The future of market research — AI meets tradition, and the middle disappears

Editorial Series | Market Research | We Tell More Than We Know

July 19, 2026 | Niche Industry Deep Dive Summer Editorial Series | Focus: "The future of market research — AI meets tradition, and the middle disappears"

The survey said the space was too loud. I sat in it for three hours on a Thursday and watched people stay longer because it was loud.

Not one of them would have told you that. Asked directly, they would have said what the respondents said — that it was noisy, that they would prefer it quieter, that noise is a negative — because that is the socially correct answer and because they had no conscious access to what the volume was actually doing for them. What it was doing was giving them permission. The noise meant no one could hear their table. The noise meant a party. The noise meant a woman could laugh at full volume without the room turning. They did not want quiet. They wanted cover, and they did not know it, and no questionnaire on earth was going to extract it from them.

I have built a career on the gap between what people say and what people do, and this month I have been thinking about it constantly, because that gap is about to become either the most valuable thing in my profession or the thing my profession forgets entirely. AI can now produce a thousand articulate respondents in nine seconds. They will answer any question you ask. They will be plausible, consistent, fast, and cheap.

And they will tell you the space is too loud.

What actually happened to research

Let me describe the state of things plainly, because it has moved faster than most founders realize.

Synthetic respondents — AI personas generated to stand in for real audiences, sometimes called digital twins or silicon samples — went from curiosity to standard line item in about two years. By the middle of this year the adoption pattern was visible and unembarrassed: agencies run synthetic panels to win pitches and populate client workshops; in-house teams use them for the first eighty percent of a project and validate the last twenty with real humans; B2B teams use them to "reach" audiences they could never realistically field, like CIOs or regulated buyers. There is genuine academic work underneath it — simulated samples in political science, homo silicus in economics, replication studies that found language models could reproduce the findings of several classic experiments. Where personas are calibrated on real prior data from the same audience and the question rewards general reasoning, published validation puts correlation with real respondents somewhere in the eighty to ninety-five percent range on directional questions.

That is a real capability and I am not going to pretend otherwise. Used narrowly, it is excellent: stress-testing a discussion guide, dry-running a screener, pre-testing twenty concepts so you take the strongest two into live testing instead of paying to learn that eighteen were bad.

But the industry itself is not comfortable, and the discomfort is worth listening to. One survey of researchers found more than four in ten explicitly not excited about synthetic respondents. And then Pew Research Center — an institution whose entire asset is credibility about what people think — put its position on the record this year in a sentence with no wiggle room in it: they only interview real people, and they do not use AI to tell them what the public thinks. Their internal testing had found something specific and damning. AI-generated responses stereotyped demographic groups and understated real disagreement.

Sit with that second finding, because it is the whole essay.

The machine flattens the very thing you were looking for

A synthetic respondent is not a person. It is a distribution. It has no purchase history, no bad Tuesday, no unresolved thing about its mother, no dissatisfaction — it has training data, which is a compressed average of what people have already publicly said. And an average, by construction, has had its edges removed.

Which means synthetic research is structurally biased toward consensus and structurally blind to deviation. It will reliably tell you what the middle of the distribution thinks. It will smooth over the fifteen percent who feel violently otherwise, because in the training data that fifteen percent is noise to be minimized rather than signal to be found. Pew's finding — that it understates real disagreement — is not a bug in a particular model. It is what averaging is.

Now recall why market research exists. Nobody commissions research to confirm the obvious. You do it to find the thing you did not already believe — the unmet need, the strange behaviour, the segment nobody is serving, the reason the product is succeeding for a reason you did not design. Growth lives in the anomaly. Every category-creating insight in commercial history was, at the moment of discovery, a deviation from the consensus: a minority behaving strangely, a use case nobody intended, an emotion nobody had named. And synthetic respondents are, by their nature, machines for removing exactly that.

I wrote earlier this season about model collapse — feed a generative system its own output and it drifts toward the bland centre, losing the rare and the strange with each generation. Synthetic research is that phenomenon applied to your understanding of your own customer. You are not learning about the market. You are learning about the average of what has already been said about the market, which is a description of the past with the interesting parts removed. It is a superb tool for confirming what you know. It is structurally incapable of surprising you, and surprise is the entire product.

