How to Measure Product-Market Fit｜A Practical Guide to Metrics and Diagnosis

Hello, this is Ryuta Hamamoto from TIMEWELL.

Whenever I talk with people building a new business, there's one phrase that always comes up: "It kind of feels like it's growing." That "kind of feels" is a trap. Still working off a gut sense they can't put into numbers, they start hiring, go out to raise money, and pour budget into ads. Then, when they look back six months later, the users they gathered have quietly slipped away and only the cash on hand has shrunk—I've watched this play out many times. "It feels like it's growing" and "it is genuinely growing" are two completely different things.

What closes that gap is measuring PMF (product-market fit, the state in which a product has been accepted by its market). This article is one installment in a series that walks step by step through how to build a new business; the overall map is laid out in the Complete Guide to New Business Frameworks. This time, we'll work through—hands on, using a fictional new business—how to measure PMF not by mood but by numbers and feel. Before you read on, it helps to take a quick measure of how far your business is really putting AI to use with the AI Literacy Check; the AI techniques I introduce later will click into place better once you have.

PMF measures not "have you achieved it" but "how likely you are to have achieved it"

A lot of people frame PMF as a 0-or-1 question—have you achieved it or not—and this is the first stumble. Things go better if you think of PMF not as something to paint black or white, but as measuring how close you are to achieving it—a degree of likelihood. Try to see it in black and white and you'll only gather the evidence that suits you, misjudging it through confirmation bias. That's why you read the gradient of likelihood with two eyes: the quantitative and the qualitative.

Untangling four definitions, gently

The term PMF has several famous definitions. Andy Rachleff, who popularized the phrase, called it "a state where a valuable market and a product that can satisfy that market mesh together." The point being that if the market isn't good, it doesn't matter how much you polish the product. Marc Andreessen was more visceral, describing it as "when PMF is happening, you can always feel it." Users show up on their own, servers scream, hiring can't keep pace. It's that sense of being pulled along.

That said, "you can feel it" leaves the person actually building the business in the lurch. When you're in the thick of it, that very sense is what gets warped by confirmation bias. This is where the 40% rule proposed by growth expert Sean Ellis, and the idea of the HXC (High-Expectation Customer, the high-expectation core customer) systematized by Superhuman's Rahul Vohra, come in. These two translate feel into numbers and direction. The practical work in this article runs on those latter two.

"A hint," "achieved," and the signals that fool you

If you look at it as likelihood, it's easier to handle in three stages. Pre-PMF, where there's still no sign of it; PMF Hint, where a portion of customers start giving you real traction; and PMF Achieved, where multiple metrics line up. Many businesses mistake the Hint stage for "achieved," rush into scale investment, and stall.

Before we go further, let's throw out the "faint signals" you must not use to judge. The one I find most dangerous in the field is a rise in active user count. You can manufacture that at will with acquisition tactics. Likewise: a high NPS (it surveys existing users, so it carries survivorship bias), praise from media or investors, a long waitlist, one big contract with a single company. Every one of these feels good, but none of them is evidence of PMF. Base your judgment on them and you'll get the whole call wrong. Clear these five off the table first, and then line up the real metrics.

Measuring quantitatively: the retention curve and the Sean Ellis test

There are two pillars on the numbers side. The retention curve, which shows whether users are sticking, and the Sean Ellis test, which measures how much it would hurt to lose the product. Let's measure them in order, using a fictional business as our example.

Our example is a fictional B2B SaaS called "GenbaNote." It's a service for regional construction firms that has AI automatically shape a site supervisor's spoken voice memos into a daily construction report; the subscription is ¥4,800 per site per month. Let's set the scene as five months post-launch, with 34 paying firms. We'll diagnose this business together all the way through.

The three shapes of the retention curve

The most reliable quantitative signal is whether the retention curve has flattened. You split users into cohorts (groups that started in the same period) by their sign-up week or month, and draw a line showing what share remains as time passes. If that line stops falling somewhere and levels into a flat plateau, it means a layer of people who "can't get by without this" has taken hold. Conversely, a curve that keeps falling forever toward zero means there's no PMF for that product. I'll say it flat out: a product whose curve doesn't flatten has no PMF.

You read the curves in broadly three shapes.

Curve shape	Distinctive form	How to read it
Flat type (never bottoms out)	Falls in a straight line	No PMF. Churn hasn't stopped
Smile type	Dips and then recovers later	Weak signal. Propped up by resurrection mechanics
Power-user type	Levels off with a stable layer after early churn	Strongest signal. A core has taken hold

Let's plug in GenbaNote's numbers. 68% remain at Month 3, and for the more recent cohorts the slope from month three onward has been getting gentler. It isn't a fully horizontal line yet, but it's at the stage where partial flattening is starting to show. What matters here is not to judge on a single curve that blends all users together. Split it by cohort and check whether newer cohorts sit above older ones—in other words, whether the product is getting better month over month. While you're at it, inspect whether your definition of "active" is too loose. If you count someone as active just for logging in, the numbers look better than reality. You should anchor on the action that, "if it went away, the core value would vanish"—for GenbaNote, one daily report being auto-generated.

