ChatGPT Lost 22 Points of Market Share in a Year. The AI Monopoly Is Quietly Over.

For three years, "AI chatbot" and "ChatGPT" meant the same thing. OpenAI didn't just have a lead — it had a monopoly on the default. When your product name becomes the verb, competitors don't beat you on features. They wait for you to get bored and start showing ads.

About that.

According to Similarweb data plotted earlier this month, ChatGPT's share of worldwide web visits to AI assistants has fallen from 76% to 54% in roughly a year. Gemini went from 6% to 28%. Claude went from 1% to 9%. That's not churn. That's a leak in a hull everyone assumed was unsinkable.

And before you ask — no, this isn't one shaky dataset. Sensor Tower's app analytics tell the same story from a completely different angle: ChatGPT's share of AI app usage slipped below 50% for the first time in January and hit 46.4% by end of May, with Gemini at 27.7% and Claude at 10.3%. Two methodologies, two measurement messes, one direction. When every way of counting agrees, the trend is real.

The numbers, honestly

Let me be precise about what these figures actually measure, because sloppy market-share math is how AI discourse rots.

The Similarweb numbers are web visits — desktop and mobile browser traffic to chatbot sites. The Sensor Tower numbers are app usage and downloads. They disagree on the exact percentages (they always will) and agree violently on the trajectory:

• ChatGPT: 76% → 54% of web visits (Similarweb); ~50% → 46.4% of app share (Sensor Tower)
• Gemini: 6% → 28% of web visits; 27.7% of app share
• Claude: 1% → 9% of web visits; 10.3% of app share
• Grok, Perplexity, DeepSeek, Meta AI: under 5% each, scrapping over the leftovers

ChatGPT is still enormous — 1.1 billion monthly users against Gemini's 662 million and Claude's 245 million. OpenAI reported 900 million weekly active users in February. Nobody is writing OpenAI's obituary.

But "still huge" and "losing a fifth of your share in a year" are both true, and the second one is the story.

Gemini didn't win anything. Google did.

Here's the uncomfortable part of this trend for anyone who believes merit drives markets.

Gemini did not go from 6% to 28% because it became 4.5x better. It went from 6% to 28% because it's Google, and Google owns the distribution rails that two billion people touch before breakfast. Gemini is preinstalled on Android devices. It's woven into Chrome, Workspace, Search. When Google decides you're going to try its assistant, you're going to try its assistant.

The Similarweb chart itself notes the quirk: Gemini has a billion app MAUs coexisting with ~28% web share precisely because so much of its usage never touches a browser at all. It arrives through the operating system, like weather.

This is the oldest pattern in tech, and it never stops working. Internet Explorer didn't beat Netscape on quality. It came with the machine. Google is running the same play with a 2026 budget, and OpenAI — a company whose entire moat was "we got there first" — is discovering that first-mover advantage has a half-life measured in quarters, not years.

The lesson isn't "Gemini is better." The lesson is that in consumer software, distribution is destiny, and the model underneath matters less than everyone building models wants to believe.

OpenAI is speedrunning the enshittification timeline

Now the self-inflicted part, because ChatGPT's decline isn't only Google's doing.

In February, OpenAI signed a deal with the U.S. Department of Defense. Sensor Tower measured a spike in uninstalls immediately after. Whatever you think about defense contracts politically, the data says a meaningful chunk of users thought about it, decided they didn't like the vibe, and left. Brand trust turned out to be load-bearing.

Also in February, OpenAI started experimenting with ads in ChatGPT. By May, an average of 17% of daily users were being served ads — with software and shopping as the biggest categories. ChatGPT is now sending referral traffic to Target, Walmart, and Costco. Amazon, sensing a competitor's shopping engine crawling its moat, blocked ChatGPT's crawlers outright.

Sit with that sequence. The most trusted product of the decade, the one people paid $20/month for precisely because it wasn't adtech, started serving sponsored answers within four years of launch. Facebook took over a decade to fully make that trade. OpenAI did it before the product's fifth birthday.

Users notice. The uninstalls are the noticing, quantified.

