You ship a feature you have spent three weeks on. You message ten of your most engaged users and ask what they think. Nine reply within a day, all of them positive. One sends a thumbs-up. You go to bed feeling good about the next quarter. Two weeks later the usage chart is flat, and the only person who actually opened the new feature is the one who left a thumbs-up.
This is the moment most founders quietly conclude that “talking to users doesn’t work.” It does work. What does not work is the kind of conversation you just had — a conversation in which the truth was never going to surface, because nothing in the way it was framed gave the user permission to say it.
This piece is for the bootstrapped or tiny-team B2B SaaS founder, post-launch, with somewhere between fifty and a thousand users. The question it tries to answer is the one you have probably been asking yourself in private: why users don’t tell you what they really think, even when they like you, even when they want your product to succeed, even when you ask nicely. The reason your interviews keep producing polite encouragement and roadmap fog is not that your users are dishonest. It is that human beings — including you, and including the engineer you onboarded last Tuesday — are running an entirely different conversation than the one you think you are having.
People aren’t lying. They’re being polite.
The word “lying” sets up the wrong frame. Almost no participant sits across from you intending to deceive. They are doing what humans do in every other social situation: managing how they come across, sparing your feelings, and reaching for the most agreeable answer that does not feel obviously wrong.
This tendency has been documented and measured for more than sixty years. In 1960, the psychologists Douglas Crowne and David Marlowe published a scale for measuring social desirability, the basis for what later became known as the Marlowe-Crowne Social Desirability Scale. The pattern it captures is mundane and universal: people inflate the things they believe society approves of and minimise the things they suspect it does not.
Nielsen Norman Group calls this the close cousin of the Hawthorne effect — the well-documented finding that people behave differently when they know they are being observed. In a research interview, the participant is not just being observed. They are also being asked to be useful, to not waste your time, and to not embarrass themselves. The cheapest way to satisfy all three is to be agreeable.
The implication for a founder is uncomfortable. Most of the things you set up to “talk to users” — the fifteen-minute Zoom, the in-app survey, the reply-to-this-email pulse check — are nearly perfect environments for producing polite answers. The participant is on display, the goal is unstated, and there is a clear interpretation of “good behaviour.” You are not collecting honest feedback. You are collecting performance.
Three answers that look like data and aren’t
Once you start watching for them, the same three answers will turn up in every interview you run. Rob Fitzpatrick’s The Mom Test is built around naming and disarming them. He calls them compliments, fluff, and hypotheticals.
Compliments cost the participant nothing. “That’s a great idea” is a polite filler that buys them ten seconds of social warmth without committing them to anything. Fitzpatrick treats compliments as bad data because they validate the founder without proving behaviour. When you hear one, the data point is not that the user likes the thing. The data point is that you have been validated, and validation is not evidence.
Fluff is the layer below that. It is a generality dressed up as a fact. “I usually check this once a day.” “I never use that other tool.” “I’d probably switch if it had X.” None of those sentences contain a real event. They are summaries the participant has assembled on the fly because they sound about right, and they are mostly invented in the moment.
Hypotheticals are fluff with a future tense. “Yes, I’d pay for that.” “Yes, I would invite my team.” Teresa Torres makes the same point in her willingness-to-pay guidance: answers to “what would you do?” questions are unreliable because the participant is not making a real choice. The classic apocryphal version is the line attributed to Henry Ford about customers asking for “a faster horse” — a quote Harvard Business Review’s fact-check finds no evidence Ford ever said. The myth survives because the underlying observation is true: when you ask people what they would do, they reach for the future they imagine themselves living in, and that future is almost always more rational, more virtuous, and more disciplined than the one they are actually living in.
These three answers — compliments, generalities, and predictions — are what most founders bring back from their first ten interviews. They feel like signal. They are noise.
Why surveys hide more than they reveal
The temptation, once interviews start producing fluff, is to retreat to surveys. Surveys feel rigorous. They scale. They produce numbers, and numbers can be dropped into a slide deck.
Erika Hall, in her book Just Enough Research and in her Awkward Silences podcast appearance, argues that surveys become dangerous when teams use them as an easy substitute for listening or observation. The danger is not the format. The danger is that the format hides its own failure modes. With an interview, you can tell when something has gone wrong — when the participant is the wrong person, when they have misread the question, when they are not engaged. With a survey, you have no such signal. The bad data looks identical to the good data.
Two response biases compound the problem. The first is acquiescence bias — the well-documented tendency for respondents to agree with whatever statement they are shown, especially when the question is binary, the wording is positive, and the participant is mildly disengaged. The second is the response-rate problem. A 5% completion rate is not a small sample of your users. It is a complete sample of the small group of users who are sufficiently energised, on a given Tuesday morning, to fill in your survey. Those users are systematically different from everyone else, usually in the direction of being more enthusiastic and more articulate. The “silent 95%” — the people who churned, the ones who never activated, the ones who use the product weekly without ever feeling much about it — are precisely the cohort whose feedback would matter most, and they are the ones the survey misses.
This is also the deep flaw with metrics like NPS used as product signal. The number is a summary of a summary, weighted toward the loudest fraction of the loudest decile, asked in a moment when the user has no specific event in front of them to anchor on. It can tell you whether a brand is broadly liked. It cannot tell you why a particular person stopped using a particular feature on a particular day, which is the only kind of information that changes a roadmap.
