The honesty problem in market research

Ask a person something that costs them a little to answer honestly, and you have started a negotiation. How much they tell you depends less on the question than on who is listening and what candor might cost them. Every research method is a different version of that bargain. A survey trades depth for reach. A focus group trades candor for nuance. Whichever one you pick quietly decides how much of the truth ever reaches you.

For decades the choice came down to two unhappy options. You could reach thousands of people through a survey and accept that you were collecting ticked boxes rather than reasons. Or you could sit a handful of people in a room, get real depth, and pay a price that kept the sample tiny. Breadth or depth, and rarely both at once.

Underneath that trade sits a second one that matters more, because it decides whether the answer is even true. The more personal the question, the more people shade their answer toward what looks acceptable. They do it differently in each setting, and the research on how they do it is more settled than most buyers assume. It is worth walking through, because it points somewhere useful.

Start with the survey, where the failure is attention. When researchers hide a small instruction inside a question to see who is actually reading, a striking share of people miss it. The study that introduced the technique found more than thirty percent^* failed, and later work has put the rate at a third to nearly half^* of all respondents. People are not so much answering the questions as moving through them.

It gets worse the longer someone answers for money. Professional panelists learn to grab the first plausible option and move on. Straightlining like this runs as high as forty percent^*, and climbs the longer they stay on the panel.

The defenses meant to catch it barely hold. In one analysis of more than sixty thousand respondents, three quarters^* of the people later flagged as bogus had passed both the attention check and the speed trap built to remove them.

And even a diligent respondent is boxed in by the format. A closed question can only hand back one of the options the researcher already thought to list. Anything outside that list, including the reason behind the choice, never appears.

The focus group solves the depth problem and introduces a worse one. Put a person in front of another person, or in front of a group, and they begin managing how they look. This is measured behavior, documented for decades.

When researchers checked self-reported grades against real records, admissions of a low GPA rose more than threefold^* once an anonymous form replaced the human interviewer. Men reported sensitive behavior three times as often^* to a private computer as to a person, in a study published in Science. Remove the audience and honesty climbs, most of all on the questions people most want to keep to themselves.

A group sharpens it further. In the classic conformity experiments, people asked to judge which of two lines was longer were wrong less than one percent of the time on their own. Surrounded by a few others all giving the same wrong answer, about a third^* went along with it.

That is roughly the size of a focus group, and a focus group is built for exactly that kind of agreement. Small wonder that what people say they will do in the room drifts well away from what they actually do later, by a third to several times over^*.

Turn the same lens on a conversational voice agent, and a second body of research answers the obvious question: does removing the human keep the honesty without giving up the depth? A 2025 review of twenty-six studies on what people disclose to conversational AI found, again and again, that people disclosed more to the machine^* than to a human, a survey, or a form, with the gap widest on the sensitive questions. In direct tests, the same person handed a human interviewer the polished answer and handed the agent the truer one.

The mechanism the research keeps returning to is anonymity. People say more when they believe the listener cannot judge them, and a machine reads as a listener with nothing to judge them for. When participants were told they were speaking with an AI rather than a person, their fear of being judged fell and their disclosure rose. Two further findings explain why this works best as a voice agent. People reveal more when they speak than when they type, and a faceless agent draws out more detail than one wearing a human-looking face, which runs against the instinct to make the thing friendlier.

The honest part of the claim is where it stops. On easy, low-stakes questions, people answer a survey, a person, and an agent about the same, and the literature says so plainly. The agent pulls ahead only once a question starts to trigger the urge to manage one's image. That is a narrow claim, and it is the right one, because that narrow band is exactly where the commercial value lives. The easy answer was never the hard part to collect.

The left column is the survey and focus-group research. The right is the self-disclosure research. The two were built independently of each other, and of us.

The reason depth has lived at small sample sizes is plain: a person had to run each interview or moderate each room, and that person was the cost. The figures below are per participant, in 2026 dollars, and loaded the same way on both sides, collection and analysis together, so the comparison is fair.

A Nexa conversation runs about $1.32 all in, in a range of roughly $1.19 to $1.45. Most of that is the voice itself; the analysis on top of it is a small fraction. Reading down the column, the gap is not a discount. It is a different order of magnitude.

Put the two halves together and the shape is clear. A survey is anonymous but shallow. A focus group is deep but exposed. A voice agent is the first method that is anonymous and deep at the same time, which is why it holds onto honesty exactly where the older methods lose it.

None of this rests on a claim we made about ourselves. It rests on two separate bodies of peer-reviewed work, built by people with no stake in the outcome, that happen to meet at one point: the sensitive, considered questions where understanding people is worth the most, and where the truth has always been hardest to get.