AI冲击下App的护城河：留存与差异化策略

I’m not going to tell you that AI is not coming for you. I know founders wake up at 2am thinking: “Why would anyone pay for my app when AI does it for free?”

But I also don’t believe apps will disappear altogether. Ethan Garr, a growth advisor, shared on a Sub Club Live about building with AI that a friend had said apps would be gone within six months. Not what you want to hear when you work in the app space. Yet here we are more than six months later: apps are still around, and more are being launched than ever before (7x more since 2022, to be specific). So I don’t believe they’ll disappear. Yes, LLM chats are replacing certain apps. Yes, the space is changing faster than ever. But the data suggests there is still hope, and I’m willing to cling to that data.

Especially as I’ve found you can win the game if you focus on the areas where an LLM chat cannot beat your app.

"...since March we’ve seen a significant spike in student interest in ChatGPT. We now believe it’s having an impact on our new customer growth rate."

When CEO Dan Rosensweig of Chegg, a homework-help subscription service, said this during an earnings call in 2023, the impact hit hard:

• The company’s stock fell 50%
• By 2025, 45% of their workforce were made redundant
• By Q2 2026, the freefall continued and they attempted to pivot away from homework toward an employability platform

Every step of the way, the blame was the same: AI.

Now take Duolingo: ChatGPT, and most LLMs for that matter, can also tutor you, for free. Google has layered live translation into its products — only a few months ago, I did a full strategy session with a subscription app startup in Portuguese (even though I don’t speak a word of Portuguese). The experience was honestly mind-blowing. I’m curious what I’m like in Portuguese.

Yet Duolingo isn’t floundering:

• Q2 2026 revenue was up 18%
• DAU rose 23% to 58.7 million
• Paid subscribers increased 17% to 12.7 million

By contrast, Duolingo’s CEO Luis von Ahn seems largely unimpressed by AI competitors:

“Just having conversations in French on something like ChatGPT gets pretty boring after a while. It doesn’t keep you there. We keep you on task with all the gamification.”

So what explains the difference between the two? Why is one being, frankly, obliterated by AI, while the other’s owl is gleefully celebrating its success?

It’s about the app itself: while both companies create educational content, Chegg’s product is essentially a one-shot answer: a lookup experience without enough built around it in terms of data, habit, and workflow. Duolingo, meanwhile, has spent 14 years building gamification systems and habit-forming product design that are much harder to replicate.

People will pay for AI apps, they just won’t stay

One thing worth separating before we dive into the numbers is what we mean by an ‘AI app’, because we often talk about these all under AI:

• An LLM chat is ChatGPT, Claude, Gemini
• An AI app is a subscription app with AI as a central feature inside it, like Cal AI or Rosebud (and quite possibly yours)

The first is the free competitor keeping you up at night. The second might be your own app. The data below is about the second.

AI-powered apps are casually out-monetizing and outperforming non-AI apps, according to RevenueCat’s State of Subscription Apps 2026:

• 41% more revenue per payer
• 52% better trial conversion

But there is a flip side to this:

• 12 month retention of 21.1% vs. 30.7%
• Refunds are about 20% higher

So consumers are willing to pay for AI apps, but the harder challenge is getting them to keep paying. When OpenAI launched the AI video app Sora, it got more than 12 million downloads, but by Day 30 retention was a miserable 8% or so compared with the 30%+ industry standard. Since then they’ve sunsetted Sora, and I was met with this depressing screen:

So the question in this article’s title is secretly the wrong one: people will pay. They’re already paying more — and faster — for AI-powered apps than for almost anything else right now.

The real question is why will they still be paying in month six? That’s a product question, not a marketing one, and it’s the question the rest of this article sets out to answer. (It’s also a much nicer question to wake up to at 2 am.)

AI alone isn't enough to win customers permanently, and Sora proves it: when the company that builds the model can't keep people, the model was never what kept them. Which is also why I don't think the other AI apps or copycat AI apps in your category are your biggest competitors.

ChatGPT is the biggest ‘window’ ever built

A while ago, I was standing in the kitchen, checking several weather apps on my phone, trying to decide whether to walk the dog now. I turned around to find my father-in-law looking at me like I’d lost it: "Why are you doing that? Just look out the window."

Weather apps’ competitors aren’t other apps; they are looking out the window, asking around, or even just going to stand outside.

In the world of apps, your window is ChatGPT, Claude, or whatever other LLM chat of your choice. The free models, for now, can do a lot, and we see growing usage in categories of apps that support them, e.g, one in five US chatbot users ask for medical advice, and as many again about diet and fitness.

The second, more meta way it is your window competitor is the ability to help you and others build faster. App launches went from about 2,000 to 14,700+ per month in four years; the volume is huge.

But in that, we still see 69% of subscription revenue going to pre-2020 apps, versus just 3% to apps launched in 2025 and 2026. Eric Seufert argues that while the cost of building is lower, the cost of distribution has increased because apps are all chasing the same attention. Rik Haandrikman echoes this: distribution is a moat.

Those who are holding on to that 69% know what a weather app can do that a window can’t, what their app can do that an LLM can’t. Those who can’t explain the differentiator of their app are the ones slowly dying under the weight of AI's growth.

Anatomy of the apps AI has actually killed

This is not a random murder spree; AI’s victims share something in common (I clearly need to stop reading Swedish noir novels). The apps replaced by LLM chats have one or more of these four things in common:

• No structure
• No memory
• No habit
• No precision advantage

Most apps being replaced are a single question or task and an answer. That was Chegg’s real problem: strip away the branding, and the product was a prompt with a subscription attached. Stack Overflow, a question-and-answer platform for developers, is on the same path, with traffic falling around 6% every month since early 2022. Why? An LLM gives you the same answer instantly, in a conversation, without posting a question and hoping a stranger replies.

Dan Layfield, Founder of Subscription Index, made a point on Sub Club that explains the pattern: your retention is dictated by how long the user has the problem you solve. Phone plans are retained for decades because the problem never goes away. An answer-lookup problem dies the second the answer arrives — and now the answer arrives in three seconds, for free. Chegg's real problem was never ChatGPT; it was that the problem they solved only lasted one homework question at a time.

The uncomfortable truth: if your app’s value fits in one prompt and one reply, you are in the blast radius.

The Blank Box Test (and six things a blank box cannot do)

I have a challenge for you called the Blank Box Test. It’ll only take two minutes:

• Open your LLM chat of choice
• Write out your user’s problem in their words
• Analyze the results vs. what your app offers