How to validate a startup idea before you build it (2026)

Validating a startup idea means getting evidence, before you build much, that specific people will change what they do or pay you to solve a problem. The evidence that counts is a commitment: money, time, an introduction, a pre-order or a signed letter of intent. Compliments, a "yes" on a survey and likes on a post don't count, because they cost the person nothing.

This guide covers what counts as evidence, six low-cost tests, the pass/fail line to write down first, and how validation flatters you into building the wrong thing. We don't sell a validation tool, so there's no score to buy here, and we won't quote conversion benchmarks, because no reliable universal one exists.

What counts as evidence, and what doesn't

Evidence is anything that costs the other person something real. YC Group Partner Gustaf Alströmer says in How to talk to users that if you ask "Will you use our product?" people will probably say yes, "but that's not going to mean anything to you". Usability researcher Jakob Nielsen makes the same point in First Rule of Usability? Don't Listen to Users (August 2001): watch what people do, and "definitely don't believe what people predict they may do in the future."

Counts as evidence:

• Money: a deposit, a pre-order, a paid pilot or a first invoice paid.
• Time: a second call, a screen-share of how they do the job today, or a trial they actually use.
• Reputation: an introduction to their boss, the person who signs, or a peer with the same problem.
• Written commitment: a letter of intent that names a price, a start date or a volume.
• Existing spend: they already pay for a workaround, a consultant or another tool.

Doesn't count:

• "Great idea, I'd use that."
• A survey where people tick "yes, I'd pay".
• Likes, upvotes, follows and comments.
• Email sign-ups with nothing asked of the person.
• Praise from friends, family and other founders who aren't the buyer.

Paul Graham's test in How to Get Startup Ideas (November 2012) is a good filter: ask "who wants this right now?" If you can't answer, he says, the idea is probably bad. YC's essential startup advice adds that 10 customers with a burning problem are much better than 1,000 with a passing annoyance.

Write your pass/fail line before you test

Decide what result would make you stop, and write it down before the test starts. If you set the bar after you see the numbers, you'll set it just below whatever you got.

A useful line has four parts: who, what commitment, how many, and by when. A made-up example of the format: "If fewer than 3 of the 15 practice managers I interview agree to a paid four-week pilot by 30 November, I drop this version of the idea." The numbers should come from what your business needs to work, not from someone else's average.

YC's Michael Seibel gives the same advice for product specs in How to plan an MVP: write your spec down, because it's easy to keep changing what you're working on when you never write it down. He also says to "hold the problem you're solving tightly, hold the customer tightly, hold the solution you're building loosely".

The Lean Startup principles call this validated learning, and say progress tracking needs actionable metrics that show cause and effect. A pass/fail line written in advance is the simplest version.

Six ways to validate an idea, compared

Pick the cheapest test that asks for a real commitment. The times below are our rough planning guide for a founder working part-time on validation, not measured averages.

Most founders run interviews first, then a test that asks for money.

Method 1: Problem interviews

Problem interviews tell you whether the problem is real, how often it happens and what people already do about it. They cost nothing but time.

How to run it: talk over video, phone or in person, not by survey. Alströmer says you can learn more from a five-minute video interview than from 500 or 5,000 survey responses, and that the most important rule is not to introduce your idea until the end of the call, or not at all. His suggested questions include "Tell me how you do X today", "What is the hardest thing about X?", "Why is it hard?" and "What do you do to solve this problem for yourself today?" Ask them to screen-share how they do it now.

He also lists what not to ask: whether they'll use your product, which features would make it better, yes-or-no questions, what a better product would look like, and two questions at once. After five or ten interviews, he suggests sorting your notes into problem buckets, writing down your conclusions and forming a hypothesis to test with the same people. To judge whether the problem is valuable, he asks whether people already pay for other solutions, and whether they're happy with a spreadsheet.

Rob Fitzpatrick's The Mom Test is the standard handbook. Its core idea, in our words: people are kind about your idea, so ask about their life and their past behaviour, where kindness can't distort the answer, and look for signs they'll actually buy. The book's subtitle says it's for learning whether your business is a good idea when "everyone is lying to you".

Stop when: most people can't describe a recent time the problem hurt, nobody spends money or effort on it today, or nobody agrees to a second step.