The Sean Ellis test and the 40% rule

The other pillar is the Sean Ellis test. You ask users, "How would you feel if you could no longer use this product tomorrow?" and have them answer on a three-point scale: "very disappointed," "somewhat disappointed," "not disappointed." If "very disappointed" comes to 40% or more of the whole, it's considered a strong signal of PMF. This is the famous 40% rule.

The reliability of the result rides entirely on how you distribute it. Send it to users who have done the core experience at least twice. Gather 40 or more responses, ideally 100 or more. Don't skew it toward friends or die-hard fans; get as close to random as you can. Break these three and the number that comes out becomes a tool for fooling yourself. And don't stop at the three-point choice—always attach the free-text questions that follow. This is what pays off later.

[Sean Ellis Test, 4 questions (ready to distribute as-is)]

Q1. How would you feel if you could no longer use "(product name)" starting tomorrow?
    ○ Very disappointed   ○ Somewhat disappointed   ○ Not disappointed

Q2. Who—what kind of role or situation—do you think gets the most value from this product?
    (free text)

Q3. In a single line, what's the biggest benefit you get from this product?
    (free text)

Q4. What could get better, and how, to raise the value it has for you even further?
    (free text)

* Timing: after the user has completed the core experience at least twice
* Target responses: n≧40 (100 or more if possible) / get as close to a random sample as you can

GenbaNote's result was 37% "very disappointed" (n=52). Just short of 40%—in the 25–39% band. Combined with the hint of retention flattening, you can read the stage as leaning toward PMF Hint. Reading 37% as "so close to 40%, basically about to hit it" is premature. The number is only a signal. The real treasure lies in the free-text answers from the "very disappointed" layer, which we'll get to in the next section.

Measuring qualitatively: identifying the HXC to set your improvement direction

Where numbers tell you your stage, feel tells you the direction—where to polish next. This is where the method practiced by Superhuman's Rahul Vohra pays off. He raised the Sean Ellis score from 22% to 58%, and the key was not to satisfy everyone.

Here's how it goes. First, define the layer that answered "very disappointed" as your HXC (High-Expectation Customer). Narrow their Q2 answers down to three or fewer shared attributes. Next, from Q3, pull out the verbs and adjectives they keep repeating and treat those as the "true core value." Then read only the Q4 of the "somewhat disappointed" layer among those who fit the HXC attributes, and turn the improvement requests written there into a high-priority roadmap. What matters here is deciding whose voices to ignore. The requests from the "not disappointed" layer, and from people outside the HXC attributes, you deliberately cut. The harder you try to be liked by everyone, the more your features scatter, and the further you drift from PMF.

Let's look at GenbaNote. Reading the "very disappointed" layer's Q2, the common thread was "site supervisors aged 40 or older who juggle multiple sites at once." In Q3, phrases like "the time daily reports used to eat has disappeared" and "I no longer have to go back to the office" came up again and again. From this you can narrow the true core value down to "reducing the paperwork after the site wraps to zero." In other words, what to polish wasn't accounting integration or photo management—it was the experience of a supervisor juggling sites finishing the daily report in three minutes, right at the site. This direction never comes out of staring at numbers. It comes only from the customer's words.

This work of putting into words "whose value, and what kind" runs continuously with the matching of customer pains and offered value that we covered in How to Write a Value Proposition Canvas. Measuring PMF is also checking the answers to your value hypothesis.

Pinning down your stage with a 20-point scorecard

Once the quantitative and qualitative are both in, you put them together on a single scorecard to lock in your stage. Judging on numbers alone, or feel alone, is like driving with one eye. I use a 20-point scorecard that scores five quantitative items and five qualitative items at two points each.

Perspective	What you look at (2 points each)
Quant 1	Whether retention has flattened
Quant 2	The "very disappointed" rate on the Sean Ellis test
Quant 3	Weekly growth rate (B2C) or NRR (B2B)
Quant 4	Share of new users from organic or referral
Quant 5	Level and trend of monthly churn
Qual 1	Whether customers say "I couldn't live without this"
Qual 2	Frequency of referrals and inquiries
Qual 3	The quality of questions reaching support (how-to, or deeper requests)
Qual 4	Whether you can state the HXC's attributes in one sentence
Qual 5	Whether you can state the true core value as a verb

For the judgment, gauge how many of the four core conditions you meet at once. The core conditions are: retention has flattened; Sean Ellis is 40% or more; organic or referral is 20% or more; and weekly growth of 5–7% (B2C) or NRR of 110% or more for four to eight straight weeks. If you have a segment that meets three or more of these at the same time, you may declare "PMF achieved" there—that's the practical working definition.