The older problem the machine inherits

Here is the part that should worry researchers most, and it long predates AI.

In 1977 Richard Nisbett and Timothy Wilson published one of the most quietly devastating papers in psychology, under a title that has never been improved on: people tell more than they can know. Asked why they did something, subjects produced confident, fluent, entirely plausible explanations — that were demonstrably not the actual causes of their behaviour. They were not lying. They had no privileged access to their own processes, so they confabulated: constructed a reasonable-sounding story after the fact and believed it completely.

Hold that against the line I used earlier this season from Michael Polanyi — we know more than we can tell. Put them together and you have the complete indictment of the interview as an instrument. We know more than we can tell, and we tell more than we know. The most important things about a customer are inaccessible to her conscious report, and the explanations she offers instead are confidently wrong. This is not cynicism about people. It is simply how minds work, and every good researcher has known it for fifty years.

Which brings us to the structural problem with synthetic respondents that no amount of model improvement will fix. They are trained on what people said. Stated preference. The least reliable data humans produce. A synthetic panel is not a simulation of human behaviour; it is a simulation of human self-report, which was already the weakest link in the chain. The machine has automated the part of research that was always broken — and it has automated it beautifully, at scale, with total fluency and no hesitation, which makes the wrongness far harder to see.

This is how New Coke happened, decades before anyone had a language model. The taste tests were extensive and the taste tests were right: in a sip, people preferred the sweeter formula. But nobody was buying a sip. They were buying an identity, a childhood, a piece of American furniture, and the research never asked because the research had already decided which question mattered. The data was clean. The question was wrong. A synthetic panel would have replicated the error at a thousand times the speed and a hundredth of the cost.

What tradition actually is

So when I say tradition, I do not mean the survey. The survey is not the tradition worth defending — it is the part most deserving of automation.

The tradition worth defending is observation. The oldest, least scalable, most consistently productive method in the discipline: going where people are and watching what they actually do. Ethnography. The researcher in the store, the parking lot, the kitchen, the dining room on a loud Thursday. The tradition that produced the famous milkshake insight — that a fast-food chain could not understand its milkshake sales until someone stood in the restaurant for eighteen hours and noticed people buying them at seven in the morning, alone, for a long boring commute, because the real job was not dessert but company. No survey would have surfaced that, because no customer would have said it. It had to be seen.

Observation solves both halves of the indictment. It bypasses what people cannot tell you, and it bypasses what they tell you wrongly. It puts you in contact with behaviour rather than narrative. And it produces the one thing a synthetic panel can never generate: surprise you did not prompt. The AI answers the question you asked. The Thursday afternoon answers the question you did not know to ask, which is where every genuinely valuable insight in my career has come from.

There is a second traditional skill, harder to name and harder to hire: reading the unsaid. The hesitation before an answer. The compliment delivered with a flat face. The contradiction between what someone says about price and what happens to her shoulders when she sees it. Qualitative researchers have always known that the transcript is the least informative part of an interview. A synthetic respondent has no body, no pause, no reluctance — it is all transcript, all the time, which is precisely the layer where humans are least honest.

The barbell: why the middle dies

Here is my actual prediction for this profession, and it follows from everything above.

The future of market research is a barbell. At one end, AI does the enormous, tireless, genuinely superhuman work: reading ten thousand open-ended responses and finding the pattern a human team would take six weeks to see; synthesizing the existing literature in an afternoon; generating fifty hypotheses to be tested rather than believed; pre-testing concepts so real money is spent only on survivors; collapsing synthesis time from weeks to hours. This is a gift and refusing it is not integrity, it is inefficiency.

At the other end, the human work becomes more valuable precisely because the first end became free: real contact with real people, observed rather than surveyed. Deep ethnography. Small samples, long exposure, actual presence. When answers become infinitely cheap, the scarce asset is not answers — it is encounter. Being in the room. Seeing the thing nobody reported.

