How Synthetic Research and AI Personas are Redefining Market Insights

(00:03) We're super excited to introduce Innovative Orchestrator, the first AI super agent for omni channel advertising. We're bringing [music] humans and AI agents together so that marketers can conduct a symphony across channels to reach their most valuable customers. >> AI is not a matter of if, it's [music] a matter of how and when.

### Welcome to AI Insiders + why this topic matters

(00:23) And we will help you solve that. Hey everyone, welcome to the final edition of AI Insiders for 2025. Wow, I I I'm still uh having trouble remembering what year I'm in. Am I in the present, the future, the past? The ghosts of all the years past and present and future all over the place right now. So, uh, but very excited to be in the present with everyone from the AI Marketers Guild community and, uh, and we've got a terrific guest who I only recently connected with, but, uh, she is deep in a space and arguably has been helping uh, pioneer and advance a space

### Introducing the theme: synthetic research (and why it is controversial)

(01:12) that uh, I've been eager to learn way way more about. And that's this issue that's come up. It's even come up on the previous week. Uh it's come up during several topics. Synthetic research. It's a little controversial. It is uh very much at the forefront of where I think conversations around AI are going.

### Setting expectations: agents next year, but deeper focus on synthetic research now

(01:35) So yes, we'll be talking about agents next year and we'll be talking about you nano banana 80,000.968 and all these other crazy things going on. But as far as like one of the things that that I think can make a big impact, but is uh a little scary, misunderstood, uh and and needs like just way more understanding about a topic I want to cover a lot more next year.

### Guest intro: Jill Axline (Mava) joins

(02:04) Uh we've got a terrific uh guest and founder and entrepreneur Jill Axline who's who's got an exciting story for how she wound up uh running Mava. And we just recently met. I was like this is great. Any chance you're free? We we we got an opening soon and I'd love to hear more. So Jill, welcome. >> Thank you. So happy to be here.

### Quick rapport + holiday banter (Yoda)

(02:29) Nice to meet all of you. So, >> happy holiday. And I love that you have Yoda up there, [laughter] >> by the way. >> Oh, we we need all the the real real guru. I mean, there are some people who are pretty close to gurus in this room, but um but but at least Luciota is like OG legit guru. >> So, uh yeah.

### Jill’s background: enterprise marketer → research-led messaging

(02:52) Uh so, so Jill, why don't you just >> introduce yourself and tell us what you're doing and what Mava is and we can dive in. >> Absolutely. So, hi everyone. I'm Jill Axine. Um, you know, I you mentioned entrepreneur. I'm kind of, you know, new to this game. I really for the last seven years have been an >> entrepreneur by fire. Right. >> That's right. Exactly. Trial by fire.

### Starting point: content strategy role and audience-first thinking

(03:15) >> And so, yes, I I've been an enterprise marketer at a financial services company for the last seven years where um it was really a chip and a chair. Like, I started as a content strategist. I had to Google what content strategy meant when I was offered that role. You know, it's kind of an amorphous term.

### Research foundation: empathy, perspective-taking, and resonance

(03:32) Um, and from there, I think kind of brought into recognition for the firm that so much of what they were trying to do was about the beeps and boops of the product um with maybe a peppering of benefits, but they were losing sight of the audience. And um, you know, way back machine, I I studied it in my dissertation really audience resonance.

### Turning research into practice: building a market research capability

(03:55) So I was looking at empathy. I was looking at cognitive perspective taking. I was looking at counterargumentation and how do we build different things within the context of our strategic messages that are going to help engage our audience and help them have empathy with the message that we're we're sending out.

### From “answers inside the firm” to real customer research

(04:15) And so, um, I really, I think, brought that into the consciousness at the last firm I was at and and ended up bridging that out and building a market research team because up until then, and maybe some of you are familiar with this, you know, the answers of what the customer wants were within the four walls of the firm and maybe a little bit of, you know, conversations with with sellers.

### The synthetic audience spark: recreating focus groups

(04:35) And so, um, building a market research muscle was so important. And then that really came into brand strategy, segment strategy and we we came down funnel with we can't say solutions marketing solutions was always in jail word jail but um yeah I mean that's that's really how I got my start in in marketing and content strategy.

### Why synthetic research felt compelling: cost, speed, and freshness

(04:59) And so from there, I remember, you know, attending maybe two years ago, uh, Content Jam, which was a conference with Andy Cresadina, if you're familiar with him, and he was talking about, or somebody at the conference was talking about, let's scan LinkedIn profiles and create synthetic versions of people. And I thought to myself, okay, well, I've just run live focus groups with um, financial advisors.

