Asking the AI if the idea is good
It will say yes, with reasons. Ask instead: "Argue that this will fail. What would have to be true for it to succeed, and how would I check?" Then have a second model attack the first one's reasoning.
Validate a business idea with AI
Validate a business idea with AI, without fooling yourself.
The danger of validating with AI is that it is agreeable. Ask one chatbot whether your idea is good and it will find reasons it is. Real validation is the opposite: a ladder of increasingly expensive tests, each one designed to prove you wrong, with AI doing the preparation and the analysis and real people supplying the evidence. This page is that ladder, rung by rung, with the prompts and the numbers that tell you whether to climb or stop.
Free to start. The latest models are sharper critics, which is what validation needs.
Rung one
Score it before you research it
An afternoon, no money. Score the idea 1 to 5 on each test. Under 17 out of 25, change the idea or the customer. This is not the decision; it is the filter that stops you spending a month on something that could not have worked.
| Test | The question | 5 looks like | 1 looks like |
|---|---|---|---|
| Painful | Does the customer lose money, time or sleep to this every week? | They complain about it unprompted. | They shrug when you describe it. |
| Already paid for | Do they spend money or staff hours on it today? | A budget line or a person exists for it. | "We just live with it." |
| Reachable | Can you name where these people gather, or ten of them? | You are already in the room. | You would need ads to find one. |
| Buildable by you | Can a first version exist in weeks, with AI, without a team? | A web tool with a few screens. | Needs hardware, licences, or a year. |
| Repeatable | Is it the same problem for many people? | Every customer wants the same thing. | Every customer wants a custom job. |
Use disagreement as the signal
Have several AIs score the idea independently, with the same description. When they agree on a score, take it. When they split, say 4 and 2 on "reachable", that test is the one you have not actually got evidence for. It goes to the top of the list for rung two.
The ladder
Five rungs, each one harder to pass than the last
Each rung costs more than the one before, in time or money, and each has a pass mark. Climb only when you pass. Stop or change the idea when you do not. The AI prepares each rung and analyses the results; the evidence comes from people who are not you.
Desk research: what do they do today?
Cost: an evening. Find what the customer uses instead: competitors, spreadsheets, a person, nothing. Read competitors' two- and three-star reviews and pricing pages, and put screenshots in the Project. Pass mark: you can name the alternative, its price, and the complaint people have about it. Fail looks like: "there is nothing like this", which almost always means nobody wants it.
Ten conversations with people who have the problem
Cost: two weeks of evenings. Not friends. The AI writes the interview guide: what they do today, what it costs, what they have tried, what they would pay, who else decides. You ask and take notes. Then paste all ten into the Project and ask for the patterns, the surprises, and the quotes. Pass mark: at least six of ten describe the pain unprompted and can say what it costs them. Fail looks like: polite interest and "sounds useful".
A landing page that asks for something
Cost: a weekend and a small ad budget or a few community posts. The AI drafts the page from your interview quotes: the problem in their words, the offer, a price, and a button that asks for an email or a deposit. Build it with the app guide loop and publish. Pass mark: of the people who reach the page, a meaningful share leave an email; if you ask for money, anyone at all pays. Fail looks like: many visits, no action. The message or the price is wrong; go back to the interviews.
Pre-sell it before it exists
Cost: your nerve. Go back to the interviewees and the email list with a concrete offer: "It will do X, it costs Y, it ships in six weeks, pay now and get three months free." The AI drafts the offer and the follow-ups; you send them and have the calls. Pass mark: three to five people pay, or sign something that commits them. Fail looks like: everyone wants to "see it first". That is a no.
The smallest thing that does the job, for those buyers
Cost: four to eight weeks. Build only what the pre-sale promised, with AI, with the buyers' real data in the Project. Measure one number that matters to them. Pass mark: they use it without being reminded, and at least one of them refers someone. Now you have a business. Everything after this is on the launch page.
