How to Tell If You Have Product Market Fit

Every founder can name the week they decided they had product market fit.

Almost none can name the number that decided it.

The number is the share of your active users who would be genuinely put out if the product disappeared tomorrow. Forty percent, read off at least a hundred responses, with a retention curve that flattens underneath it.

The problem sits in where that number usually gets read. The standard test surveys the people still using you, then asks whether they would have stayed. That is the most flattering sample in your business.

So read it the way an auditor would. A sample big enough to carry weight. A segment cut out of it, and one behaviour that outranks every answer anybody gives you.

Five readings follow. Each has a threshold, and each failure has a bill attached to it.

Product market fit is a ratio with a denominator

Marc Andreessen's definition has held since June 2007: "being in a good market with a product that can satisfy that market" (pmarchive, June 2007).

Read it twice. It names a market and a product, and most founders spend the entire measurement budget on the product half.

The market half has a body count. CB Insights puts poor product market fit at 43% of the reasons failed startups give, ahead of bad timing at 29% and unsustainable unit economics at 19% (CB Insights, read 26 October 2026).

The versions of that study disagree, so here is the disagreement. Earlier cuts, working from about a hundred post-mortems, reported roughly 35% citing no market need. The current analysis covers 385 companies and counts poor fit more broadly at 43%.

Trust the newer count for the size of the problem. Trust neither for the size of yours.

A ratio is what you can actually compute. Write both halves down before you survey anybody:

people who would be very disappointed this vanished, divided by people who used the product in the last 30 days

Then fix the boundary, because the boundary is where the first number goes wrong. Used the product in the last 30 days. Not signed up, and not a trial that never opened the first screen.

A denominator of people who never used the thing produces a beautiful number about nothing, and it usually produces it fast.

The mistake looks like this. A tool has 4,000 emails collected over two years, 900 people who signed up, and 210 who logged in during October. Survey all 1,900 and the responses come back mostly from the curious and the long gone. The share who would be very disappointed comes back at 18% and the founder concludes the market is cold.

The same company, asked of the 210 who used it, might read 38%. Same product. Same month. Different denominator.

Write your line on the sheet you already keep for revenue. One date beside it, one number. A number you cannot recompute later is a mood with a spreadsheet under it.

The 40% test, run with a sample size that holds up

Sean Ellis's version is one question with four answers. How would you feel if you could no longer use this product? Very disappointed, somewhat disappointed, not disappointed, not applicable.

He set the 40% line after comparing results across close to a hundred companies. The essay that made the method famous walks a company through it from 22% upward (First Round Review, May 2018).

Run it in an hour, in five steps.

1. Pull every account that used the product in the last 30 days. Export the list rather than guessing at it.
2. Send one email. One question, four answers, one link. Not a twelve-screen survey tool that asks for the company's headcount first.
3. Add a second question: what is the main benefit you get from it? Those answers become your homepage copy this week.
4. Wait for 40 responses or seven days, whichever is slower. One reminder at day three, to non-responders only.
5. Divide very disappointed by responses, then write that number and the sample size on one line with the date.

Step four is where most of these numbers die, because the sample size decides what the reading is allowed to mean.

| Responses | Range around a 40% reading | | 25 | plus or minus 19 points | | 40 | plus or minus 15 points | | 100 | plus or minus 10 points | | 400 | plus or minus 5 points |

At 40 responses, a reading of 40% runs from 25% to 55%. You cannot separate fit from no fit with that in front of you, and neither can anybody you send it to.

A hundred responses is the smallest sample that lets you act. Four hundred turns it into a figure you can quote without a caveat.

Under a hundred active users you have a real constraint and one honest way out. Widen the window to 90 days, treat the score as directional, and let the retention table carry the verdict. Say which of the two you did when you report it.

The email itself matters more than it should, and the rules for getting it answered are the same rules as a launch email. This one gets replies:

Subject: one question about your account

We're trying to make this better and I need a straight answer from you, not a survey.

How would you feel if you could no longer use it? Very disappointed, somewhat disappointed, not disappointed, or you never really got started.

One more line if you have it: what is the main thing you use it for?