Method 2: A landing page with a real price and a pre-order or deposit

A landing page tests whether strangers will pay, but only if it shows a real price and asks for real money. An email box tells you about curiosity; a payment tells you about demand.

How to run it: write one page that names the buyer, the problem in their words, what they get, the price and when it ships. Add a payment button. Stripe's Payment Links docs say you can create a payment page without code, share it by link, QR code or a buy button on your site, sell a one-off product or a subscription, limit the number of purchases and refund from the dashboard. That limit is handy for a "first 20 founding customers" offer.

Cost: Stripe's pricing page says it charges no setup or monthly fees, and lists 2.9% + 30 cents per successful domestic card payment in the US, plus 1.5% for international cards (checked 26 Oct 2026). Fees differ by country. For the page itself, Carrd Pro starts from US$9 a year, and the plan with forms and a custom domain is US$19 a year (checked 26 Oct 2026). If you sell to buyers in many countries, our guide to merchant of record vs Stripe covers the tax side.

Be honest on the page about what exists and when it ships, and refund anyone who asks. If you take money, the rules on supplying what you sold apply:

• Australia: the ACCC's page on supplying products or services that are paid for says businesses must not accept payment if they don't intend to supply, intend to supply something different, or know (or should know) they can't supply by the promised date or within a reasonable time.
• United States, physical goods: the FTC's Mail, Internet, or Telephone Order Merchandise Rule says you must ship within the time you state, or within 30 days if you state none, ask the customer's consent to any delay, and offer cancellation with a full refund. Refunds are generally due within seven working days of cancellation. The FTC says the rule doesn't cover services.

Stop when: the people you sent to the page match your buyer, and almost none of them pay. Getting traffic to the page is its own problem, and our guide to getting your first customers without ads covers it.

Method 3: Concierge or manual delivery

Concierge validation means delivering the result by hand for a few customers before writing the software. You learn what the job really involves, and whether they value the result enough to pay.

Paul Graham's essay Do Things that Don't Scale (July 2013) describes a variant where you don't just use your software but "are your software", and notes that with a small number of users you can do by hand things you plan to automate later. He describes Stripe's founders manually signing merchants up for traditional merchant accounts behind the scenes, and Meraki and Pebble assembling their first products themselves. YC's essential advice says to get your first customer by any means necessary, "even by manual work that couldn't be managed for more than ten" customers.

How to run it: pick three to five customers, agree on the outcome, deliver it with spreadsheets, email and your own hours, and charge something. Seibel says some startups start with just a landing page and a spreadsheet.

Stop when: customers don't come back for a second round, won't pay once the first round is done, or the manual work shows the problem can't be solved at a price they'll pay.

Method 4: A fake-door test, disclosed honestly

A fake-door test shows a button, plan or feature that doesn't exist yet and counts who clicks it. It's useful for choosing between options, and weak as proof anyone will pay.

How to run it: add the option to an existing site, product or email, at its real price. When someone clicks, tell them straight away that it isn't built yet, why you asked, and offer a waitlist or a short call. Then ask the clickers for a commitment, because a click is a cheap signal.

Don't take payment behind a fake door. Under the ACCC guidance above, accepting payment for something you don't intend to supply breaches the Australian Consumer Law.

Stop when: the option gets clicks but nobody who clicked will take the next step.

Method 5: A working prototype built in a day with AI tools

AI coding tools mean you can often put a rough working version in front of buyers within a day, which beats describing it. Our guide to AI coding tools for founders compares the options and their prices.

Seibel's advice still applies: "launch something bad, quickly", build in weeks, not months, and time-box the spec to what fits the deadline. With AI tools, a one-day time box is realistic for a narrow prototype. Alströmer's tip for testing prototypes is to give people a goal, not instructions, then watch and ask them to think aloud.

Keep AI on the right side of the line. It can draft interview questions, summarise transcripts and build the prototype. It can't stand in for a buyer. Asking a model whether your idea is good is the most flattering validation there is. If an assistant helps run the admin, see how solo founders use AI agents.

Stop when: people can't complete the core task without your help, or they complete it and still won't pay or commit time.

