Lifelike, emotive and rigorous: How we know synthetic personas really work
Editor's note: This article is an automated speech-to-text transcription, edited lightly for clarity. For the full session, please watch the recording.
On January 29, 2026, during the Quirk’s Virtual – AI and Innovation series, Toluna shared an overview of synthetic personas and how they really worked and were created.
Renee Smith, EVP innovation at Toluna, walked attendees through a series of examples to prove various aspects of Toluna Synthetic Personas. From proving the personas have emotions to ensuring cultural differences are included, Smith worked to persuade researchers of synthetic personas effectiveness.
Session transcript
Joe Rydholm
Hi everybody, and welcome to our session, “Lifelike, emotive and rigorous: How we know synthetic personas really work.”
I'm Quirk's Editor, Joe Rydholm. And before we get started, let's quickly go over the ways you can participate in today's discussion.
You can use a chat tab to interact with other attendees during the session, and you can use the Q&A tab to submit questions to our presenter today during the session, and they'll answer as many as we have time for at the end during the Q&A portion.
Our session today is presented by Toluna. Renee, take it away.
Renee Smith
Thanks. Thanks for having us, Joe. Thanks everybody for joining.
Today I'm going to talk about Toluna Synthetic Personas and much of the validation work that we've done, parallel tests and other sorts of analysis that we've done. This is sort of a companion piece to something we did in December where we talked about how they were created. I'm not going to spend as much time on how personas are created because I really want to talk about validation today.
The one thing to know is that on any page where you see this symbol during this presentation, it will indicate that all or some of the results are from Toluna Synthetic Personas, so we don't want there to be any misunderstandings. We will show results from parallel tests, and you'll see this symbol, but I'll call out, ‘okay, this is both synthetic and humans being compared.’
Let's level set though in terms of the word personas gets used in a lot of different ways. So, let me tell you what Toluna Synthetic Personas are and what their key attributes are.
First of all, we set out to work on quantitative research and being able to put individual level personas together in a sample. Our personas do not reflect a segment or a group. They actually have an individual level diversity and I'll explain in a minute why that is.
We set out to solve the quant challenge. We probably will move to qual at some point. It's on our roadmap, but we really wanted to try to solve the challenge of quant and synthetic survey taking and response generation.
This is where I can explain that part of the way we get the diversity, we're not creating digital twins. We are not taking observed full profiles of our panelists. We're actually using a small snippet of a few, about 10 or 12 variables, an anonymized seed profile, and from that, using retrieval augmented generation, large language models, actually also really important machine learning model that we've built, and we're predicting the additional attributes.
So, that is one difference as well. We're not creating digital twins. This doesn't reflect a specific known person, but it is because it's the retrieval augmented generation with guidance and guardrails, it does reflect a plausibly existing individual.
And then in terms of how we generate the responses, our personas operate within an agentic system. So, it's not prompt based, it's not a single prompt, it's not just one or two prompts. There's this full agentic system that includes a lot of checks and balances before the answer is even given.
So, just a little bit of background, but again, I can't on this presentation go too much into right now how they're created. We can talk about that in the Q&A if you'd like.
What do we currently offer in terms of synthetic personas? This isn’t a sales pitch. It's just so you know where the validation testing has been, most of it that we've done.
We have Act Instant AI, which is a copy pre-testing solution that uses 100% synthetic personas. It's automated end-to-end. And we have a rapid claims screening solution that also uses 100% synthetic personas. So, much of the validation, not all, but much of it that you'll see will come from these as we validated these two solutions.
We often get asked, can personas have emotions? So that's what I want to address first.
The first thing, and people kind of get a kick out of this, it's also a little disconcerting to realize that they, behind the scenes, our personas grumble about poorly designed surveys just the same way humans do.
So, think of it almost like a whiteboard. They have a place, the personas have a place where they can say things about what's occurring while they're taking the survey. You can see this one that talks about, “I'm getting very, very, very annoyed with answering the same question.”
We intentionally gave the personas the same question and the persona just kept adding the word very each time they were getting more and more frustrated. Now, they will always answer honestly because that's one of their goals, but they do have emotions.
You can also see them grumbling about a survey being too long. My bad back is starting to act up again. So that was a persona that had arthritis among their profile attributes.
We weren't really expecting this, but we also had a persona that was a non-drinker as a profile attribute who said, "I really think this branch should avoid showing a bottle of beer inside its product because it's triggering." So, they have emotions, they have reactions as humans could as well.
Another aspect, if you want to think about emotions and personas, this is showing for advertising, emotional resonance. So, what you have on the left-hand side is you have the percent of global ads that were coded by humans. This was some work we did entirely separately about a year ago, coded by humans as to whether those ads have an emotional storyline.
And as you would expect, television ads, a higher percentage of them tend to have a more emotional storyline. There's often more room for the narrative and the narrative arc.
What you see on the right-hand side, and you see the symbol down there for the synthetic personas. What you see on the right hand side is that the synthetic personas, when they were asked to provide information about whether this ad was resonating emotionally, a similar pattern, basically TV ads. They recognize that those have the highest emotional narrative content, TikTok, Instagram Reels, they recognize as having the least.
So just another angle on if they handle emotions.
The third piece of evidence I'll show you, this was a particular study we did. This used that claim screening and you can actually screen other things, you can screen messages. It's got a MaxDiff trade off, two MaxDiff trade off exercises in it.
In this particular one, the client wanted to understand emotional associations to the brand and gave them a trade off exercise with a list of 30 emotions or feelings, depending on how you define the word emotion. When we looked at the parallel test with humans, we had in the rankings that came out of the Max Diff, we had a 0.86 correlation between the way the humans made the trade-offs about the emotions and the way the personas made the trade-offs about emotions. In this particular case, we had the top two were also exactly the same.
And what's particularly impressive to me about this one is, I can't reveal the brand, but I can tell you that it makes sense that they would want to know what emotions are associated because there's been some challenges for them.
The emotions range from everything like positives, satisfied, appealing, joy, all the way down to disgust and contempt.
So, just three pieces of evidence coming from different directions. Yes, personas can have emotions, they can recognize emotion and they can associate emotion with brands.