No incentive, no logo, no five star rating request. The people who reply to that email are the ones who care enough to reply, which is exactly the bias the retention curve corrects for.

The segment you cut before you believe the number

Superhuman's published numbers are why this step comes second and not last. Their first survey came back at 22%. They split the responses into segments, found the group where fit was already strongest, and were at 33% with no product change at all.

Three quarters of product work later, 58%.

That first jump was a measurement decision, and it is the cheapest thing available to you this week.

Cut the responses five ways, in this order.

1. Acquisition source. Newsletter, paid search, a partner's audience, a referral link.
2. Use case, taken from their answer to the main benefit question.
3. Size or plan, whichever splits your base more honestly.
4. Time to first value, same day against same week.
5. Who answered, the owner against the person who inherited the login.

Keep every cut with at least 20 responses and read the share inside it.

Worked example. A scheduling tool at $29 a month has 340 signups across 90 days and 96 survey responses, 34 of them very disappointed. That reads 35.4%, with a range from 26% to 45%. It tells you nothing you can spend money against.

Cut by acquisition source and the picture separates.

| Segment | Responses | Very disappointed | | Newsletter readers | 62 | 27 (44%) | | Paid search | 34 | 7 (21%) |

The fit sits in the newsletter crowd. The paid crowd bought the same product and got a weaker version of it, which usually means the landing page promised something wider than the product delivers. That page needs a narrower promise and a call to action that matches it.

At 62 responses the newsletter interval runs roughly 32% to 56%. That cut is worth acting on because it agrees with the retention table, not because it is precise. Precision needs the next two months of responses.

Here is the guardrail, and it is the difference between segmenting and lying to yourself. Choose the cut before you look at the answers. Write down "acquisition source first, then use case" on Monday and read it on Friday. A founder who tries five cuts after seeing the numbers will always find a segment above 40%, and it will always be an accident.

Your move after that is not a feature. Rewrite the promise in the newsletter crowd's own words, using their benefit answers verbatim, and point the paid landing page at that narrower promise. Then watch the paid segment's score at the next read.

The retention curve that outranks the survey

A survey is a claim. A curve is what people did.

Build the table in the spreadsheet you already have, in four steps.

1. One row per signup month, six months back at minimum.
2. One column each for day 7, 14, 30, 60 and 90.
3. Each cell holds the share of that month's cohort that did the core action in that window. Logging in does not count. The action that delivers the value counts.
4. Read the day 30 column for cold arrivals first, then read the shape of the whole row.

Two cohorts, same product, same ninety days, different doors in.

| Day | Paid search | Referral | | 0 | 100% | 100% | | 7 | 22% | 64% | | 14 | 11% | 48% | | 30 | 6% | 34% | | 60 | 3% | 32% | | 90 | 2% | 31% |

Read the last two rows, because that is where the verdict is.

The paid cohort is still falling at day 90. Every month discards most of what it bought, so the next month has to buy more, and the month after that buys more again.

The referral cohort lost two thirds of itself in a fortnight and then stopped losing. That flat tail is the thing you are looking for. A group stays, and they stay without being reminded.

Now the arithmetic that makes the shape worth more than any survey score. Take a hundred signups a month, $29 a month, and $6 of delivery cost, so $23 of gross profit per active account.

The referral cohort settles at 31 active users, which is $713 of gross profit every month, and it keeps producing it while they stay. The paid cohort settles at 2, worth $46, and next month's hundred is already needed to replace last month's.

Same price. Same product. One of those two cannot afford to be advertised to, and no survey will tell you which is which.

Two ways this table gets built wrong, both common.

The core action is set too loose, so the curve looks flat because "opened the app" is easy. Set it at the step where the customer first gets the result they came for: the first invoice sent or the first report shared.

The cohort is mixed, so warm and cold arrivals sit in one row. The average of a flat tail and a cliff is a gentle slope, which is the most misleading shape in the table. Split by source before you read anything.

