Vibe Coding Flooded the App Stores. PMF Still Kills 90% of Them

Andrej Karpathy coined "vibe coding" in February 2025 — the practice of describing what you want to an AI and shipping whatever it generates without fully understanding the code. Twelve months later, the App Store has more apps than ever. Indie hacker communities are full of people who shipped their first SaaS in a weekend. And the failure rate at 90 days has not changed at all.

The build problem is effectively solved. The distribution and retention problem never was.

What Vibe Coding Actually Changed

Before Cursor, Claude Code, and GitHub Copilot reached their current capability level, building a production SaaS required either strong engineering skills or the budget to hire engineers. That barrier filtered out a lot of ideas — some bad, some potentially good — before they reached users.

AI coding tools removed that filter. A non-technical founder can now ship a working product with Supabase, Vercel, and a Stripe integration in 48–72 hours. The code quality varies, but the product ships and functions. That is genuinely new. Three years ago it was not possible at this speed and cost for most people.

What did not change:

• Whether users will pay for your product
• Whether users will return after the first session
• Whether there is a scalable channel to acquire users at unit economics that work
• Whether the problem you are solving is painful enough to displace an existing solution

These are not engineering problems. AI cannot solve them for you because they require talking to potential users, understanding behavior, and iterating on positioning — not code.

The Quality Problem Is Real But Not the Main Problem

Technical debt from vibe-coded products is a genuine issue. When you do not understand the code you are shipping, you cannot maintain it safely. Security vulnerabilities get introduced because the developer does not recognize them. Database queries get written that work fine at 100 users and destroy a server at 10,000.

The Claude incident from early 2026 where an AI autonomously wiped a German founder's production database — documented and confirmed — illustrated the extreme end of this: shipping AI-generated infrastructure changes without understanding what they do.

But the quality problem is largely a problem for products that get traction. Most vibe-coded SaaS products fail before they encounter scale issues. They fail because nobody uses them.

The "bad code" critique misses the actual failure mode. The typical trajectory:

1. Build in a weekend (cost: near zero)
2. Post to Product Hunt, Hacker News, relevant subreddit
3. Get 200 signups from launch day traffic
4. Active users at day 30: 12
5. Active users at day 90: 3
6. Product abandoned

The code quality was irrelevant to that outcome. The problem was that the founder built something without validating that the pain was real, urgent, and worth paying to solve.

What Product-Market Fit Actually Requires

Y Combinator's acceptance rate is approximately 1.5–2% of applicants. This has not changed significantly in the AI era despite the lower barrier to shipping. YC partners have noted that the flood of AI-built demos has made it easier to identify founders who cannot articulate user problems clearly — because they have a polished demo but no user insight.

PMF is not a product state. It is evidence from user behavior: strong retention curves, high NPS, organic referral, users being genuinely upset when you suggest taking the product away. These signals cannot be faked, engineered, or generated by AI tools. They come from users who have a real problem and found that your product solves it better than alternatives.

The specific things AI tools cannot do for you:

User research: AI can help you analyze transcripts once you have them. It cannot replace the actual 30-minute customer discovery calls where you hear someone explain their workflow in their own words and realize your initial assumption was wrong.

Pricing discovery: What users say they will pay and what they actually pay are different. Finding the price point where conversion and retention hold requires real experiments with real users.

Positioning and messaging: The words that make your target user say "finally, someone built this" versus "that's basically X but worse" come from deep category knowledge and user interviews, not from AI-generated landing page copy.

Distribution channel validation: Whether SEO, cold outreach, paid acquisition, or community distribution works for your specific product is an empirical question. AI tools can execute on a channel once validated; they cannot tell you which channel will work.

The SaaS Market Is Not Saturated — Your Category Might Be

A common complaint from failed vibe-coded products is that "the market is saturated." This is usually wrong. The SaaS market is not saturated. Specific poorly-differentiated categories are saturated.

"AI writing assistant" has hundreds of entrants. "Automated SOC 2 evidence collection for infrastructure teams running on Kubernetes" has a handful of players with clear moats. The narrower and more specific the problem, the less saturation, the higher willingness to pay, and the more tractable user acquisition becomes.

Vibe coding combined with a specific, validated niche is a genuine competitive advantage — you can ship a version 1 in a weekend and iterate at a speed that a well-funded team with slower development cycles cannot match. The problem is that most vibe-coded products are built for broad categories where the founder picked the niche based on what seemed easy to build, not what users actually needed.

Retention Is the Real Scorecard

The metric that distinguishes products with PMF from products without it is retention — specifically, the shape of the day-1, day-7, and day-30 active user curves.

A flat retention curve after day 14 — where roughly the same percentage of original signups are active — means you have PMF. The absolute number does not matter. If 15% of week-one users are still active at month three, and that curve is flat (not still declining), you have something.

A retention curve that keeps declining to near zero by day 30 means the product does not have PMF regardless of launch day signup numbers. AI tools built this product perfectly to spec. The spec was wrong.

What kills retention in AI-built products specifically:

Unexpected behavior at edge cases: AI-generated code handles the happy path. When users do unexpected things, the product often breaks or behaves confusingly. The founder, not understanding the code, cannot debug quickly.

Missing the workflow: AI tools are good at building features in isolation. They are less good at building the workflow that real users follow, because workflow understanding requires user research that the AI tool did not have access to.

Performance at real usage: A product that works fine in development with synthetic data often has latency or reliability issues with real user data at real usage patterns.

The Opportunity That Is Actually There

The vibe coding era creates one genuine advantage that most founders are not exploiting: the ability to run experiments at near-zero cost.

Instead of building one product and launching it, the correct strategy is to build five small experiments in five weekends, each targeting a specific validated user pain, and measure which one gets genuine retention signals. Kill four, double down on one.

This is how AI coding tools should change your strategy: not "ship faster once," but "experiment faster continuously." The founders who understand this are using AI tools not to build their main product but to build the validation experiments that tell them what the main product should be.

Key Takeaways

• Andrej Karpathy coined vibe coding in February 2025 — 12 months later, the App Store has more apps than ever, but day-90 failure rates have not changed
• AI tools solved the build problem — they did not solve user acquisition, retention, or product-market fit, which are behavioral and positioning problems
• YC acceptance rate is still ~1.5–2% despite the lower barrier to shipping — partners can now identify founders without user insight more easily because polished demos without user evidence are common
• Retention curves, not signup numbers, indicate PMF — a flat day-30 curve at 15% beats a spiking launch day at 5,000 signups with 98% churn
• Saturation is category-specific: "AI writing assistant" is saturated, "automated compliance evidence for Kubernetes infra teams" is not
• The real opportunity: use AI tools to run five cheap validation experiments instead of building one product without evidence

Build with AI tools, track AI model costs with LLM API Pricing, and evaluate your own career exposure to automation with Will AI Replace Me.