Scoring GenbaNote, it met 0–1 of the core conditions. NRR was 106%, short of 110%; referral share was 16%, under 20%; and flattening was still only a hint. The scorecard total also stays mid-range. The verdict: PMF Hint. The next move is to narrow the HXC to juggling supervisors, run the Vohra-style improvement loop, and pull the Sean Ellis rate above 40%. What you must not do at this stage is pour budget into paid ads or CAC optimization.

Why avoid ads before PMF? Because a product whose retention hasn't flattened is a leaky bucket. Even if ads efficiently pour in new users, they don't stick and drain away. The more you spend to gather them, the faster they leak and the sooner the money runs out. So at the Pre-PMF or Hint stage, don't spend time on acquisition-efficiency metrics like CAC, LTV, or ROAS. Look only at retention, Sean Ellis, and the unfiltered voice of your core customers. The cue to step into acquisition efficiency is when Sean Ellis clears 40%, retention has flattened, and growth has held for several weeks. Just keeping to that order lets you avoid burning the money. If you're uneasy about whether the market itself is big enough, it settles your judgment here to first check the size of the vessel with Market Sizing (TAM, SAM, SOM).

Pivot or continue: decide your exit criteria in advance, as numbers

Of course, there are times when no sign appears no matter how many loops you run. What kills a business in that moment is usually the emotion of "just a little more effort." An intended 18 months stretches to 24, and before you know it the cash has run out. The only way I know to prevent this is to write your exit criteria on paper, as numbers with a deadline, before you're backed into a corner.

Let me tell you about myself. TIMEWELL used to run an AI-assistant business. User reactions were by no means bad, but retention wouldn't flatten, referrals didn't grow, and the Sean Ellis band never climbed all the way. Judging against the criteria I'd set in advance, I concluded this was a problem with the mesh between market and value, and pivoted to ZEROCK in a way that preserved the AI foundation we'd built up. If I hadn't set those criteria back then, I'd probably have kept repeating "just three more months." Exit criteria are a mechanism to protect your future self from emotion.

For the judgment, first score on the five axes that gauge whether to continue or pivot. Has the founder's passion run dry; has the market itself shifted; have customers entered "willing-to-pay mode"; is there a regulatory wall; and has there been no sign of PMF for six months. Paint each red, yellow, or green. If red keeps coming up, next use the Pre-PMF four-branch diagnosis to separate out the cause. Is it a product problem, a market problem, a go-to-market (GTM) problem, or is it all of them? If the cause is the product or the way you sell, you can fight it with improvement, but if the market is bad, no amount of polishing the product pays off. Which brings us back to Rachleff's point about a "good market."

Once you decide to pivot, choose the pattern that preserves the most of what you can keep. Offer different value to the same customers, deliver the same value to different customers, or change only the business model? Rather than starting over from zero, look for a direction that can build on the customer understanding and technology you've stacked up so far. Even when you decide to continue, always rewrite your next exit criteria on the spot. Set numbers and a deadline together—"if Sean Ellis is below 30% at quarter-end, reconsider," and so on. Then re-measure quarterly and gauge the real PMF on the trend of whether you keep exceeding 40%, not on a momentary gust.

Making AI your partner in measuring PMF (a copy-paste prompt collection)

Every step up to here goes far faster when you use AI as a sparring partner. But the output is raw material for hypotheses, not fact. Always verify the numbers and confirm the customer's voice with your own hands. Below are prompts you can paste straight into Claude or ChatGPT and use. Like GenbaNote, plug in your own business's numbers and spar with them.

First is a current-state diagnosis. Deliberately have it push hard on your self-serving interpretations.

You are a PMF-diagnosis expert who has seen dozens of seed to pre-Series A startups.
From my business data below, judge the PMF stage as "Pre-PMF / PMF Hint / PMF Achieved."

- Business overview: {B2B or B2C / what the service is / pricing model / months since launch / number of paying users}
- Numbers
  - Sean Ellis test "very disappointed" rate: {__}% (responses n={__})
  - Retention curve: {whether it has flattened / what % remains at which week or month}
  - Weekly growth rate (B2C) or NRR (B2B): {__}
  - Share of new users from organic or referral: {__}%
  - Monthly churn: {__}%
- Feel (if available): customers saying "I couldn't live without this," frequency of inquiries or referrals, quality of questions reaching support

What I want you to output
1. How many of the four core conditions are met, in a list
2. Any place I'm mistaking a "clear signal" for a "faint signal" (rise in active count, NPS, media praise, waitlist, one big single contract, etc.)
3. Score me on the 20-point scorecard of 5 quantitative + 5 qualitative, and give the total and verdict
4. The top 3 actions to prioritize over the next 90 days

Push hard, without holding back, on the weak spots in my numbers and any self-serving interpretations.

