Market Research Is Being Redone from Scratch by Generative AI

A while ago, I saw a set of numbers that genuinely shook me.

A research firm found over 170 people who actually work in market research and asked them: Are you using generative AI?

The result? 45% said yes, already using it. Another 45% said they're getting ready to.

Add those together? 90%.

You read that right. 90% of practitioners have either gotten their hands on it or are on their way there.

What does that mean? It means generative AI in market research is no longer a question of "should we try it" — it's a question of "if we don't, we'll be left behind."

Today, I want to talk to you about how this is actually happening, and where it's going to take market research.

First, a Baseline Judgment: It's Not Here to Help — It's Here to Redo Everything

A lot of people imagine generative AI in market research as "help me tidy up interview transcripts, help me write reports."

Sure, that's something it can do. But if that's all you see, you're selling it way short.

A group of researchers at Columbia Business School spent two years deep in the trenches with a bunch of companies in this space. They arrived at a judgment:

Generative AI's transformation of market research unfolds across four layers, each one deeper than the last.

What are the four layers?

Layer 1: Make the current work faster and cheaper. For example, turning hours of interview recordings into a summary of key points in minutes.

Layer 2: Use synthetic data to directly replace human research. AI generates synthetic people that simulate real human responses.

Layer 3: Surface insights that were previously impossible to obtain. For instance, customer segments you can't reach — use AI to simulate them and test your product ideas.

Layer 4: Create entirely new types of data that never existed before. Like a "digital twin" of every customer — a virtual replica that never gets tired, never gets annoyed, and can be tested anytime you want.

These four layers escalate in ambition, one by one. Let's look at them one at a time.

Layer 1: Make the Current Work Faster and Cheaper

This one's the easiest to understand.

In the survey conducted by research firm GBK Collective, among those already using AI, 62% use it to condense lengthy interview transcripts; 58% use it to analyze data; 54% use it to write reports.

Think about it — what used to take a researcher a full week, reading through 50 interviews, now produces conclusions in a single afternoon.

But here's what's interesting: the real breakthrough here isn't about speed. It's about depth.

A startup called Outset.ai did something very clever. Instead of having AI generate answers, it has AI generate questions — dynamically producing follow-up questions based on the respondent's previous answer.

And the result? Respondents are actually more willing to open up to an AI than to a human interviewer.

Why?

Because when you're talking to a real person, you worry about being judged. With an AI, you don't carry that burden.

The head of research at WeightWatchers said something that really stuck with me. He said researchers used to be forced to choose between "the depth of an interview" and "the breadth of a survey." And now?

"We want both."

You see, that's what this layer is really changing. It's not about saving time — it's about making the old "you can't have it both ways" trade-off suddenly unnecessary.

Layer 2: Replacing Humans with Synthetic Data

What is synthetic data?

Simply put, it's people conjured up by AI — but the responses of these "synthetic people" are statistically close to those of real humans.

It sounds far-fetched. But one story genuinely blew me away.

A startup called Evidenza partnered with Ernst & Young on a double-blind test. Ernst & Young handed its annual brand research questionnaire to Evidenza — but withheld the real results. Evidenza used AI to generate over a thousand virtual CEOs — CEOs of companies with a billion dollars or more in annual revenue — and had these "synthetic CEOs" fill out the questionnaire.

Then the reveal.

95% of the findings matched the real-human results almost exactly.

Many of the numbers were nearly identical.

The CMO of Ernst & Young's Americas division said: "The results were stunning."

Impressive, right?

But this layer has its ceiling, too. In the survey, only 31% of respondents felt that AI-generated data was "highly valuable" — the lowest-rated application. Why? Because AI can't simulate the full unpredictability of human behavior. It can simulate "how most people would choose," but it can't simulate "that one unexpected choice."

Here's the key finding, though: if you feed AI past survey samples along with some Q&A examples, the quality of its synthetic data improves dramatically. It's not inherently smart — it depends on what you feed it.

Layer 3: Surface Insights You Previously Couldn't Get

This layer, I privately think, is severely underestimated.

Many companies love to say they're "data-driven," but the reality? Columbia's researchers found that the vast majority of decisions are made without any formal research backing them up.

Why? No budget. No time.

