Is market research dying? Not of natural causes, and nobody pushed it off a cliff. You killed it slowly, one reasonable compromise at a time, and I stood next to you and helped.
By “you” I mean everyone who makes a living from asking people questions: the researcher at the agency, the insights manager at the brand, the procurement team that buys sample by the kilo. I have been all three kinds of guilty. The trade still has its name, its conferences and its awards. What it is losing is the thing it was paid for, which is telling a company something true about people that it didn’t already know.
The murder weapon isn’t AI, and it isn’t shrinking budgets. It’s a long list of choices, each of which looked sensible on a Tuesday afternoon.
Is market research dying?
The appetite for understanding customers has never been bigger. Companies hold more data about people than at any point in history. The craft of asking people good questions and listening to the answers is in far worse shape than the spending suggests.
At Pew Research Center, the response rate of a typical telephone survey was 36% in 1997 and 9% in 2012. By 2018 it had fallen to 6%. Pew blames robocalls and spam filters as well as the old complaints about time and privacy, and it moved most of its US polling to an online panel. Commercial research moved online much earlier, and for a simpler reason: it was cheaper.
Online didn’t solve the problem of people not wanting to talk to us; it hid it. An opt-in panel has no response rate to embarrass anyone. It has a supply chain, and the supply chain will deliver 1,000 completes by Friday whether or not 1,000 honest people want to fill in your survey. Most of what follows is about how we learned to stop asking where the completes came from.
Saying yes to every brief
The first bad choice comes before a single question is written. A client wants a segmentation of the US market in two weeks, for roughly the price of the board’s Christmas dinner. A good researcher knows it can’t be done properly. Most of us say yes anyway, because the agency down the street will.
The industry’s own rulebook says otherwise. The 2025 ICC/ESOMAR International Code tells researchers to design research that is fit for purpose, and if it isn’t, “the client must be informed and the issues resolved.” I have read very few proposals that say “this can’t be done well at this price, and here is what you’ll get instead.” I have written proposals that should have said it.
Clients deserve their share. Procurement compares research quotes the way it compares printer toner, by cost per unit. Cost per complete is a fine way to buy toner. It is a terrible way to buy the truth about why your customers leave, because the cheapest complete is the one where nobody checked who completed it.
Why are surveys so long?
Then comes the questionnaire, which starts at 10 minutes. Brand wants a few image statements, the regional team wants its own module, finance wants price sensitivity, and legal adds a page. Nobody owns the respondent’s time, so everyone spends it, and the 25-minute survey assembles itself one reasonable request at a time.
We know what this does from a 2009 experiment in Public Opinion Quarterly, in which Mirta Galesic and Michael Bosnjak told people a web survey would take 10, 20 or 30 minutes. The longer the announced length, the fewer people started and the fewer finished. Answers to questions placed later in the questionnaire were faster, shorter and more uniform than answers near the start. Your lovingly worded question 47 is being answered by someone who stopped reading at question 30.
This is what treating respondents as lab rats looks like in practice. Nobody intends cruelty; we simply assume their attention is free and unlimited. People respond the way anyone responds to being mined: they speed, they click the same column all the way down, or they leave.
The ones who stay are increasingly the ones who are there for the points. Getting someone to the last page and getting them to think are two different achievements (participation is not engagement), and we only measure the first.
The cheapest sample is the most expensive data
Online sample is sold by the complete, and in my experience the only question anyone asks about the price is whether it can go lower. In 2016 Pew Research Center gave the same 56-question survey to nine opt-in samples from eight vendors and found that their accuracy varied widely. The two least accurate samples ranked very highly on how well their unweighted demographics matched the population. A sample can look exactly like the census on age and gender and still be wrong about the things you wanted to know.
None of this was secret. In 2010 the American Association for Public Opinion Research told researchers to avoid nonprobability online panels when one of their goals is to estimate population values accurately. The industry read the report and kept buying the cheapest completes, because the client wanted a number, not a lecture on sampling theory.
The bill arrived in April 2025, when federal prosecutors charged eight people connected to the panel companies Op4G and Slice MR in a conspiracy to bill $10 million in fraudulent survey data. By October 2026, four of them had pleaded guilty. I go through that case, and through the young adults who claim to be licensed to operate a nuclear submarine, in Movin’ on up. The short version is that our samples are full of submariners in space, and a market that pays for completes without asking who completed them will eventually be sold completes that nobody completed.
Are synthetic respondents market research?
