Anyone Can Build Your Product. Not Everyone gets Chosen.

Written by: Raziel Bhandari (UQBA Sub-executive)

August 27, 2026

In July 2025, a company called Pickle cloned Cluely in four days and gave it away for free, source code included. Cluely was charging up to $200 a month for the same product.

It didn't dent Cluely at all.

Cluely was a rocket ship at that point, $5.3 million raised in April, then a $15 million Series A led by Andreessen Horowitz two weeks before the clone appeared. The product watched your screen during a video call and fed you AI-generated answers in an overlay nobody else could see. Pickle's version, Glass, did exactly that, for nothing, forever.

Glass went nowhere.

The supply of software exploded. Demand didn't move an inch. What's left to compete on was never in the code.

What Cluely had that Glass didn't was 70,000 signups in its launch week, a founder suspended from Columbia for building a tool to cheat technical interviews who turned the suspension into a personality, and a launch video of him using the product on a date to invent facts about himself.

The obvious lesson is that marketing wins now. The real lesson is stranger, and it only shows up if you follow Cluely past the funding announcement.

Everyone Can Build

The supply shock isn't a vibe. It's in the numbers.

In February 2025, Andrej Karpathy coined "vibe coding", where you "fully give in to the vibes" and forget the code exists. A month later, Y Combinator said roughly a quarter of its current batch had codebases that were 95% AI-generated. By mid-2026, Lovable, which builds working apps from a written prompt, was doing $500 million a year with a million new projects a week, with four out of five of its users unable to code.

Apple's App Store took 557,000 new app submissions in 2025, up 24% in a year, the biggest wave of new releases since 2016, reversing almost a decade of decline.

That's the supply side. The demand side is where it gets strange.

In May 2026, economists at MIT and Wharton matched more than 100,000 developers, showing which AI tools each one used. Developers using autonomous coding agents wrote 180% more code. That led to 50% more projects started, which became 30% more software actually shipped. The tools clearly work.

Then the economists went looking for all that software in four app marketplaces to see whether anyone was using it.

The conclusion: More was made. None of it was consumed.

The Bottleneck Changed

180% more code becomes 50% more projects, becomes 30% more shipped, becomes no measurable increase in anything used. Economists call this a weak-link problem: a pipeline speeds up only as much as its slowest human step.

METR ran an experiment where sixteen experienced developers worked through 246 real tasks inside codebases they knew intimately. With AI tools, they were 19% slower, while estimating afterwards that AI had made them 20% faster.

They weren't slow at writing; they were slow at checking. A mature codebase carries knowledge that lives nowhere a model can read: why a function is shaped oddly, which shortcuts are safe, what quietly breaks three files away. The model proposes something plausible, and a human has to reconcile it against everything the model doesn't know.

Generation got cheap. Verification didn't.

Nobody has More Time

Total apps grew 24% in 2025. The total hours people spent inside apps that year grew 3.8%. Per person, per day, the increase was 1.1%.

Herbert Simon saw this coming in 1971, when he wrote that a wealth of information creates a poverty of attention. Supply compounds. Attention can't, because it's capped by how many people are alive and how many hours they're awake. Double the world's software in a year, and you have not moved the world's willingness to use it.

Lovable's own numbers show it perfectly. 50 million projects have been built on the platform, and together they draw about 720 million visits a month; 14 visits each per month, and because a handful of winners take most of the traffic, the median project is closer to 0.

Mobile apps, which long predate AI and vibe coding, tell the same story. The median app keeps about only 4% of its users thirty days after install with close to 0 in-app purchase revenue, whilst the top 1% of publishers take roughly 92% of all in-app purchase revenue.

The only change is how many people are now at the bottom.

The Customer Acquisition

It's easy to say brand matters more now. It's more useful to know why it has to.

George Akerlof won a Nobel Prize for the used-car problem. The seller knows whether the car is any good; the buyer doesn't. So the buyer only offers an average price, and at an average price, the honest seller with the good car walks away. Quality that can't be verified stops earning a premium, so it stops being worth supplying. What pays instead is a signal the buyer can trust without checking.

That is the same bottleneck as before but just moved to the other side of the transaction. Developers can't cheaply verify what the machine wrote. Buyers can't cheaply verify what the developers shipped. Generation is cheap at both ends and checking is expensive at both ends, which is why the pipe stays narrow no matter how much you pour in.

The natural objection is that apps are usually free, so what’s the cost for buyers? The currency here isn't money; it's attention. Downloading a hundred apps to find the good one costs nothing yet is impossible. Unverifiable quality on a crowded shelf is a used-car market, whatever the price tag says.

Thus, nobody evaluates. Everybody uses a shortcut: a name they've heard of, a person they follow, a friend's recommendation. Akerlof named the fixes himself: brand names, licensing, guarantees, and every one of them is a way of letting a stranger trust you without checking.

For thirty years, "We can make good software" was a moat, because engineers were scarce and expensive. Every founder who couldn't code needed someone who could. AI removed that barrier and, with it, changed the moat.