Five readings, one verdict

No single number carries this on its own. Read all five onto one sheet, once a month, and let the count decide what you do next.

| Reading | Pass | Fail | Where to read it | | Very disappointed share | 40% or more at 100 responses | under 25% | the one question survey | | Retention shape | flat tail by day 30 | still falling at day 90 | the cohort table | | Unbought arrivals | 30% or more of new accounts | every arrival has a paid touch | source tags on signup | | Named referrers | 1 in 10 new accounts names a person | nobody does | the referral field | | Payback | cheapest channel inside 12 months | no channel pays back at any budget | spend against gross profit |

Five passes and you stop testing and start spending. Three and you fix the two that failed before you buy anything else.

Two or fewer and you have a distribution problem wearing a marketing hat. More spend buys the same leak at a higher price.

Set the review on the same date every month. The readings move at different speeds, and a comparison against last month is worth more than any single column.

The payback line carries one caution. Use gross profit rather than revenue. A $29 subscription with $6 of delivery cost pays back a $180 acquisition cost in eight months, and the same maths against revenue says six. That gap is the whole decision on whether a channel is worth running.

What to do between 25 and 40 percent

Most products sit here, and most advice stops short of it.

Four decisions cover the whole zone.

If the somewhat disappointed group is bigger than the very disappointed group, narrow the promise. The product is close. The sentence explaining it is aimed at everybody.

If the very disappointed group all share one attribute, that is your market. Sell to it and stop apologising for how small it sounds.

If the very disappointed group is scattered across segments, you have three products sharing one login. Pick one for the next quarter and leave the other two alone.

If every churned customer said price, check whether they used the core action twice. Price is the polite reason. Non-use is the real one.

Now the two moves worth making this week, with the numbers attached.

Raise the price 20% for new signups only and hold it for 14 days. Revenue comes out even if conversion falls by less than 17%, because 1 divided by 1.2 is 0.83. Past a 17% drop you have lost money and learned something more useful than the price test: at this price, this promise does not hold. The structure you move toward is a separate question, covered in how to choose a pricing model.

Cut the time to first value in the same fortnight. Timestamp the signup, timestamp the core action, subtract. A median over ten minutes is a fit problem you can fix without touching a roadmap, and every step between the two timestamps is a candidate.

There is a third lever that looks like a distraction from growth work. Do the unfashionable manual thing to get the first hundred users to their first real result, the call and the import done by hand (Paul Graham, July 2013). The conversations in that work are where the narrower promise comes from.

The rest of the growth reading sits on the growth page, and the handover to a full plan is in a go-to-market strategy, for the month the number clears and the question changes to how fast to spend.

The bill arrives in your cost per customer

Fit arrives on the invoice for attention.

When the retention tail is flat, paid traffic pays for itself and $500 buys customers who stay. When it is not, you rent attention and keep none of it, and the cost per customer climbs every month while the product looks unchanged.

Two numbers make that visible on one line. What a visitor costs on the channel you are buying, and what CAC includes once you count the tooling and your own hours.

Ranked attention is priced in public now, which makes the comparison easy to run. The Board is ours, and every listing on it is a paid placement ranked by nothing but the amount paid, so a rank buys impressions and never a conversion. A dollar starts a listing.

What a listing sends is a fit question, and the click count printed beside each one is where the answer shows up. A rank can be bought. The response to it cannot.

The bill that follows a market you do not fit arrives as the cost of buying the same stranger twice, then once a quarter for a year, with each purchase filed under growth.

The number, on one sheet

The sheet is five rows and twelve columns, one column a month. Fill the first column tonight and you have a baseline, which is more than most companies have ever had.

Set the review for the same day each month and read the trend rather than the level. Two declines in a row on one reading is the signal to stop buying acquisition and go back to the segment work.

The survey answers lag the curve by about a quarter, which is why the curve gets the deciding vote on whether to spend.

Here is the failure at this end of the work, and it is common. A founder reads 42%, tells the team fit is solved, and stops measuring. The market shifts, the promise drifts, the score slides to 31% over two quarters, and nobody notices because nobody opened the sheet.

Fit is a ratio you keep, and it goes stale in the quarter your market moves.

Send the email first. One question, four answers, a hundred users, and a segment chosen before you believe what comes back.

Then build the cohort table, because the curve is the only part of this that cannot be talked into a better answer.

The number is not the point. What you do on the Monday after you read it is the point, and now you know which three columns to open.