The product-market fit survey: how to run the Sean Ellis test

What the product-market fit survey measures

The survey measures how much your current users would miss your product, as the share who say they would be very disappointed without it. It comes from Sean Ellis, who led marketing at LogMeIn from launch to its IPO filing and later worked with Dropbox and Eventbrite. In a 2009 interview with Venture Hacks, he said he asks existing users how they would feel if they could no longer use the product, and that fit needs at least 40% of them to say “very disappointed”. He set the line after comparing results across nearly 50 startups; later accounts, including Superhuman’s, say nearly 100. He also called the threshold somewhat arbitrary: startups that struggled for traction came in under it, and most that grew strongly came in over it.

It asks about loss rather than satisfaction, which makes it harder to answer out of politeness. That is the difference from the metrics it is often confused with:

Andreessen defined product-market fit as being in a good market with a product that can satisfy that market. The survey is one way to see whether you are there yet. It works best next to retention, not instead of it.

The questions to ask

Ask four questions, in this order: the Sean Ellis question, who would benefit most, the main benefit, and how to improve. These are the four Rahul Vohra describes in Superhuman’s account of its product-market fit engine. Two optional questions follow for when you want more.

1. 1

   “How would you feel if you could no longer use [product]?”

   Offer four answers: very disappointed, somewhat disappointed, not disappointed, and I no longer use it. Keep the wording exactly the same every time, and ask it first, before any question that could color it.

   Multiple choice

2. 2

   “What type of people do you think would most benefit from [product]?”

   Users describe your best customer better than your own persona documents do. Superhuman used these answers to picture the person to build for.

   Short text

3. 3

   “What is the main benefit you receive from [product]?”

   This is the reason people stay. Spoken, with one follow-up asking for a recent example, it turns “speed” into the task that got faster and what it replaced.

   Best spoken

4. 4

   “How can we improve [product] for you?”

   Read these answers by group. From very disappointed users they show what to protect. From somewhat disappointed users they show what is holding them back.

   Best spoken

5. 5

   “What would you use instead if [product] were no longer available?”

   Optional. It names your real competition, which is often a spreadsheet, a habit or nothing.

   Best spoken

6. 6

   “Have you recommended [product] to anyone?”

   Optional. A yes is a behavior, not a promise, and it tells you whether word of mouth is starting.

   Multiple choice

How to run the survey in Feedbaq

Feedbaq is built for surveys like this one, where both the numbers and the reasons behind them are critical to your success. You send one link to your active users, and they answer in their own time.

Why it works better than a traditional survey

• The reasons come out spoken. The Sean Ellis question stays multiple choice, so you can count it. The questions about the main benefit and what to improve are answered by voice, and spoken answers are about three times longer than typed ones: 44 words against 14 in a 2024 study with 1,001 people.
• An AI asks the follow-up. When someone picks “very disappointed”, it asks what exactly they would miss, guided by a line you write.
• People can show you. Add a screen share, and the people who would miss the product walk you through how they actually use it.
• The score and the reasons in one place. Click “very disappointed” in the report to read only those people’s answers. Every spoken answer is transcribed and labelled.
• Your AI agent can do the work. Connect Claude, ChatGPT or another agent, and it can build the survey and read every answer, split by segment.

Set it up in four steps

1. Start from the template. The product-market fit template asks the Sean Ellis question and the three questions that follow it, in order, with a follow-up on each.
2. Add what you need to segment by, such as a multiple-choice question about role or plan, if you do not already know it.
3. Send the link to people who have used the core of the product recently. They need no account.
4. Read the answers. Filter by “very disappointed” and by segment, or ask your agent to calculate the score for each group and quote people.

Building the survey is free and you can use our product-market fit template to start collecting feedback today.

Who to ask, and how many

Who to send it to

Send it to people who have used the core of your product recently enough to judge it. Superhuman, citing Ellis, used at least twice in the last two weeks. Match the window to how often your product is normally used: a monthly reporting tool needs a longer one than an email client. The rule keeps out the people who cannot judge yet and would drag the score down for the wrong reason.

• Leave out signups who never reached the core of the product, people in their first week, and your own team, friends and investors.
• In B2B, ask the people who use it, not only the person who bought it. Tag each answer with the account, role, plan and company size from your own data, so you can segment later without asking.
• Ask each person once, and survey new qualifying users as they arrive. Asking the same people again and again teaches them to answer quickly.

How many answers you need

Aim for at least 40 answers to see a direction, and 100 or more for a stronger basis for decisions. Vohra writes that results start to become directionally correct at around 40 respondents. The reason for caution is the margin of error. Jim Lewis and Jeff Sauro at MeasuringU point out that a 40% score from 50 answers has a 95% margin of error of about ±13 points: the true share could be anywhere from about 27% to 53%.