The Exception?

Cursor is the strongest argument against all of this. The AI coding tool passed $100 million in annual revenue about a year after launch, faster than any software company in history, with no stunts and essentially no marketing. Just a good product. In August 2026, SpaceX closed its acquisition of Cursor's maker for $60 billion in stock, the largest purchase of a venture-backed startup on record, putting it in the elite list of decacorns.

So can a product still win on merit alone? Yes, but on one condition.

Cursor sold to developers, and developers are one of the few audiences who can judge the quality of a product with minimal costs. That’s the condition: cheap verification.

Where buyers can cheaply verify, quality still pays. Where they can't - which is most consumer software - they fall back on who they trust instead.

Attention is NOT Retention

Roy Lee, Cluely's founder, wasn't a better engineer than the people at Pickle. He was a better broadcaster. The suspension, the "Cheat on Everything" tagline, the deliberately obnoxious video, that was the marketing, and it produced 70,000 signups in a week.

Being good at building used to exempt you from needing a personality, because building could be the moat itself. That exemption is expiring.

By November 2025, Cluely had quietly repositioned as an ordinary AI meeting assistant, as its sign-up and revenue numbers were gradually declining after its initial boom.

Personality is an acquisition channel. It is not a retention mechanism.

Software is used to hold users through the product itself, with switching costs manufactured out of engineering difficulty. However, when someone can rebuild your features in four days, those costs evaporate.

So, what actually holds anyone? Let’s look at Wordle.

In November 2021, it had 90 weekly users, and two months later, it had more than two million weekly users. It’s one web page, and the cost of leaving was a single click.

Naturally, it got cloned instantly, and the best-known clone was better on paper: the same game, plus longer words and unlimited plays. Yet nobody moved.

What held those players wasn't in the software, but that everybody got the same word on the same day. This made playing it a shared event that everyone did.

This is the difference between what Lee built and what Wordle built. Lee's audience had a relationship with Lee. It ran through him, it needed constant feeding, and it transferred almost nothing to the product, which is why the revenue never caught up to the attention. Wordle's players had a relationship with each other, and the product was only the place it happened.

So it isn't that marketing beat engineering. It's that getting people and keeping people now rests on the same asset: the relationship between you and the people who use your product.

What Survives Being Copied

The natural response is that AI writes marketing too. It does, which is exactly why it confers no advantage. Since January 2025, Google's own guidelines have told the contractors who grade its search results to give the lowest rating to mass-produced AI content.

So the advantages sort cleanly enough. Virality, a launch moment, a clever stunt, being first: all of these can disappear overnight. An installed base, a licence, data nobody else can gather, a community that would notice if you vanished, a reputation: these last.

Buffett, whose castle-and-moat metaphor started all this, said: “most moats aren't worth a damn”. Then he named the exception: a powerful worldwide brand, the barrier that makes high returns last. Not the ability to make the product.

Anyone can build your product. Someone probably will, and they might give it away free. What they can't copy is the reason somebody picked you in the first place, and the reason they didn't leave.

Making things is nearly free now. Being chosen costs what it always did.

Sources

1. Mert Demirer, Leon Musolff & Liyuan Yang, "Writing Code vs. Shipping Code," NBER Working Paper 35275, May 2026 — 100,000+ developers; the 180% / 50% / 30% / no-increase chain and the weak-link finding.

2. METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity," July 2025 — 16 developers, 246 tasks, 19% slower against a 20% perceived speed-up.

3. METR, "We Are Changing Our Developer Productivity Experiment Design," February 2026.

4. George Akerlof, "The Market for 'Lemons': Quality Uncertainty and the Market Mechanism," Quarterly Journal of Economics 84(3), 1970 — counteracting institutions at pp. 499–500.

5. Herbert Simon, "Designing Organizations for an Information-Rich World," 1971.

6. Appfigures, "Biggest Release Year in Nearly a Decade," December 2025 — 557,000 App Store submissions, up 24%.

7. Sensor Tower, State of Mobile 2026, January 2026 — app hours growth; top 1% of publishers at ~92% of in-app purchase revenue.

8. Pushwoosh mobile retention benchmarks — ~4% median day-30 retention. Industry data, not peer-reviewed; corroborated across sources but softer than the citations above.

9. Google, Search Quality Rater Guidelines, 23 January 2025 — lowest rating for mass-produced AI content.

10. TechCrunch, "A quarter of startups in YC's current cohort have codebases that are almost entirely AI-generated," March 2025.

11. TechCrunch, "Lovable says it has hit $500M in annualised revenue with 1 million new projects a week," June 2026.

12. TechCrunch, "Cluely's ARR doubled in a week to $7M, founder Roy Lee says," July 2025 — the figure later corrected to $5.2 million.

13. CNBC, "SpaceX to acquire the AI coding startup Cursor for $60 billion," June 2026; deal closed 14 August 2026.

14. Macworld, "Wordle clones are removed from the App Store," 2022.

15. Berkshire Hathaway 1995 annual meeting; 2007 Chairman's Letter.