The quiet winner is Claude, and it's not even close

Here's the number nobody's putting in headlines: 13% of Anthropic's users pay for a subscription. That's the best paid-conversion rate in the industry — by a lot. Sensor Tower flags it, notes Claude is closing in on ChatGPT's user-retention rate, and moves on, which is insane, because that's the most important stat on the page.

Downloads are vanity. Share of attention is rented. Revenue per user is the only thing that survives winter.

Think about what Claude's numbers actually say: fewer people use it, but the people who use it reach for their wallets. That's the profile of a tool that has become part of someone's workflow — not a toy you poke when a link goes viral. Claude grew from 1% to 9% of web share doing essentially zero consumer marketing, on the strength of "it's genuinely good at work stuff."

Meanwhile the industry overall is on pace for 2.3 billion AI app downloads and $4.2 billion in in-app spending in the first half of 2026 — more than double the $1.83 billion from a year prior. The pie is exploding, and the slices are fragmenting. Asia actually recorded its first download decline (down 3.3% in Q1). Growth is decelerating even as absolute numbers climb. The land-grab phase is ending.

The endgame isn't winner-take-all. It's three players and weather.

One more stat from the Sensor Tower report worth your attention: the top three assistants now command 89% of all time spent in AI apps, and total hours are on pace to more than double — from 17.2 billion hours in the first half of 2025 to roughly 36 billion in the same period this year.

Read those together. Usage is exploding, but it's pooling around three apps. Everyone else — Grok, Perplexity, DeepSeek, Meta AI, plus the long tail of companion apps and wrappers — is fighting over 11% of a growing pie. That's a brutal place to build a consumer business, and it explains why so many AI startups are pivoting to enterprise, to agents, to infrastructure: the consumer chat layer is ossifying faster than anyone predicted.

But notice what "three players" actually means here. It's not three winners. It's two companies with distribution (OpenAI's brand, Google's everything) and one company with the best conversion rate in the category (Anthropic). Each has a different kind of moat, none of them is the one VCs pitch — "the smartest model" — and any of the three could look silly in eighteen months if the ground shifts again. It shifted 22 points in twelve.

The floor, meanwhile, keeps rising underneath all of them. Every open-weights release narrows the gap between "free-ish and yours" and "$20/month and someone else's." When the gap gets thin enough, the middle of the market doesn't pick a winner. It stops caring. That's the endgame nobody at the top wants to say out loud.

What this actually means for you

Zoom out and the real story isn't about three companies. It's about you, and it's about a word that's doing a lot of quiet damage: default.

The collapse of ChatGPT's share proves something everyone in AI suspects but hates to say out loud: switching costs between AI assistants are effectively zero. No lock-in, no network effect, no re-learning curve worth mentioning. You type the same question into a different box. Users are switching on the strength of a news cycle, an ad, a preinstall. That cuts both ways — it's how Gemini gained 22 points, and it's how Gemini can lose them too.

Loyalty in consumer AI is fictional. The moment a better or cheaper or less creepy option appears one tap away, people take it. Which means any strategy built on "our model" being the moat is a strategy built on sand — and that's good news if you're not one of the three giants, because the same zero-loyalty dynamic applies to open models.

Here's where I'll stop pretending to be neutral: this is exactly why the open-weight crowd has the wind at its back. When nobody's loyal to a brand, they'll be loyal to whoever gives them the best deal on capability — and open models running on hardware you own are the terminal form of that bargain. The same Qwen or GLM or DeepSeek weights that cost providers pennies to serve can run on your machine, answerable to no roadmap, no ads department, no defense contract. The fragmentation at the top of the market is the pressure that pushes everything else open.

If your AI workflow is one company's app, you've built on someone else's distribution — and OpenAI just spent a year demonstrating what that's worth. Roughly 22 points.

---

If you're rethinking the one-app-for-everything approach: CopperRiver is a desktop AI assistant for Mac that runs on open-source models — GLM, DeepSeek, Qwen, Kimi, MiniMax — and switches between them on the task, not the marketing. Your files, your terminal, your browser, your machine. Plans start at $9/mo, which is less than one assistant subscription and none of the sponsored answers.