There's a new trick going around. You ask a model to "act as our target customer" and then you interview it. What are your pain points? What would make you switch? What do you think of our pricing page?
It answers instantly, confidently and in full sentences. It's also mostly telling you what the internet thinks a customer like that would say, averaged out and politely phrased.
That's not research. That's a horoscope with a nicer font.
Why synthetic customers feel so convincing
The answers sound plausible because they're built from plausible material. A model has read a lot of reviews, forum threads and marketing blogs, so "a busy marketing manager" sounds exactly like the busy marketing manager in every marketing blog.
But your customers aren't the average. The useful stuff in research is almost always the weird bit: the objection you didn't expect, the word they use that you never would, the competitor you didn't know you had. Averages sand all of that off. You end up confirming what you already believed, faster.
What actual research looked like, and still should
The best work I've been part of started with people who were not us. At Global Citizen, the campaigns that performed were built on what the audience actually responded to, not on what we assumed someone who cares about climate or health wanted to hear. The research was boring and slow and it was the reason the work landed.
The old-school toolkit hasn't changed:
- Talk to customers. Five honest calls beat fifty synthetic ones.
- Read what they already wrote. Reviews, support tickets, sales call notes, comment sections, the "why did you cancel" form.
- Watch what they do. Search terms, the pages they leave from, the emails they never open.
The only thing that's changed is how fast you can get through the reading.
Where AI is a genuine superpower
This is the part I love. The slow bit of research was never the talking. It was the sorting. AI is brilliant at the sorting.
- Clustering real feedback. Paste in 400 reviews and ask for the recurring themes, with a direct quote for each. Then check the quotes exist.
- Pulling the language. Ask for the exact phrases customers use to describe the problem. That's your headline copy, written by the people you're selling to.
- Finding the outliers. Ask what shows up rarely but strongly. That's usually where the insight is hiding.
- Prepping the interviews. Have it draft a discussion guide, then cut the leading questions it inevitably writes.
- Summarizing calls. Transcripts in, decisions and objections out, in a format sales will actually read.
Notice the pattern. Real customer input goes in. AI does the heavy lifting. A human decides what it means.
One rule that keeps you honest
Every insight that makes it into a strategy deck should trace back to a real person. A quote, a ticket, a recording, a number. If the source is "the model said so," it goes in the bin, or at best in the hypothesis column to be tested.
That one rule will save you from the most expensive mistake in marketing: building a whole campaign for a customer who doesn't exist.
So, use the mirror if you must
I'll admit synthetic personas have one use. They're a decent warm-up before you write a discussion guide, a way to spot the obvious questions so you can skip past them.
Then close the chat, pick up the phone and talk to a human who has actually paid you money. They're less articulate than ChatGPT. They're also right.