Treat a synthetic audience as a rehearsal, not a verdict. Use it early to surface objections and sharpen your questions, watch for the biases it hands back as insight, and take the surviving ideas to real customers before you commit real money.
One of the stranger things AI can do for marketing now is let you test an idea on customers who don’t exist.
You build a set of synthetic personas, grounded in real data about a market, and then you ask them things. What do you make of this proposition? Which of these two offers would you pick? What would put you off? You can run something close to a focus group with participants who are entirely modelled, in hours, for the cost of the compute rather than the cost of a recruited panel.
It sounds like a gimmick until you’ve watched it surface an objection nobody in the room had thought of. It also sounds like a shortcut to certainty, which it absolutely is not. The trick is knowing which of those two it’s being on any given day.
What are synthetic audiences?
A synthetic audience is a group of AI-generated personas, built from real data about a market, that you can question and observe as a stand-in for real customers. Marketers use them to rehearse a reaction, test messaging, explore how demand might shift with price, or map which products people would trade between, all before committing to real research or real spend. The point is speed and cheapness at the early, exploratory stage. The catch is that a model of a customer is not a customer, and treating one as the other is how you get confidently misled.
So the honest framing, and the one worth holding onto, is this. A synthetic audience is a rehearsal, not a verdict. It narrows the field and sharpens the questions you take to real people. It doesn’t replace them.
Where it genuinely helps
Used at the right stage, this is a real capability, not a toy.
Before a launch, you can build rich, data-grounded personas and interrogate a concept against its intended audience, so the idea meets some version of its market before you’ve spent anything making it. You can run moderated-style sessions with synthetic participants to surface the obvious objections, the confusing language and the bits that don’t land, and fix them cheaply while fixing them is still cheap.
It’s also useful where real research would be slow, costly or hard to reach. Generating synthetic survey-style data can help you explore a market you can’t easily run fieldwork in, as a first read rather than a final one. You can model how demand might move as price changes, to find roughly where the sweet spot sits before you commit to a position. And you can explore substitution, which options customers might trade between, and how a change to one product could pull demand towards another.
None of that is the answer. All of it is a faster way to work out which questions are worth paying to answer properly.
Can AI replace focus groups?
No, and it’s worth being clear about why, because the temptation to say yes is strong when the synthetic version is so much cheaper.
A synthetic participant is a model of how people like this tend to respond, based on the data it was trained on. That’s genuinely useful for surfacing the predictable. It’s weak exactly where real research earns its money: the unexpected reaction, the thing nobody knew mattered, the emotional response that doesn’t show up in the training data, the specific quirk of your actual customers rather than the average of everyone. A synthetic audience will happily give you an answer to anything, delivered with the same even confidence whether it’s insightful or invented. Real people go quiet, contradict themselves, and surprise you. That friction is often where the value is.
There’s a quieter risk too. Synthetic data reflects the patterns it was built from, which means it can bake in the same blind spots and biases, and hand them back to you looking like fresh insight. If the model under-represents a group, so will your synthetic market, and you won’t see the gap unless you go looking.
How to use it without fooling yourself
Treat it as the first step, never the last. Use synthetic audiences to explore, narrow and sharpen at the start, then take the surviving ideas to real customers before you bet real money. Be as interested in what the model can’t tell you as in what it can. And keep the same governance you’d apply to any AI in marketing: know what data the personas are built on, and don’t confuse a confident output with a reliable one.
The teams who get value from this are the ones who use it to ask better questions, not to avoid asking real people. The ones who get burned are the ones who liked the speed so much they forgot it was a rehearsal.
What to do this week
Take a decision you’ve got coming up where you’d normally go on gut, because proper research felt too slow or too expensive to justify. A new proposition angle, a price point, a choice between two campaign ideas. Use a synthetic audience to rehearse it, purely to surface the objections and the questions you hadn’t considered. Then decide, with clear eyes, which of those questions are worth taking to real customers before you commit.
You’ll have done in an afternoon what used to wait weeks for a slot in the research budget, and you’ll know more precisely what you actually need to find out. That’s the win. Not skipping the real work, but arriving at it better prepared.
Weighing up where AI fits in your marketing?
If you’re weighing up where AI could take cost and delay out of your marketing without taking the judgement out of it, tell me what you’re trying to decide and we’ll talk it through.