We read the app store reviews of ChatGPT, Gemini & Claude so you don't have to. Each one is failing differently.

There are two kinds of AI benchmarks. The kind with acronyms — MMLU, SWE-bench, Arena Elo — where models battle over decimal points. And the kind nobody publishes: the app store review section, where a million people who have never heard of a benchmark tell you exactly where it hurts. We build a tool that exports app reviews, so we did the obvious thing: we pulled the 10,000 most recent reviews of ChatGPT, Gemini and Claude from both app stores — 60,000 reviews — and read the room.

The headline numbers look like a boring race. The complaints underneath look like three completely different products having three completely different crises.

One in four Claude reviews in our sample is a one-star. Case closed, ChatGPT wins? Hold that thought — because when you sort 12,145 negative reviews by what they're actually about, the leaderboard stops meaning what you think it means.

01The star ratings: a race with three different finish lines

First, the shape of the vote. Star ratings on app stores are famously U-shaped — people show up to worship or to riot, rarely to shrug. But the proportions of worship to riot differ wildly:

So Claude is the worst product, right? That's what the stars say. It's not what the words say. Which brings us to the fun part.

02The complaint fingerprint: three apps, three diagnoses

We took every English-language 1–3 star review — 12,145 of them — and tagged the complaints. Not sentiment mush. Specific, named failure: wrong answers, usage limits, login walls, billing, memory loss, crashes. Then we asked one question: what share of each app's angry reviews mention each problem?

The result is the closest thing we've seen to a fingerprint. No two apps fail alike:

messages keep disappearing for no reason within the same chat, and gpt has no memory of them, losing days of work. — actual 1★ ChatGPT review, July 2026

03Only one company answers its reviews. It's not who you'd guess.

Here's the stat we double-checked three times because we didn't believe it. Of the three richest AI labs on earth, exactly one replies to app store reviews:

Read that again: in 20,000 ChatGPT reviews — including thousands of people reporting lost work, broken subscriptions, and refund requests — OpenAI responded to none of them. Google, the company with a reputation for killing products before answering support tickets, is out there replying to one in seven iOS reviewers. The AI race apparently doesn't include the reply button.

04What the love letters reveal

Flip to the 5-star side and each app has a signature use case — and they are not the same product at all:

ChatGPT is the default — work, homework, and, strikingly, companionship: 13.2% of its five-star reviews are some version of "it listens to me." Claude is the specialist — its fans say "coding" ten times more often than Gemini's, and it leads in "better than ChatGPT / I switched" mentions. In fact, 12.1% of Claude's negative reviews name a rival, triple ChatGPT's rate. Nobody compares ChatGPT to anything; it's the thing others get compared to. And Gemini? Gemini's five-star reviews are as diffuse as its complaints. It ships on every Android phone on earth — our sample clocked its Android review velocity at roughly 3,300 reviews per day, versus ~790 for ChatGPT and ~190 for Claude — and still hasn't found the thing its users would riot over losing.

it's a slow assistant. it's alright, but not as quick as doing it on your own — 2★ Gemini review. The shrug, verbatim.

05The scoreboard nobody's watching

Put it together and the star ratings invert into something more interesting:

ChatGPT wins the ratings but is quietly accumulating a trust debt. One in five of its angry reviews says "it was wrong." For a product people use for homework, work, and increasingly as a confidant, that's the complaint that compounds. Claude loses the ratings for reasons that have nothing to do with intelligence. Rate-limit anger is demand wearing a disguise — you don't riot over the limits of a product you don't want more of. If Anthropic fixed capacity and login, our data says its rating problem mostly evaporates. Gemini has distribution instead of devotion. 3,300 reviews a day and not one thing its users can't live without — the exact profile of a product that wins shelf space and loses hearts.

06Do this for your own app (it takes 60 seconds)

Here's the part where we tell you why we did this. Every number above came out of a CSV — the same CSV you can pull for any public app: yours, or the competitor whose users you'd like to steal. Export the reviews, filter to 1–3 stars, count the repeated complaints. That list is a roadmap someone else's customers wrote for you, for free.

We'll re-run this in six months and see whether OpenAI patched the trust hole, whether Anthropic bought enough GPUs to stop the limit riots, and whether Gemini found something to be loved for. Follow along — or run it yourself before we do.