Method 6: Sell to 3 to 5 people before you build

The strongest test is a signed order before the product exists. We suggest three to five, because one buyer can be a favour while a handful start to show a pattern.

How to run it: take what you learned in interviews, describe the outcome, the price and a delivery date, and ask for a deposit, a paid pilot or a letter of intent that names a price and start date. Seibel's point that "launch" just means getting any customers applies here.

Stop when: buyers who confirmed the problem won't sign anything, even at a lower price or a smaller first scope. If they sign, build only what you promised them.

How to validate an idea without spending money

You can run four of the six tests for nothing but your time: problem interviews, concierge delivery, a fake door on something you already have, and selling before you build. A landing page with a payment link costs close to nothing, since Stripe charges per successful payment rather than a monthly fee.

The real cost of validation is building before anyone has committed.

The validation that flatters you

Most bad validation isn't fake, it's flattering. Three patterns cause it.

• Leading questions. The Nielsen Norman Group's guide to avoiding leading questions (December 2017) defines one as a question that includes or implies the answer you want, and notes that people tend to echo the interviewer's wording. Its example: "How well would this save time for you during your workday?" assumes it saves time. Ask "How might this affect your efficiency, if at all?" instead.
• Friends, family and other founders. Alströmer says people you know are the most likely to reply but can be less honest so as not to offend you. Founders around you are rarely your buyer either.
• Measuring the wrong thing. Nielsen says self-reported data is "typically 3 steps removed from the truth". Sign-ups, survey ticks and page views measure attention. Count commitments instead.

Graham warns about a related trap: ideas that sound plausible but turn out to be bad, which he compares to a sitcom writer's made-up startup. If the evidence only exists in your notes, it isn't evidence yet.

Is there a benchmark for a good validation result?

No reliable universal conversion benchmark exists for idea validation. Landing page and pre-order rates depend on where visitors came from, how closely they match your buyer, the price and the offer. A number that's strong for a warm list of interviewees could be meaningless for cold traffic, so check where any figure you see comes from and what traffic it measured.

The best-known number in this area doesn't apply either. In First Round Review's Superhuman article (November 2018), Rahul Vohra describes Sean Ellis's survey: ask users how they'd feel if they could no longer use the product, and aim for at least 40% saying "very disappointed". It surveys people who've used the product at least twice in the last two weeks, so it measures product-market fit after launch, not demand for an idea.

Set your own line from your economics instead. Work out how many paying customers at what price make the next step worthwhile, then decide what share of the right people you speak to must commit.

Cold outreach rules

If your tests need cold email, follow the rules for each recipient's country: our guide to first customers without ads covers the US, UK and Australian rules with their official sources.

Common mistakes

• Building first. It's the most expensive way to find out.
• Pitching in interviews. It biases every answer after it. Ask about their past.
• Counting compliments. "I'd use that" isn't a commitment.
• Moving the goalposts. Without a written pass/fail line, every result looks like a pass.
• Testing on the wrong people. Friends, other founders and anyone who isn't the buyer.
• Hiding the price. A page without a price tests curiosity, not demand.
• Quoting someone else's benchmark. Their traffic, price and buyer aren't yours.
• Taking money you can't deliver on. It's a legal problem as well as a trust one.
• Asking AI instead of buyers. A model's opinion isn't market evidence.

It's free to request an invite to Remote Builders, which is invite-based, and its builder dinners in Bali and Australia are a good place to sharpen a problem statement, as long as you remember fellow founders usually aren't your buyers.

Checklist: validate a startup idea

1. Write one sentence naming the buyer, the problem and what it costs them.
2. Answer Graham's question: who wants this right now?
3. Write the pass/fail line for your first test: who, what commitment, how many, by when.
4. Book problem interviews with real buyers, not friends.
5. Ask about past behaviour and keep your idea out of it until the end.
6. Sort your notes into problem buckets after five or ten calls.
7. Pick a test that asks for money or time: a priced landing page, concierge delivery, a prototype or a direct sale.
8. If you take money, promise a delivery date you can meet and refund on request.
9. Disclose any fake door straight after the click and never take payment behind it.
10. Count commitments, not sign-ups, likes or survey answers.
11. Compare the result with the line you wrote, and stop, change or continue.
12. Build only what your first committed customers need.