And the middle collapses. The commodity survey panel. The generic focus group with a two-way mirror and a bowl of M&Ms. The undifferentiated quant report that tells you what you already suspected in a deck nobody reads. That entire tier — expensive, slow, and no better than what a model can now produce in seconds — is the part of this industry that is about to disappear. Not because AI is smarter than researchers, but because AI is exactly as good as mediocre research, and mediocre research was most of the market. The industry's own trend data already shows budget migrating away from quant tracking and toward qualitative depth, which is the barbell forming in real time.

The strategic instruction for a founder is therefore simple and uncomfortable: use the machine for volume, and buy yourself hours of real human contact you cannot automate. Ten conversations you conduct yourself, watching faces, are worth more than a thousand synthetic responses — and they cost less than you think, because the expensive part of research was never the talking. It was the infrastructure around it, and that infrastructure is what just got cheap.

Listening was never the soft part

There is a hierarchy inside research that I want to name, because it is about to invert.

Quantitative work has always carried the prestige — the numbers, the significance, the confidence interval, the vocabulary of rigor. Qualitative work carried the apology: just anecdotes, just a few conversations, soft data, insufficiently robust, the thing you did before the real study. And it will not surprise you that the qualitative side of this industry has always skewed toward women, or that the skills it requires — sustained attention, emotional reading, sensing what someone almost said, holding a room without steering it — are exactly the ones the culture files under "intuition" and declines to price.

Watch what just happened to that hierarchy. The quantitative layer — the surveying, the coding, the tabulation, the pattern-finding — is the part that automated first and most completely. It has been substantially commoditized within two years. And what has survived, what Pew is defending, what the industry's own budget shifts are moving toward, is the exact layer that was dismissed: real conversation with real people, and the trained human capacity to hear what is underneath the answer.

Listening was never the soft part of research. It was the part that could not be counted, which is a very different thing, and the profession confused the two for fifty years. It is going to spend the next five discovering that the skill it undervalued is the only one it cannot buy off a shelf.

The bridge

This is how I work and it is what I sell. The machine reads everything; I go sit in the room. The analysis is fast and cheap and I use it without ceremony. The interpretation — deciding which anomaly matters, which contradiction is a market, which strange behaviour is the beginning of a position — is the part I will not automate, because it is the part clients are actually paying for whether or not the invoice says so.

That interpretive layer is the spine of the Strategic Positioning Audit: not a data pull, which is now free, but the trained reading of what the behaviour means and where the consensus is wrong. It is what the 90-Day Brand Positioning Intensive installs — a brand built on an observed truth rather than a stated preference. And Positioning Sprint in A Box is the method for founders who want to run their own first pass this week. The founders in the community are the ones who stopped asking their audience what they want and started watching what they do.

Closing reflection

The room was too loud. Every respondent would have said so. And the volume was the reason they stayed, and the reason they came back, and eventually the reason the place worked at all — a finding that existed nowhere in the data and entirely in a Thursday afternoon.

We are about to get very good at generating answers and we are at real risk of forgetting what a question was for. The machine will give you the average opinion of everything anyone has already said, delivered with total fluency and zero hesitation, which is a magnificent tool and a terrible oracle. It knows the consensus. It cannot know the anomaly. And your entire future is in the anomaly.

So here is the question I would leave with every founder about to buy a research report generated in nine seconds:

When was the last time you were surprised by your own customer — and would your current method have let you be?

If the answer is that nothing has surprised you in a year, you do not have an insight problem. You have a proximity problem. Close the laptop. Go sit in the room. Watch what they do with their hands.

The machine will tell you what people say. Tradition — the real one, the one worth keeping — was always about learning what they meant.

B0LD is a cultural intelligence agency disguised as a marketing firm. We use the machine for scale and our own eyes for meaning. Start with the Strategic Positioning Audit, or explore the work at b0ld.ca.

SEO keywords: future of market research, AI market research, synthetic respondents, AI in consumer insights, qualitative research vs AI, market research trends 2026, ethnographic research, consumer insight strategy, say-do gap, human insight in market research.