Why the cheapest validation method in 2026 is also the one that cannot fail you
A post going round today gives founders three prompts to run before they build. One of them is called a Willingness to Pay Test. It is a good prompt. It is well written. At no point in it does anyone who could hand over money get contacted.
That is not a small omission. That is the entire test.
Can AI validate a startup idea?
It can validate that your idea is coherent. It can tell you the category exists, that other people have solved adjacent problems, and roughly what such products have charged. All of that is real information and none of it is demand.
A model answers from what has already been written down. Your buyer's decision has not been written down yet, because it has not happened. Asking a model what your customer would pay is asking a very well read stranger to guess on their behalf.
The uncomfortable part is that the guess arrives with confidence, formatted, in seconds. Confidence is exactly what a founder is shopping for at that moment.
What is the difference between stated interest and revealed preference?
Stated interest is what someone says when saying it costs them nothing. Revealed preference is what they do when it costs them something.
A founder posted today that he had a hundred users on his app before launch, no revenue and no monetisation plan. Usage was proven. Payment was never tested. Those are two different questions and only one of them was answered.
Another founder discovered today that his analytics numbers were mostly bots. The metric he had been steering by was not measuring people at all. Same failure in different clothes: a number that felt like demand and was not.
Why do synthetic respondents feel so convincing?
Because they are fluent, and fluency reads as knowledge.
A company pitching investors this week describes millions of hours of past customer interview data used to help entrepreneurs validate ideas before they build. That corpus is real. What it contains is what people said, to someone else, about something else, some time ago.
The expensive part of a customer interview is the same thing that makes it worth having: a specific person spending real time today, on your thing, with the option to walk away. Remove that and you have kept the format and thrown out the substance.
What actually counts as payment evidence?
Something a person did that cost them.
Money changing hands is the cleanest version. A pre-order, a deposit, a signed pilot, a switch away from a tool they already pay for. Below that, time given up on purpose: an hour on a call they gained nothing from, a detailed complaint written unprompted, a referral to a colleague.
Note what is missing from that list. Survey responses. A poll on a founder forum. A thumbs up from your co-founder. An estimate from a model. A landing page signup that took four seconds and cost nothing.
One founder today asked a forum of other founders to pick his positioning for him. Every person answering was someone who would never buy the product.
What should you do with the model's answer?
Treat it as a hypothesis with a name on it.
If the model says your buyer would pay forty a month, that is now a claim you can go and test against a person. Take the number to someone who fits the description and watch what happens to their face. The value of the model output is that it gives you something falsifiable to carry into a real conversation.
The failure mode is treating the output as the conclusion and skipping the conversation, which is what happens roughly every time, because the conversation is slow and awkward and the output is instant and flattering.
Key takeaways
- A willingness to pay test with no payer in it measures your prompt, not your market.
- Models can report what was said in the past. They cannot report what your buyer will do next.
- Stated interest costs nothing to give, which is precisely why it predicts nothing.
- Usage without payment is a proven habit and an untested business.
- Real evidence is behaviour that cost someone money or time they will not get back.
- The right use of an AI estimate is as a hypothesis to falsify, not as a verdict to celebrate.
FAQ
Is it wrong to use AI for idea validation at all?
No. It is fast and useful for mapping a category, sharpening your questions and pressure testing your own reasoning. It stops being validation the moment it replaces contact with a person.
How many real people do I need before I trust a signal?
Fewer than most founders think, if the people are right and the behaviour is costly. One buyer who pays teaches you more than two hundred survey responses from people who will not.
What if I cannot reach my target buyer?
That is itself a finding worth taking seriously. If a buyer is unreachable at the idea stage, they will not become reachable when you have a product and a burn rate.
Do landing page signups count as willingness to pay?
They count as willingness to type an email address. Useful as a first filter, never as a pricing signal.
How do I ask about price without leading the answer?
Ask about the past instead of the future. What did they buy last time this problem got bad enough, what did it cost, and what happened after.