live in $100K+ households. That's radically top-heavy for any consumer tech product.
The second is that Meta AI's distribution looks almost exactly like the general US population. That makes sense the moment you think about it. Meta AI is embedded in Facebook, Instagram, WhatsApp, and Messenger. Hundreds of millions of Americans use at least one of those every day, and Meta AI has slid into the interface next to the search bar whether they asked for it or not. Its user base is whoever uses Meta's apps, which is basically a cross-section of the country.
The middle band is the interesting one. 56–64% of ChatGPT, Gemini, and Copilot users are in $100K+ households. Well above the population average, well below Claude's 80%. Those products all have some consumer gravity (free tiers, browser integration, Google Workspace embeds) along with real professional adoption. Claude stands alone at the top.
Meta AI's chart is the inverse of Claude's, and that's the story
I want to linger on this for a second, because it's where the data gets most interesting.
Meta AI is one of the most heavily distributed AI products on earth. If you count total impressions, it's probably ahead of Claude by an order of magnitude. Zero friction. No signup. No new app. No subscription. You open Instagram and it's just there.
And yet: only 37% of its users are in $100K+ households. Which means 63% are below that line. The product's distribution mirrors the distribution of the country.
That's what a consumer AI looks like when it wins consumer distribution.
Now look at Claude. Claude has no consumer distribution surface. There's no Facebook-for-Claude, no Gmail plugin by default, no WhatsApp bot, no pre-loaded Instagram integration. You don't stumble into Claude. You open a browser, go to claude.ai, create an account, and start a conversation. Or you subscribe at $20 a month for Pro. Or you wire the API into a product you're already building.
Every Claude user is a user who went looking for Claude. That's the self-selection you don't get with Meta AI. And when you let a user base self-select, the distribution tells you who needed the thing badly enough to go find it.
What “$100K+ household” actually means
I want to be careful here, because income-bucket demographics are easy to misread.
A $100K+ household in the US doesn't mean rich. It means roughly the upper third of the income distribution. It's a single professional making $110K a year. It's a two-earner household with each partner pulling $60K. It's a plumbing contractor doing good work in a metro area. It's a senior engineer, a mid-level lawyer, a high school principal in an expensive district.
What it correlates with, and this is the part that matters for a business conversation, is knowledge work. Jobs where the main output is written decisions, documents, code, analysis, spreadsheets, and judgment. Jobs where an AI that's actually good at those things can double your throughput on a Tuesday afternoon.
It also correlates with decision authority over your own tools. People in that income band are more likely to pay the $20 a month for Pro themselves, without waiting for IT to approve a license. They're more likely to evaluate three AIs and pick one. They're more likely to care that the one they pick is any good.
Claude's user base looks more like the Wall Street Journal's subscriber base than like Facebook's user base. Meta AI's looks like Facebook's. That's the shape of the signal.
The operator class has already picked
Here's what I keep seeing at Settle.
We deploy Claude into traditional mid-market businesses. Manufacturers, distributors, ops-heavy service companies. My job usually starts with a CEO or a COO who is somewhere between skeptical and curious about AI. And almost every single engagement, there's one pattern I can predict before I walk in.
The people at the company who “get” AI the fastest, the ones who see the demo and immediately start connecting it to workflows I haven't even pitched yet, are almost always already using Claude. Personally. On their own subscription. For their own work.
The people who are still skeptical after the demo, or who tell me “we tried AI and it wasn't impressive,” are almost always running either ChatGPT's free tier from a year ago, or a Copilot license IT set up and nobody really uses.
I didn't design it this way. It just keeps happening.
It's not that Claude users are smarter or better. It's that they're the people who already went through the work of finding a serious AI tool. They already know what the good version feels like. So when I show them a production system, their brain has a reference point. The skeptics don't. They're still comparing what I'm building to a chatbot from 2023.
The Epoch chart is the first time I've seen that pattern show up in aggregate data.
The honest counter
I should be honest about what this chart isn't.
Income skew can mean a couple of things, and “this tool is better for serious work” is only one of them. The other obvious story is “this tool has access friction that keeps lower-income users out.”
That explanation is partially true. Claude Pro costs the same $20 a month as ChatGPT Plus, so it isn't really about subscription price. But Claude's free tier is thinner, its rate limits are tighter, its surface area is narrower. You don't bump into Claude by accident the way you bump into Meta AI. To become a Claude user, you have to care enough to go find it.
So yes, some of the 80% is self-selection driven by access friction, not pure preference. If Anthropic shipped a Gmail plug-in tomorrow and a Facebook integration next week, the distribution would slide toward the middle. I'm reasonably sure of that.
But here's the thing. That self-selection isn't a bug for a business conversation. It's the whole point.
If I'm a mid-market CEO trying to figure out which AI to deploy across my operations team, I don't care much what the average US consumer picks when the AI is handed to them for free. I care what the people doing serious knowledge work pick when they have to go find it themselves. The Epoch data is exactly that filter. It isn't a popularity contest. It's a revealed-preference signal from the subset of the population that most looks like my future operators.
So what does this mean if you're picking an AI for your business?
Here's the part I'd want a mid-market CEO to hear.
Benchmarks are useful. Model quality matters. Tool use matters. Managed-agent architecture matters. All of that is real, and you should push any vendor, including me, on it.
But there's another kind of signal that matters too, and it's usually the one missing from AI buying decisions. The question of who else is using this tool, and why.
When 80% of a major AI's user base is the slice of the country that does knowledge work and chooses its own tools, that's telling you something benchmarks can't. It's telling you the model is being trained, refined, and prioritized around the use cases that population cares about. Long-form reasoning, writing, coding, analysis, structured business workflows. The feedback loops between the users and the model are dense in exactly the work your operations team is going to do.
The opposite is also true. A tool with a consumer-mirror user base is probably being tuned for mass consumer chat. Which is fine, if that's your use case. It's just not the use case most mid-market businesses actually have.
So when I sit across from a CEO and he asks why we use Claude, I point to the benchmarks, the architecture, the deployment track record. Increasingly I also point to the chart.
If you're still running a Copilot license IT set up a year ago because the seats were free, the chart is a cheap second opinion. You don't have to trust me that Claude is the right tool for operators. You can trust the operators themselves.