So a score of 38% from 40 answers is not a miss, and 43% is not a pass. If you have fewer than about 30 qualifying users, which is normal for an early B2B product, the percentage will not settle. First Round’s levels of product-market fit describe the earliest stage as three to five customers. At that size, ask the same questions in conversation and read every answer, rather than computing a score.

How to calculate your score

Divide the number of “very disappointed” answers by the number of people who answered very, somewhat or not disappointed. If your survey offers “I no longer use it”, leave those answers out of the total by default: they tell you about churn, not about fit among current users. Whichever you choose, choose it once and keep it, so scores stay comparable over time.

Enter your own numbers below. The example is 44 answers with a score just under 40%, and the range shows why that result is too close to call.

Segment before you decide

Your overall score averages people the product was built for with people it was not, so score each segment on its own before you decide anything. This is the heart of Superhuman’s method. Vohra reports that only 22% of users first said they would be very disappointed. Segmenting down to the kinds of people who loved the product most (founders, managers, executives and business development) raised the score by about ten points. Building for them raised it further: within three quarters, the score reached 58%.

1. Score each segment separately: role, company size, use case, plan or how people found you.
2. Find the segment with the highest share of very disappointed users, and describe that person as specifically as you can.
3. From that group’s answers to question 3, find the main benefit in their words.
4. Among the somewhat disappointed, keep the people who name that same benefit. Their answers to question 4 are the gaps to close.
5. For now, set aside everyone else, politely. They may be a market for later, but not the one to build for now.

Superhuman then split its roadmap in two. Half went to what the very disappointed users loved: speed, shortcuts, automation and design. The other half went to what held back the somewhat disappointed users who also valued speed, such as mobile, integrations, calendaring and search. Try the same split on invented data:

In this example the whole group scores 36%, founders alone score 64%, and engineers score 15%. The founder segment is the one to build for, but 14 answers is a small sample. Its range is wide, which is a reason to collect more answers from founders, not to celebrate.

Read the reasons, not only the number

The score tells you whether; the open answers tell you why and what to do next. In 2015 Hiten Shah and a group of volunteers ran the survey independently on Slack’s users. Of 731 people, 51% said they would be very disappointed without Slack. The more useful findings were in the words: the main benefit people named was less email, the people who depended on Slack used it as their main internal channel, and every group asked for video calls.

To read your own answers:

• Group the answers to questions 3 and 4 into themes, separately for the very, somewhat and not disappointed.
• Count the themes per group. The benefit that very disappointed users name most often is what your product is for.
• Keep the exact words. How your best users describe the benefit is often better copy for your homepage and onboarding than anything you would write.
• Use AI to group and summarize, then read a handful of answers in full to check that the themes are right.

This is where spoken answers help. A typed answer to “what is the main benefit?” is often one word. Spoken, with a follow-up asking when it last mattered, the same question gives you a story you can use:

What to do with your score

Treat the score as a guide to your next move, not a grade. The bands below are common rules of thumb rather than research findings, and at small samples your score can move between them by chance.

What the test cannot tell you

The survey is a useful signal, not proof. Lewis and Sauro found little research evidence behind the item itself and note that the 40% line rests on its originator’s intuition. Surveys also carry the usual biases: questions asked before it can shift the answers, and the people who reply tend to be your more engaged users.

A high score can also sit next to weak retention. Sean Ellis wrote in December 2024 about a startup where 55% of surveyed users said they would be very disappointed, yet retention was poor. That is why Lenny Rachitsky lists the survey as one signal among several, alongside retention curves that flatten, organic growth and efficient growth. Read them together.

Run it this week

1. Pull the list. Everyone who used the core of the product recently, such as twice in the last two weeks, tagged with role, plan and company size.
2. Send the four questions. The product-market fit template has them in order, with spoken answers and follow-ups on the open questions.
3. Wait for about 40 answers, then calculate the score and its range.
4. Segment, read the reasons by group, and write down the one segment to build for and the two or three gaps to close.
5. Set a date to measure again, a month out, with new qualifying users.

Questions people ask

Sources

1. Sean Ellis, interviewed by Venture Hacks 19 November 2009. The 40% benchmark and the “very disappointed” question.
2. Rahul Vohra, “How Superhuman built an engine to find product-market fit” First Round Review, 2018.
3. Marc Andreessen, “The only thing that matters” 25 June 2007, mirrored by Stanford.
4. Jim Lewis and Jeff Sauro, “Is the product-market fit item a good measure?” MeasuringU, 2022.
5. Hiten Shah, “Slack product-market fit survey” An independent survey of 731 Slack users, 2015.
6. Lenny Rachitsky, “How to know if you’ve got product-market fit” Lenny’s Newsletter, 2020.
7. Sean Ellis, “High PMF score but low retention” December 2024.
8. First Round Capital, “Levels of product-market fit” 2024.
9. Kromatic, “Survey: product-market fit” On order effects and survey bias.
10. Lenzner, Höhne and Gavras, spoken versus typed answers to open-ended probes Journal of Survey Statistics and Methodology, 2024.