### Desktop “always-on” audience and segmentation drift

(05:23) What if I could recreate them and set them into a synthetic focus group and then look at the disparity and convergence in the response and start, you know, thinking about what does it mean to have a synthetic audience and how can I build a desktop advisor or desktop asset manager that I can have all the time and I can put not just, you know, strategy and messages and understanding of trends and channel preferences, but I could also ask them about every piece of content or UX or anything that I wanted to know because when I was sending things out to market,

### Time-to-insight problem in traditional research

(05:55) not only was it incredibly cost prohibitive, we were talking to our audience maybe once, maybe twice a year at $50,000 or more, but I also noticed that it took so long to come out of fields that by the time I was really presenting the analysis, it already felt stale. And I also thought that our segmentation expectations, which is something we'd set in October or November and then maybe refresh quarterly, those were also stale and they changed across the funnel.

### Skeptic’s mindset and meeting Mava

(06:25) So if I was writing something highly topical at that top of the funnel, where I would draw those like segments and subsegments and how I would want to change my message and send it out through different channels would be different than at the product marketing level. So that's what really got me thinking like is there something to the synthetic audience? But at the same time being an academic I kind of walked into it as a skeptic.

### LLMs mimic tone; Mava aims to model thinking + feeling

(06:51) So cut two I mentioned this idea to um an agency partner I was working with and he's like you got to meet these guys over at Mava and um they were doing something that I hadn't seen. We were working with Jasper. We were working obviously with chat GBT um early stages of Claude and while those allowed me to create custom GBTs or spring up a persona even Qualrix had something on offer there.

### Swarm of models + governance for population representation

(07:18) Um all of them were large language models that were mimicking the language of my audience. They were coming forward with tone. But as you know, a social scientist who studied empathy, I'm much more concerned with how people are thinking and feeling because it becomes a lot more predictive of their behavior. And so, how can we get to some um representation of that in the models that we're working with? And it's not just one model that's going to do it.

### “Decisions shouldn’t move faster than evidence” + adversarial stance

(07:44) It'll be a swarm of models that are then governed and deployed um to create a more representative view of my aggregate population. Um, so that's the one thing. I think the second thing that Mava kind of brought into my awareness is that decisions shouldn't move faster than evidence. And so Mava is kind of a built skeptic.

### Measuring confidence, spread, and stability (vs. sycophancy)

(08:07) It has an adversarial model rather than a synthetic one or a sycopantic one. So I don't know if you guys have had this experience where you're talking to Chad GBT or some other um widespread novel model and it's always going to agree with you and it's always very certain of its of its output >> right >> whereas working with Mava it really opens up the kimono and provides an analytics blade that's going to tell me what is the confidence level here what is the spread of opinion in the audience that I think I'm talking to >> because maybe there's a huge spread and

### Using stability to decide: quick social vs. long-tail investment

(08:39) I need to break up that audience a little bit more and and really render my message differently. And then it also tells me what's that response stability. So, am I going to have a reliable response for my audience over time and build this into a longtail campaign or is this something that's really due to today's signals and news and um just, you know, what's in the ether today from a social listening perspective? because that's going to delineate whether I'm going to push something out in social today or build it into something I'm

### Controls to avoid “unexamined AI influence”

(09:10) investing in for the long term. And it kind of just goes on from there. Um I I like the idea that there are controls in place that would allow us to not um not get bound up in the kind of unexamined AI influence that that AI outputs are having on some of the work we're doing as marketers. Yeah. >> Yeah.

### Host reaction + transition into objections

(09:36) That's that's really what it's about. >> I love all this and and and even one of my favorite examples that I've I've mentioned this uh interactive journaling app, Rosebud, that I uh I've been using more the past month and uh and a future speaker here, David Levy, recommended to me. And actually the other day, I I shared something about something I was working on and it actually said to me, I call And I was like, like I didn't know you were coded that way, but this is fun now, right? Like when you when you see that, then it's like you can actually

### Big objection: can “past-based” synthetic audiences judge novelty?

(10:12) get in to what's going on here and you know it's like when it does agree with you, it actually like there's a reason behind that too. But what I'd love to hear cuz so just getting right into objections cuz I I think synthetic research and audiences are like like I like I I think it's you know it's it's it should be scary.

