You can rebuild Airbnb's interface in a weekend with AI tooling, and it will be worth nothing, because the product was never the asset. The asset is two-sided demand: millions of hosts with listings and millions of travellers with intent, plus the trust, reviews, payments rails and regulatory position that took fifteen years to accumulate. Software is now the cheapest part of any marketplace.
This matters far beyond marketplaces. If AI has made building cheap, then defensibility has moved somewhere else — and knowing where is the difference between a startup and an expensive hobby.
What actually got copied, and what did not
A clone reproduces the visible layer: search, listings, calendar, checkout, reviews UI. What it cannot reproduce is the inventory, the demand, the trust signals accumulated over a decade, the payment and insurance infrastructure, the local regulatory relationships, and the operational muscle for handling a bad stay at 1am in a city you have never visited.
Every one of those is expensive, slow and human. None of them get cheaper because a model can write code.
Where defensibility lives now
- Distribution: an audience, a partnership, or a channel that competitors cannot buy their way into cheaply.
- Proprietary data: information generated by usage that improves the product and does not exist outside it.
- Workflow depth: being embedded in how work is done, with integrations and history that make switching painful.
- Trust and brand: the reason a buyer picks you when the feature list is identical.
- Network effects: value that grows with each additional participant.
- Regulatory and operational position: licences, certifications, insurance, and processes that took real time to establish.
- Speed of learning: not building faster, but running more validated cycles with real customers than anyone else in your niche.
What this changes for founders
It inverts the sequence. When building was expensive, the plan was to raise money, build the product, then find demand. When building is nearly free, you find the demand first and treat the build as a consequence.
Practically: get the waiting list, the pilot commitment, the partnership or the community before you build. Not because building is hard, but because it is now so easy that having built something says nothing about your position.
The clone test
Ask yourself: if a competent team cloned my product tomorrow with AI tooling, what would I still have that they would not?
If the answer is "nothing yet, but we would be ahead on features", you have no moat — feature leads last weeks now. If the answer names customers, data, a channel, an integration nobody else has, or a relationship that took a year to build, you have something. Run this test before your next planning cycle; it reorders most roadmaps immediately.
How to build an asset instead of a copy
Pick a niche narrow enough to dominate. Get into the workflow rather than beside it. Generate data that only exists because your product runs. Build one distribution channel deliberately — a community, a content position, a partner who already owns the audience. And accumulate proof: named customers with numbers attached.
None of these are software problems, which is precisely why they remain defensible in a world where software is generated on demand.
What this looked like in practice
Wocal was not defensible because of its features — the booking flow was replicable in weeks. It was defensible because 300+ venues had their inventory, their staff habits and their customer history inside it. That is what supported a £2.7M pre-money valuation: switching cost and supply, not screens.
If you are working out where your own advantage sits, the Crossley Method starts from exactly that question, and why most AI products never generate revenue covers the adjacent failure.
Frequently asked questions
- Can you clone a successful app with AI?
- You can clone the interface and much of the functionality in days. You cannot clone the customers, supply, trust, data, integrations or regulatory position that make the original valuable.
- What is a moat when software is cheap to build?
- Distribution, proprietary usage data, workflow depth and switching cost, trust and brand, network effects, regulatory position, and speed of validated learning. Features are no longer a moat because they can be reproduced in weeks.
- Should I still build an MVP if competitors can copy it?
- Yes, but treat it as a way to acquire customers, data and learning rather than as the asset itself. Secure demand first — a waiting list, a pilot or a partnership — then build to serve it.
- How do I test whether my product is defensible?
- Ask what you would still have if a competent team cloned your product tomorrow. If the answer names customers, data, a channel or an integration, you have a moat. If it names features, you do not.