The traps
Four ways people validate wrong, and the fix for each
Interviewing people who like you
Friends and family validate everything. Interview strangers who have the problem, and never pitch during the interview. Ask about their week, not your idea.
Counting interest as demand
"I would totally use that" costs nothing to say. Only three things count: an email address, a calendar booking, or money. Design every test so that it asks for one of them.
Building first "just to show them"
Weeks of building before anyone has committed. If you cannot pre-sell it with a description and a price, the product will not sell it either. Rung four exists to protect rung five.
Write the decision rule before the evidence arrives
In the Project, before rung two: "If fewer than six of ten interviews confirm the pain, I change the customer. If fewer than three people pre-pay, I stop." Deciding after the results come in is where people talk themselves into it. The AI can hold you to the rule; ask it to.
Free is the start. The latest models are the difference.
Validation needs a sharp critic, not an agreeable one.
The whole point of this page is to find the flaw before the market does. Older and smaller models are agreeable; they find reasons your idea works. The latest models, asked to argue against you with your interview notes and competitor material in the Project, find the objection you had not seen. For a decision that costs months of your life, that is the cheapest insurance available.
Newer models hold your whole Project in view instead of the last few messages, reason through trade-offs instead of picking the first plausible answer, and are wrong less often and with less confidence. For the questions on this page, that is the difference between advice that sounds right and advice you can act on.
Pro from $20 a month, cancel any time. The free plan stays free.
- Your first Project and file uploads
- The major AI families to try the workflow
- A small daily message allowance
- The latest ChatGPT, Claude, Gemini, Grok and Perplexity models, all on the same Project
- Far more messages a day, so a working session does not stop halfway
- Modes where several models draft, challenge and verify each other's answers
- Sourced web research for competitors, prices and market numbers
- Larger uploads, so all ten interview transcripts fit in the Project
Start here
First messages for a validation Project
Add the idea in one paragraph and anything you already know. Then start here.
ScoreScore this idea 1 to 5 on painful, already paid for, reachable, buildable by me, repeatable. Explain each score and name the test you are least sure about.
AttackArgue that this idea will fail. List what would have to be true for it to succeed and how I could check each one this month.
InterviewWrite a 20-minute interview guide for someone who has this problem. No pitching. Questions about their week, what it costs, what they tried, who decides.
PatternsHere are ten interview notes. What patterns appear in at least six? What surprised you? Pull the five quotes I should put on a landing page.
Landing pageUsing these quotes, write a landing page that states the problem in their words, makes one offer at one price, and asks for an email or a deposit.
Pre-sellWrite the pre-sale offer and two follow-up emails: what it does, the price, the ship date, and why paying now is worth it. Plain language, no hype.
Honest answers
What people ask
Can AI validate a business idea on its own?
No, and anything that claims otherwise is selling you comfort. AI prepares the tests and analyses the results; the evidence has to come from people with the problem, through interviews, sign-ups and money. What AI does well is make each rung faster and make you a harder critic of your own idea.
How many people do I need to talk to?
Ten is enough to see a pattern; twenty if the first ten split. Strangers who have the problem, not friends. The guide on this page keeps the conversation about them, not about your idea.
What if there are already competitors?
Good: it proves people pay. Read their worst reviews to find the gap. What should worry you is a competitor customers love, or no alternative at all, which usually means no demand.
How long does validation take?
Rungs one to four take about four to six weeks of evenings if you keep moving. Rung five, the smallest product, takes four to eight more. Skipping rungs does not save time; it moves the failure later, where it costs more.
Why several AI models instead of one?
Because a single model agrees with you. Independent answers from several show where the idea is solid and where it rests on an assumption. Disagreement on a score is the most useful output you will get.
What is free?
Your first Project, file uploads and the major model families. The latest models, sourced research and larger uploads are on paid plans. See plans.
Score it tonight. Interview ten people this month.
Put the idea in a Project, ask several models to argue against it, and write your decision rule before the evidence arrives. Free to start.
Also: AI business ideas · start a business with AI · launch a product with AI