Why your involvement in the room makes it worse
Even if you avoid the survey trap and run interviews like you mean it, there is one variable you cannot easily change about your interviews: you are in them. And the founder is the worst possible person to ask their own users to be honest with.
Erika Hall’s broader point is useful here: people are unreliable narrators of their own behaviour, and the people most prone to self-serving interpretation about a particular product are the people who have spent the last eighteen months building it. Confirmation bias does not announce itself. It does its work quietly, in how you frame the question, in which follow-ups you bother to ask, in which sentences you write down and which you let pass. You are not consciously filtering the conversation. You are just, like everyone else, more attentive to evidence that supports the path you are already on.
The participant feels this. They can usually tell quickly that you are the founder. Once they know, the social contract changes. They are no longer giving feedback to a researcher; they are giving feedback to the person whose work it is. The cost of being honest goes up. The cost of being kind goes down. The mental shortcut almost everyone takes — including the most experienced researchers, when they are interviewing users of a product they helped build — is to soften.
Jakob Nielsen made this point uncomfortably plainly in 2001 in his article “First Rule of Usability? Don’t Listen to Users”. His argument is not that users have nothing useful to say. His argument is that what they say in the room is a poor predictor of what they will do outside it, and that the role of a researcher is not to transcribe stated preferences but to observe behaviour. NN/G’s own follow-up piece, “Why User Interviews Fail”, warns that interviews are a poor fit for questions about future behaviour or detailed behavioural recall: people forget details, speculate, and rationalise.
The implication is not “stop interviewing.” It is “stop being the interviewer of record on your own product.” When the maker is in the room, the participant is not talking to a researcher. They are talking to the maker. That is a different conversation, and it has a different ceiling on honesty.
How to hear what users really think
The good news is that candour is not a personality trait of the participant. It is a property of the conversation. When the conversation is set up well, even ordinary, polite, time-pressed people will tell you things they would never write in a survey. Four moves do most of the work.
The first is to anchor every question in a specific past event. This is the central technique behind Teresa Torres’s story-based customer interviews. Instead of “how do you usually plan your week?” — a question that almost guarantees a generality — you ask “walk me through the last time you sat down to plan a week.” Torres uses the prompt “tell me about the last time you watched Netflix” as her teaching example. The participant cannot answer that question with a hypothesis. They have to remember an actual evening, an actual device, an actual interruption. The answer becomes a story instead of a position, and stories contain the texture — the friction, the workaround, the moment of doubt — that positions strip out.
The second is silence. Steve Portigal’s Interviewing Users is, more than anything else, a book about how to stop talking. Most novice interviewers, faced with a pause, fill it. They reframe the question, supply a possible answer, or move on. Portigal’s argument is that the interesting answer is almost always the second one — the one that surfaces after the participant has finished the polite version and started reaching for the real one. The interviewer’s job, in those few seconds of silence, is not to rescue them. It is to wait.
The third is to reframe what the conversation is for. Indi Young calls these “listening sessions” rather than interviews. The relabel matters. An interview is something you do to extract data; a listening session is something you do to understand how someone thinks. The participant senses the difference. So does the interviewer. You ask different questions when the goal is to understand a person’s reasoning rather than to verify your own.
The fourth is to remove the loaded relationship from the room. The maker should not be the interviewer. The clearest experimental evidence for this comes from a 2014 study by Gale Lucas and colleagues at USC’s Institute for Creative Technologies. They had participants interact with a virtual human interviewer, but told one group the agent was fully automated and another that it was puppeted by a human in the next room. In reality everyone was talking to the same system. Participants who believed they were speaking with a computer reported less fear around disclosure, managed impressions less, and shared more sensitive material. The same pattern shows up in the SimSensei kiosk, DARPA-funded research into virtual interviewers for psychological-distress assessment. People disclose more readily when they believe no human is judging them on the other side.
The deeper finding, for our purposes, is not “use a robot.” It is that the felt absence of judgement is the active ingredient, not the species of the interviewer. Human researchers can replicate part of that condition through training and discipline; tools can replicate another part by removing the maker from the room. The point is to design the conversation so that the easiest thing for the participant to do is also the most honest.
What users will tell you, when the conversation is built for it
The first time most founders run an interview that works, the experience is disorienting. The participant says something that flatly contradicts a thing the founder believed yesterday. There is a small silence. The founder, on instinct, wants to clarify, to argue, to point out that the user has misunderstood. Resisting that impulse — sitting with the contradiction, asking “tell me more about that” instead — is roughly the entire skill of qualitative research.
The goal is not to catch users in a lie. It is not to “extract the truth,” as if the truth were something the participant was withholding on purpose. The goal is to design a conversation in which the truth becomes the easy thing to say. That is a craft. It involves picking the right participants, asking past-tense questions, leaving silences alone, framing the conversation as listening rather than extraction, and — when the maker is too close to the work to hear the answer cleanly — getting the maker out of the room.
This is also the problem Maren is built to address. We talk to your users so you do not have to, with no investment in any particular answer and no social cost to the participant for being honest. But the principle is older than the tool. Even without us, the founder who internalises one thing from the canon of user research is ahead of most: people will tell you what they really think when the conversation is designed for them to. Most of the time, it is not.