How do you find product/market fit for an AI product?
Short answer: The same four variables, in the same order: ICP, value proposition, product, business model. What changed is the cost of building, which collapsed, and the noise in early signals, which exploded. AI products need harder evidence than usage (people paying, coming back, committing), a sharp answer to "why not just use ChatGPT?", and a product that holds context a general model can't.
I can build an MVP in a day now. So can you.
That's exactly the problem.
AI reduced the cost of building. It didn't eliminate the cost of being wrong. When building takes a weekend, it's very tempting to skip the boring part, figuring out who it's for and why they'd care, and go straight to the fun part. Building is fun. Being wrong is not.
The order hasn't changed
P/MF is still four variables, solved in order:
- ICP: who is desperate enough to try you
- Value proposition: the promise, in their words
- Product: how you deliver it (this is where AI lives)
- Business model: how value comes back to you
AI makes step three faster. It doesn't let you skip steps one and two. If anything it makes them more important, because now everyone can build the same thing you can.
Why early signals lie more for AI products
AI products are novel. Novelty attracts curious people who try everything and stay for nothing.
So the usual early signals get noisy:
- Sign-ups are cheap. Curiosity isn't demand.
- "Wow" in a demo is a reaction to the technology, not to your value.
- First-month revenue can be tourists paying to play.
Validation doesn't necessarily mean sales, and sales don't necessarily mean validation. Look for movement instead. Real pain moves. Phantom pain complains.
| Signal | What it tells you |
|---|---|
| Tried it once, said "cool" | Nothing yet |
| Came back the next week without a nudge | Possible fit |
| Uses it for the same job every time | There's a real job here |
| Pays, renews, or asks for more seats | Value |
| Complains loudly when it breaks | They rely on you |
The question you have to answer: why not just ChatGPT?
Every AI founder will hear it, from customers and investors alike. If your honest answer is "we have a nicer prompt", you don't have a product yet. You have a feature that a model update can delete.
The answers that hold up are about context.
AI is only as good as the context it has. A general assistant knows a little about everything and nothing about your customer's situation. Products that win own a piece of context the general model doesn't:
- They hold it: your customer's data, history, workflows, decisions.
- They manage it: pulling scattered sources into one place the AI can use.
- They apply it: to one specific job, for one specific ICP, better than anyone.
Ask yourself: what context does my product hold, manage, or miss? If you can't name your context category, you don't understand your own product yet.
Be careful with your own demos
AI makes it easy to show something that looks finished.
I recently pulled AI-generated mockups out of a client's demo because they were too good. They promised more than the product could deliver, and the buyers would have judged the real thing against the fantasy.
Excitement isn't a business outcome. Set expectations you can meet, then beat them.
A real example
One client built AI product photography for online sellers. The assumption was that fashion brands would pay the most.
We ran the same offer to several segments. Drop shippers converted far better, which also opened Latin America as a market nobody had planned for.
The technology didn't change. The ICP did. That's the order working as it should: the market told us who it was for.
Common mistakes
- Building first because building is cheap. Cheap to build isn't cheap to be wrong.
- Leading with the model. Nobody buys "powered by GPT". They buy their annoying problem gone.
- Mistaking tourists for customers. Measure who comes back and who pays, not who signs up.
- No answer to "why not ChatGPT?". If you can't say what context you own, a model release can erase you.
- Pricing like SaaS when your costs scale per use. Make sure the business model captures value at the rate you spend it.
What to do next
- Answer three questions before you build: who is this for, why should they care, and what's the minimum to test it?
- Write your answer to "why not just ChatGPT?" in one sentence. If it's about prompts, keep working.
- Name the context your product holds or manages.
- Pick one movement signal (return usage, paid pilot, renewal) and set the number that means GO before you launch.
Not sure what to build first? Read what a product team should build before P/MF.
Building an AI product and not sure the market is pulling yet? Tell me what you're seeing.