Is Your AI Startup Actually Defensible, or Just Another ChatGPT Wrapper? A 7-Question Test

On Hacker News and in most serious investor conversations, "it's just a ChatGPT wrapper" is the fastest way to end a pitch. The dismissal is both lazy and partly correct. Lazy, because some of the most successful consumer software companies are essentially interface layers on top of commodity infrastructure — nobody calls Stripe a "bank API wrapper." Partly correct, because a large proportion of AI startups launched in the last two years genuinely are thin interface layers on top of foundation models, with no proprietary data, no durable differentiation, and no reason for a user to stay when a well-resourced competitor ships the same feature.

The question worth asking is not "am I a wrapper?" It's "am I defensible?" Defensibility and technical novelty are not the same thing. You can build a durable business on top of someone else's model if you have the right combination of other factors. You can also build something technically impressive that has no moat at all. The seven questions below are designed to help you work out which camp you're in.

Question one: would my best customers lose something they can't easily replace if I shut down tomorrow? The test here is specificity. If the honest answer is "they'd just use ChatGPT directly" or "they'd switch to the next tool in the category," you have a thin moat or none. If the answer is "they'd lose a specific workflow, integration, or data layer that would take weeks to rebuild," that's a moat. It's narrow, but it's real.

Question two: do I collect data that improves the product in ways my competitors can't access? This is the flywheel question. Every use of your product should either produce data that makes it better, create a switching cost, or both. If your product works the same way for user number one as it does for user number ten thousand — if there is no compounding effect — you have built a feature, not a business.

Question three: is my distribution a moat? This is underrated. If you have a genuine distribution advantage — deep relationships in a specific industry, an existing audience, a partnership that gives you preferential access to a customer segment — that is a real competitive advantage even if your product is technically undifferentiated. The model that is best in class today may not be best in class in eighteen months. Your relationships, reputation, and positioning in a specific community are harder to replicate than any AI integration.

Question four: would replacing the underlying AI model require rebuilding a meaningful part of my product, or just a configuration change? If swapping GPT-4 for Claude or Gemini would be a one-day task, your product is model-agnostic in a way that is both flexible and fragile. Flexible because you can update quickly. Fragile because your product provides no lock-in at the AI layer, and a competitor can match your product with similar ease.

Question five: do I have customers who have integrated my product into their workflows in a way that creates genuine switching costs? Switching costs are not the same as user satisfaction. A customer can love your product and still switch if a better-priced alternative arrives. Switching costs exist when the cost of moving — in time, in lost data, in workflow disruption — is high relative to the benefit of switching. Integrations, data exports, team onboarding, and custom configurations each multiply switching costs.

Question six: is my product solving a problem specific enough that a large AI company would not consider it worth their attention? This is the defensibility of neglect. GPT and Claude are horizontal products. They do many things adequately. The opportunity for vertical AI is to do specific things extremely well for specific users — better than a horizontal tool ever will, because the horizontal tool will never optimise for your user's precise context. A tool built for the compliance needs of small fleet operators is not something OpenAI will ship. The specificity is the moat.

Question seven: do I understand my user's workflow well enough to anticipate what they need before they ask for it? This is the defensibility of expertise. If you have genuine domain knowledge — not just AI product knowledge, but understanding of the specific problem your user is solving — you can build something that a competitor without that knowledge will struggle to replicate. Domain expertise embedded in product design is slow to copy even when the technology stack is easy to copy.

Score yourself honestly. If you can answer yes to three or more of these questions, you have a working thesis for defensibility. If you can answer yes to five or more, you probably have a genuine moat. If you struggle to give a meaningful yes to more than one, you are building a feature that a better-funded competitor will ship within a year.

The wrapper dismissal is not really about the technology. It's about the business model. A wrapper is only a wrapper if it has no compounding advantage — no data flywheel, no distribution moat, no switching costs, no domain depth. If you have those things, the fact that you're calling an API rather than training your own model is irrelevant. Most of the software industry runs on APIs. The question is what you built around them.

If you're working through whether your AI startup idea is defensible, Kooio's validation stage is structured specifically to stress-test competitive dynamics and differentiation assumptions. The goal is not to talk you out of your idea — it's to identify the places where the thesis is thin before you've spent six months building.