Next is designing the Sean Ellis test and analyzing the responses you collect. The trick is to have it propose in the direction of narrowing down the HXC.

You are an expert in designing and analyzing Sean Ellis-style PMF surveys. Do (A) and (B) below.

(A) For my product "{product name / one-line description}," create the 4-question Sean Ellis test,
    with distribution timing (after doing the core experience {__} times)
    and distribution copy ({B2B email or in-app modal}).

(B) Analyze the following responses:
    - Q1 tally: very disappointed {__} / somewhat disappointed {__} / not disappointed {__}
    - The free-text Q2-Q4 answers from the "very disappointed" layer (paste as-is): {__}

Analysis output
1. PMF score (very disappointed ÷ total × 100)
2. Extract up to 3 shared attributes of the HXC (the "very disappointed" layer) from Q2
3. From the HXC's Q3, the "true core value" = top 3 most frequent verbs / adjectives
4. From the Q4 of people within HXC attributes among the "somewhat disappointed" layer, a prioritized improvement roadmap
5. Explicitly state the voices to ignore (the "not disappointed" layer, requests from outside HXC attributes)

Do not make "satisfy everyone" proposals. Propose in the direction of narrowing down the HXC.

You can hand off the retention cohort analysis too.

You are a cohort-analysis expert. Diagnose the retention data below.

- Business: {B2B or B2C, industry}
- Definition of "active": {e.g., logs in at least once a week + performs the key action}
- Retention table by cohort (paste)
Cohort, W0, W1, W2, W3, ...
2026-W01, 100%, 45%, 32%, 28%, ...
2026-W02, 100%, 42%, 30%, 26%, ...

Output
1. Curve-type verdict (flat / smile / power-user, with rationale)
2. Whether it has flattened (judge B2C from week 12 onward, B2B from month 6 onward, by the slope)
3. Whether newer cohorts sit above older ones = whether the product is improving
4. Three hypotheses for the "aha moment of habit formation" from the gap between early churners and stickers
5. A check on whether the "definition of active" is too loose and is making the numbers look better than reality

Last is a sparring session to nail the pivot-or-continue decision down with data rather than emotion.

You are a sparring partner for exit / pivot decisions. Push me with data so I don't stall on "just a little more effort."

- Current state: {results of the PMF diagnosis / each metric / how many times and how many months I've run the improvement loop}
- My honest feelings: {why I want to continue / what I'm unsure about}

How to proceed
1. Score the five axes (passion / market shift / whether customers are in willing-to-pay mode / regulation / no sign of PMF for six months) as red, yellow, green
2. If Pre-PMF, identify the cause with the four-branch diagnosis (product / market / go-to-market / all of them)
3. If pivoting, propose the pattern that preserves the most of what I can keep
4. If continuing, put exit criteria in writing together with me, as numbers with a deadline (decided at quarter-end)

If I'm scoring an axis too kindly out of self-interest, be sure to point it out.

How to build this kind of AI fluency into each phase of building a business is exactly the territory where we ride shotgun through WARP, our AI consulting service. The precision of your prompts also changes the answers you get back the more you tune them to your business's context.

Wrap-up

That was a quick run, but we've laid out how to measure PMF from end to end. Let me pull the key points together.

• PMF isn't 0 or 1—it's "how likely you are to have achieved it." Look with both eyes: numbers and feel
• The strongest quantitative signal is retention flattening. A product that doesn't flatten has no PMF
• The Sean Ellis 40% is a signal; the free-text answers (the HXC's voice) are the roadmap
• Lock in your stage with the 20-point scorecard, and don't touch ads before PMF
• Write your exit criteria on paper as numbers with a deadline, before you're backed into a corner

Let me add just one thing. Every metric listed here is nothing more than a translation, from a different angle, of "how the customer feels." If, while chasing the numbers, you've cut down on the number of times you actually talk to customers, that's putting the cart before the horse. When I made the call to shut down the assistant business, what pushed me forward wasn't the total on the scorecard—it was dozens of customer voices telling me "this is the wrong market." Metrics are the compass; what tells you where to go is always the customer's words. If you'd like to measure your own business's PMF together and get concrete all the way to the next move, feel free to bounce ideas around with us at a WARP one-on-one consultation.

---

Sources for the ideas referenced.

• Andy Rachleff's definition of PMF (a valuable market and a product that satisfies it)
• Marc Andreessen's description of PMF in "The Only Thing That Matters"
• Sean Ellis's PMF survey and the 40% rule of thumb
• Rahul Vohra's (Superhuman) HXC-centered PMF engine method
• Paul Graham's benchmark for weekly growth rate in "Startup = Growth"