Why this matters
Most AI startups that fail never had product-market fit. Many of them believed they did, because they mistook excitement for need. This course teaches one method per level, using a single running case:
Running case: an AI assistant for small accounting firms (2–15 staff). It categorises client bank transactions and drafts month-end reconciliations for a bookkeeper to review.
In this level you will:
- Learn a working definition of product-market fit (PMF).
- Sort ten signals from the case into "evidence of fit" and "not evidence of fit".
- Write the case's starting PMF hypothesis.
Task 1: Learn a working definition
The term was popularised by Marc Andreessen in a June 2007 essay. He credited the idea to the investor Andy Rachleff and defined it as "being in a good market with a product that can satisfy that market." He also described what it feels like when fit is missing: customers don't quite get value, word of mouth doesn't spread, usage grows slowly and sales cycles drag.
That definition is useful, but it is too broad to test. In this course we use a sharper working definition:
Product-market fit means a specific customer repeatedly uses, and pays for, your product to solve a specific painful problem, and would be clearly worse off without it.
The definition has three parts you must name:
- A specific customer. Not "accountants", but "owner-managers of accounting firms with 2–15 staff that do monthly bookkeeping for small businesses".
- A specific problem. Not "admin is slow", but "month-end bank reconciliation takes each bookkeeper two to three days".
- A repeated behaviour. Not "they liked it", but "they run every month-end close through the product and keep paying".
Analogy: a gym gets plenty of January sign-ups. Fit is shown by the members who are still coming in on a wet Tuesday in March, not by how many people joined.
Do this: write the definition in your own words and underline the customer, the problem and the behaviour.
What PMF is not
These signals feel like progress, but none of them proves fit on its own:
- Launch buzz. Upvotes, social posts and waitlist numbers measure curiosity, not need. AI products draw extra curiosity simply because they are new.
- Investor interest. Investors are betting on what might happen. A meeting request is not a customer.
- Pilot sign-ups. A free pilot costs the customer almost nothing to say yes to. The pilot only tells you something when it turns into repeated use and payment.
- Demo applause. A polished demo on clean, hand-picked data is not daily use on messy client bank feeds.
- Compliments. "I'd definitely use this" is a prediction about the future, and people are polite.
The test that separates the two groups: does the signal show a real customer changing their behaviour and giving up time, money or habit, again and again?
Task 2: Sort ten signals from the case
Do this: for each signal, ask "Is this repeated behaviour by a target customer, at a cost to them?" Cover the answers and try it yourself first.
| # | Signal | Verdict |
|---|---|---|
| 1 | 2,000 waitlist sign-ups after a launch post | Not evidence |
| 2 | A venture fund asks for your pitch deck | Not evidence |
| 3 | A conference audience applauds your live demo | Not evidence |
| 4 | 12 firms sign up for a free 30-day pilot | Not evidence (yet) |
| 5 | A firm partner says "This is amazing, I'd definitely use it" | Not evidence |
| 6 | 7 of 10 pilot firms run every month-end close through the product for three months in a row | Evidence |
| 7 | A firm pays full price after the free trial ends, with no discount | Evidence |
| 8 | Bookkeepers complain straight away when the tool is down for an afternoon | Evidence |
| 9 | A firm owner refers another firm without being asked | Evidence |
| 10 | Use spreads from one bookkeeper to the whole team at a firm | Evidence |
Notice the pattern. Signals 1–5 cost the customer nothing and say little about next month. Signals 6–10 show habit, payment, dependence and spread.
Task 3: Write your starting hypothesis
PMF is something you prove, not something you claim. Start with a written hypothesis you can show to be wrong.
Do this: fill in the four blanks.
- Target customer: who exactly?
- Problem: what painful, frequent task?
- Product: what does it do for them?
- Observable result that would prove fit: what behaviour, how much, by when?
Example for the case:
- Customer: owner-managers of UK and US accounting firms with 2–15 staff that do monthly bookkeeping for small-business clients.
- Problem: month-end bank reconciliation takes each bookkeeper two to three days, and errors surface late.
- Product: an AI assistant that categorises transactions and drafts reconciliations for a human to review.
- Proof of fit: at least 6 of 10 paying firms run their month-end close through the product for three consecutive months, and renew without a discount.
Rules for a good result statement:
- It names a behaviour you can observe, not an opinion.
- It has a number and a time window.
- It could clearly fail. If no result could prove you wrong, it is not a hypothesis.
Experiment: audit your own signals
- List every positive signal your product (or idea) has received so far.
- Mark each one "evidence" or "not evidence" using the test above.
- Write your four-part hypothesis.
How to read the result: if most of your signals are buzz, interest, sign-ups or praise, you do not yet know whether you have fit. That is normal at this stage. Your hypothesis now tells you what behaviour to look for next.
Key takeaways
- PMF means a specific customer repeatedly uses and pays for your product to solve a specific painful problem.
- Andreessen's original definition: being in a good market with a product that can satisfy that market.
- Buzz, investor interest, free pilot sign-ups, demo applause and compliments do not prove fit.
- Real evidence is repeated behaviour that costs the customer something: habit, payment, dependence, referrals, spread.
- Write a hypothesis that can fail: customer, problem, product and an observable result with a number and a deadline.