The newest way to kill the trade is to stop asking people at all. “Synthetic respondents” are language models told to play a 35-year-old suburban mother or a college gamer and fill in your questionnaire. They never speed, never get bored and cost almost nothing. They are on sale now, and in 2024 the UK’s Market Research Society published a report on using them.
The case for them deserves a hearing. A model trained on oceans of human text can be a cheap way to pressure-test a questionnaire, generate hypotheses, or spot an obviously doomed concept before you pay for fieldwork. Used that way, I have no quarrel with it.
The trouble starts when the output is presented as a measurement of what people think. Five political scientists tested exactly that in Political Analysis in 2024 (Bisbee and colleagues). Asked to answer as personas, a commercial language model produced average feelings toward social and political groups that looked close to the American National Election Studies. Underneath, the answers varied far less than real people’s did, 48% of the regression coefficients differed significantly from the human data, and the same prompts gave different results in April, June and July 2023.
The Market Research Society’s report warned of “normative results which can flatten diverse groups.” Flattening diverse groups is a fair summary of what bad research has always done. Now it can be done without fieldwork.
Worse, in Sean Westwood’s 2025 study in PNAS, AI agents taking surveys could infer the researcher’s hypothesis and produce data that confirmed it, even with no malicious intent. Research that tells clients what they hoped to hear is nothing new. It used to require a weak researcher, and now it can be ordered by the thousand.
The 2025 ICC/ESOMAR Code now requires researchers to tell clients when synthetic data or synthetic personas are used, and to state how much human oversight was involved. Disclosure is the minimum. Calling the output research is a separate decision, and I think it’s the wrong one.
Dashboards instead of decisions
Suppose the data is real. The next way to waste it is to deliver it as a dashboard. Dashboards are wonderful at showing that a number moved. They have nothing to say about why it moved or what anyone should do about it, and those were the questions the research was paid to answer.
The quarterly tracker is the purest form. It runs this quarter because it ran last quarter. Its main output is a slide saying brand awareness is stable within the margin of error, which it usually is, and which nobody acts on. I have presented that slide and watched a room nod at it before moving on to the next agenda item.
This is the gap between information and understanding. Information is “63% agree.” Understanding is knowing why the other 37% don’t, whether it matters, and what the company would do differently if it believed you. A 90-page deck contains a great deal of information, and often it means the researcher never decided what mattered and handed that job to a reader with 10 minutes.
Hiding in the herd makes it worse. We use the metric everyone uses, because nobody gets fired for it, and Net Promoter Score is the clearest case.
A 2007 study in the Journal of Marketing used data from 21 Norwegian firms and found that the score predicted revenue growth no better than other customer metrics. Companies kept adopting it, and I suspect a big reason was that their competitors had. A benchmark tells you where you stand among everyone measuring the same thing, which is comforting and rarely the question.
My take
I don’t think the trade is dead. I think it’s hollow, and hollow things can be filled again. That takes researchers who refuse to take part in the hollowing, and clients who are willing to hire them.
Here is what such a researcher looks like:
- She asks what decision the research will inform before she asks about the budget, and if nobody can name one, she suggests not doing the study.
- She turns down the timeline that can’t be met and puts the trade-off in writing.
- She cuts the questionnaire until it fits in 10 minutes and tells the regional team its module didn’t make it.
- She knows where her sample comes from, pays for verified people when the decision needs real estimates, and labels the result as directional when it doesn’t.
- She doesn’t call a language model’s guesses research.
- She delivers a one-page answer with a recommendation, then checks three months later whether anyone acted on it.
None of that needs new technology. It needs ordinary critical thinking: how do we know this, what would change our minds, and who benefits from the answer. It also needs a willingness to lose a client now and then.
I have failed most of these tests at some point. I suspect you have too. The good news about a trade killed by its own choices is that the same people can make different ones, starting with the next brief.
Questions people ask
Is market research dying?
Demand for understanding customers is growing. What is dying is trust in cheap, fast research: response rates have collapsed, opt-in samples carry a worrying share of fake respondents, and too many reports change nothing.
Can synthetic respondents replace real survey respondents?
Not for measurement. A 2024 peer-reviewed test found that language-model personas matched average opinions but showed far less variation, led to different statistical conclusions and changed over time. They can help draft and pretest questions.
How long should an online survey be?
As short as the decision allows. Experiments show that surveys announced as longer get fewer starts and completes, and that later answers get faster, shorter and more uniform. My rule of thumb is 10 minutes or less.