During my time running an analytics agency, nearly every client walked through our door believing the true size of their convertible customer base was much larger than it actually was. To help them measure their business more effectively, I would map their use cases to align product and marketing teams on “value” from the customer’s point of view. 

This approach did more than just inform tracking plans—it sharpened the definition of prospects that were genuinely in-market for their current offer. I’d watch stakeholders go from “oh crap”–as they realized they’d wasted millions of dollars chasing the wrong prospects–to a rush of excitement as they finally understood the problem and how to address it. 

The way you think about your market influences how you try to serve it, whereas a misread market can lead to absurd situations—like Ryanair selling pork sandwiches on flights to Morocco during Ramadan (it happened). That’s why there is a direct connection between the perspectives you hold about your market, and your bottom line. But getting people to change their mind can be easier said than done. 

In fact, in all the years delivering data insights and dashboards to teams, even when armed to the teeth with evidence, some opinions are hard to shift. Imagine my surprise when sharing a screenshot asking virtual GenZers for their opinion on a brand name—one that a founder was adamant about rebranding too–and watching him abandon his own idea almost instantly—a reaction that dashboards or A/B tests rarely create.

As one customer put it, chatting with role-playing agents feels like having a “thinking partner” who doesn't hold their punches. Unlike traditional interviews—where social pressures can lead people to give filtered answers—LLM role-play creates a safe space for genuine insights. This environment encourages unexpected anecdotes and associations that can spark creative breakthroughs.

Want to know the best part about it? The polls don’t even need to be accurate. So long as the responses feel real enough, you can stress test your ideas and avoid groupthink. And because AI agents role-play better when fed with more context, the benefits of virtual audience testing can help justify more effort into high quality traditional research… because one helps strengthen the value potential of the other. In this article, I’ll show how even a basic virtual audience can help you tap into the wisdom of the crowd, and demonstrate how providing rich context improves role-play.

Can a Basic Audience Be Helpful?

15 minutes isn't enough time to build an advanced audience that mirrors lots of nuanced characteristics. Even though you can model a niche market in Rally, during onboarding demos I keep it simple by using one of the default audiences loaded into your account (e.g. general US population) with the basic OpenAI model, so i’ll do the same here.

Let’s imagine you’re the owner of a gym for a moment, and want to know how people relate to positioning for recovery, functional fitness and mobility workouts.



A quick scan of the responses immediately reveals common themes.

“...I’ve been to those high-intensity gyms where everyone is just trying to outdo each other and it’s exhausting. A place that prioritizes how your body moves and feels? That’s where I want to be.”

“...we’ve been programmed to believe that if we’re not sweating buckets, we’re not doing it right. But focusing on flexibility and mobility? That’s smart.”

“...Too many gyms are all about pumping iron and chasing that perfect body image.”

I hadn't considered the cultural impact of a gym focused on recovery and flexibility. Nor the perspective that if we’re not sweating, it ain’t working. Right out of the gate I’m having ideas on messaging; “train smarter, not harder”. But if we continue to dig, there’s more. Little cues on how these role-playing agents relate to the idea that make me rethink about potential market segments. Ideas that could justify running a real market study to validate.

But first let’s ask Rally another question using vote mode, which gets these agents to choose between a list of options. I asked them to rate purchase intent for a $250/month gym subscription, but you could poll brand names, search results, business ideas, specific offers, messaging, or upload some selfies wearing tie / shirt combos for your next TedTalks. 

We also get to see the inner reasoning behind why these AI personas voted the way they did. So you could compare the responces of personas who voted 2 vs those who voted 5 to tease out any signals of resonance or list out objections. Both can be incredible useful when thinking up new features or hooks for your marketing campaigns.



In this case, Kevin the high school teacher appears to be price sensitive. A few of the pensioners also noted the “steep price”. This already makes me think about targeting tech workers… perhaps more likely to be in need of mobility workouts to offset long hours behind laptops. I’d imagine them having higher earning power.

But what if we wanted to tap into those price sensitive markets and tackle some of these objections? We could put forward some ideas and continue getting these virtual personas voting. But what if you’re out of ideas? Then just ask Rally what they’d need to vote higher. Since Personas maintain memory within each session (threads you can return too later), those who voted 2, will share with us what they’d want to see to vote 3. 

In the example above, Kevin who originally coughed at $250 is now saying he’d consider it for $150, but wants to see functional fitness workouts that targets different age groups, community, and tailored advice. Or Emily the Nurse mentions more holistic classes on wellbeing, mentions “nutrition”, “stress management” and “yoga”. Reviewing two responses already gives me a bunch of ideas on how to go about strengthening an offer. From trial periods and buddy systems to live streaming super niche workouts (e.g. “yoga for retired”).

At this point we’re not just rethinking our segmentation, but ideating offers, entire solution paths... ideas you can export to discuss with your team. The major advantage of speed running qual discovery like this, is we can ask questions until our eyes bleed, without burning social equity. Losing arguments over and over again. But then winning the crowd when it matters.

"In the end, people will judge you not by how many times you failed, but by how well you played your final hand."— Unknown

Does More Context Make Responses More Realistic?

I've shown the power of tapping into a crowd of role playing AI agents as your thinking partner. But does spending more time with audience design improve responses? To answer this I wanted to see how well LLMs could role-play fictional characters. 

So I used Rally’s clone via text to generate some virtual Avengers (yes MCU characters) using data from a Marvel wiki. But you could replicate this with data about real people (e.g. transcripts, comment threads, reviews, autobiography, LinkedIn profile). Go nuts.

Let’s ask Rally. “Describe your unique method for unwinding.”

What do you think? I suspect Bruce Banner might unwind by “...retreating into the wild”. Makes sense. When he gets angry heavy objects start getting thrown about and everyone suffers. And nothing gets people riled up like sitting in a traffic jam.

What about spider man?



“...perched on a rooftop” makes sense too. And this is just with the basic OpenAI model which seems to be role-playing pretty well. 

Let’s see if we put them in a situation where they might choose to use super powers. 

Good to see Wanda using chaos magic here and creating a barrier. That’s one way to stay clean in a food fight.

Now imagine what you could learn by asking virtual audiences questions, especially when getting creative with cloning using data you have about personas of interest in combination with smart prompting. This is where I anticipate AI adopters will gain an early mover advantage... gaining skill and experience in figuring out the right combination of audience design and prompt engineering to calibrate responses and make simulated environments more useful for a given use case. If you’d like to stay ahead on these topics, I've created a community for us to gather and share notes. 

Feel free to join so we can figure things out together.