But generative AI can serve as an always-on market advisor. You have a product idea — toss it in, and it'll test it for you. You want to know what your competitors are up to — hand it over, and it'll piece together intelligence for you. 30% of people surveyed are already doing exactly this.

The head of innovation at General Mills said something especially grounded. She said they're exploring synthetic data to accelerate the creative screening of product ideas — so that genuinely good concepts get identified earlier.

You see, the logic has shifted here. Research used to be "in support of a decision that's already mostly made." Now it's become "explore the full range of possibilities before you make up your mind."

This is a shift from rubber-stamping decisions after the fact, to navigating possibilities before you decide.

Layer 4: Digital Twins

This is the most science-fiction layer — and it's already happening.

What is a digital twin? It's a virtual replica of you, built from your public or private data. This replica doesn't get tired, doesn't get annoyed, and can be tested over and over again.

40% of people in the survey are already experimenting with digital twins. Another 42% are planning to.

Ogilvy is already using digital twins to test advertising creative — before sending it out, they show it to virtual customers to see whether it would move them.

But what really made me think "the game is about to change" is the project underway at Columbia Business School.

They're building a virtual panel of 2,500 people. Each person is a digital twin of a real human. The real humans complete an extensive battery of psychological, behavioral, cognitive, and attitudinal tests, which serve as "ground truth." Then, AI uses this ground truth to generate virtual replicas.

What's the point? So that future research and surveys won't need to recruit real people — you just run them on the virtual panel.

Is this reliable?

Stanford and Google DeepMind conducted a joint study. They first interviewed real humans for two hours, then used the interview transcripts to generate digital twins. Then they had both the real humans and their digital twins answer the same questionnaire.

Two weeks later, the real humans answered again.

The result: the digital twins' answers matched the real humans' follow-up answers with 85% accuracy.

In other words, a digital twin simulates a real person almost as accurately as that person taking the survey again themselves.

But Don't Celebrate Just Yet

This section, I have to include.

Generative AI in market research has its hard limits.

First: Bias. A 2023 study by Columbia and Stanford found that OpenAI's models express opinions more aligned with liberals and the well-educated than with people over 65 or the religiously devout. And the newer the model, the more pronounced this bias becomes.

Why? Because models are shaped by their training data, and the training data is itself skewed. On top of that, when humans fine-tune the models, they bring their own leanings into the mix.

Second: Poor stability. A 2024 study by Vanderbilt University found that while synthetic respondents produce answers close to those of real humans, they exhibit less variability (answers tend to converge), are extremely sensitive to question wording (change the phrasing, and the conclusion shifts), and produce unstable answers when re-asked three months later.

Third: Inability to simulate experiments. Columbia's researchers ran their own test: they had AI simulate consumer purchase intent at different price points, and the resulting demand curves not only diverged from real-human data but were logically inconsistent.

What does this mean?

It means synthetic data can be used for a lot of exploratory, supplementary work — but when you need to make a judgment call that truly determines the fate of your company, human research cannot be replaced.

It's a good copilot, but you can't hand over the steering wheel entirely.

After Looking at All These Cases

I'll tell you — after going through this whole round of cases, I felt a bit conflicted.

On one hand, this is really happening. Not a future that lives in a slide deck. 90% of practitioners are already on their way, and if you haven't touched it yet, you're falling behind.

But on the other hand, something didn't sit right with me.

Think about it: an AI-generated CEO fills out a questionnaire and matches real humans 95% of the time. Sounds impressive. But then again, brand research questionnaires are inherently formulaic sets of questions — and formulaic questions naturally produce formulaic answers. A 95% match might just mean the questionnaire itself wasn't asking much of real substance.

So my judgment on all this is: don't underestimate it, but don't be seduced by the flashy percentages, either.

What's truly valuable isn't knowing how to use AI — too many people already know that. It's being someone who understands AI's temperament, has done real research, and knows when to trust the machine and when to turn it off. People like that are incredibly rare in the market right now.

The longer I work in this field, the more I believe one thing:

Tools keep getting more powerful. But judgment — judgment has never lost its value just because the tools got better.

In fact, the more powerful the tools become, the more you have to think clearly about: what should you hand over, and what must you absolutely never hand over?

Figure that out, and only then can you truly get on board.

My wish for you: think it through first, then get on board.