### How prediction works: affinity modeling beyond “I like ice cream”

(10:40) Like if it's not scary in some way probably not thinking hard enough about it. But um uh but uh for for all this, one of the biggest questions is say I've got like a new ad campaign idea or something like that. Like, how the heck can a synthetic audience that's based on past information and past inputs possibly be able to determine uh and analyze something new that hasn't been seen before? >> Yeah, and I think that is a great question and I think what we're getting to here is that a that past information is coded and we create a synthetic layer

### Emotion benchmarks (Harvard OASIS) and model accuracy gap

(11:24) on top of it. So you have all of those signals and everything that's coming from firstparty data. But here's an example. If your audience were to say, you know, in mass, I like ice cream. We could spin up, I like cold things. I like cold wet things. I like cold sweet things.

### Where synthetic fits: not pricing or conversion, but messaging lanes

(11:41) And with a greater preponderance of this data, we're reaching data in the billions of data points to start to model affinity and again emotion and cognition that becomes more predictive of behavior. So, we're about prediction. Additionally, because we're not just mimicking tonality, >> we're really we're really drilling into emotion, we've set benchmarks against Harvard's Oasis model to take a look at how audiences typically react to different types of stimulus in images.

### Using synthetic to narrow concepts before real market tests

(12:13) And we look at veilance and we look at arousal. And when we do that, we're finding a disparity between models like ours that are looking at emotion and cognition, and they're staying roughly within, let's call it. 02 um points um from a real human rating on average. When we look at the same thing with models like ChachiBT and Claude, they're roughly closer to like one to two and a half points off.

### Topic: does it replace focus groups or complement them?

(12:43) So when you're getting closer to that emotional resonance and that emotional response, that's how you can look forward and have more of a relevant scenario analysis of given this context, this is how the audience is likely to think, feel, and act. >> So, so with where this then fits into the toolkit or the stack. Yeah. >> Uh, another question I'm sure you get more than I I uh I hear is is like does it replace traditional focus groups? Is it in addition does it replace or be in addition to other kinds of like polling and surveys and creative testing and all

### What it does not do + what it does well (differentiation across funnel)

(13:24) kinds of testing? Like where does it fit? >> Let's let's start with where it doesn't. So I would rather say when it comes to actual conversion prediction, pricing, trust, all of this is settled by humans and it's going to set be settled in market. But when we're talking about um understanding themes that are overplayed in the competitive set and how to um you know cut through the noise and make sure that you're maintaining a less crowded lane with your message, >> uh this is a great place to find that differentiation across the funnel or

### Case example: brand campaign tweaks and lift

(14:00) throughout the buyer's journey. So that's number one. Number two, I would say yes, obviously creative and UX pre-esting are a good expression of that strategy. So those are the things that you can start get to the starting gate with the best possible set of concepts to then test and market and then prove out what's true what's not true.

### Synthetic as signal gathering → execution → optimize

(14:21) So I would say the synthetic audience is really to gather those signals build the right research and intel for how you even want to go into market and then execute against or operationalize some of those tactics build it pre-est it and then push it into market. Um, and having done that at my last firm, I can tell you that it gave lots of feedback about how we should come forward with our brand commercials um, across Chicago, New York, and London markets.

### From skepticism to adoption across functions

(14:50) And we made those tweaks after the second wave because we were all very skeptical in the first wave and saw a huge uptick in brand awareness, affinity, recognition for capabilities that we that weren't often known. and then engagement in that London market after we made the tweaks that the synthetic audience suggested.

### Beyond marketing: talent acquisition, EVP, RevOps

(15:08) So, I mean, again, I I think I'm probably the greatest skeptic of this and then um actually seeing it come to fruition both in my own brand team, but then also in RevOps and marketing ops. Um our talent and acquisition teams are using it to test um you know, prospective candidates, their employee value propositions.

### Multi-persona interaction: buying team in conversation

(15:31) anywhere where you want to plug in human intelligence into your process whether that's strategy and planning or execution or optimization I think there is a case to be made for the use of this tool or this type of tool >> so so this then like like with how these audiences are structured now are we at the point where where the different uh synthetic personalities uh uh I don't know it's defer to you how you refer to each uh individual um instance here like can they interact with each other do they need to is that

### Comparison mode now; orchestrating agent soon

(16:18) coming like is that unnecessary >> so I I think that that's coming right now we have a comparison mode where you can look and in my case at my last firm I would speak to an asset manager and an asset the donor. So, different parts of the value chain and I really want to understand where they're aligned on a a topic and where they're really coming apart so that I know how to message and build content differently for these audiences.

### Best first use case: simulate buying team roles

(16:42) >> Um, or get more bang for my buck, right, in a campaign that's going to be cohesive. Um, I think what we're seeing now with our our newest release is the ability of an orchestr like an orchestrating agent to call upon the various personas and put them into company with each other so that they can actually have an open dialogue.

### Validity question: side-by-side with humans?

(17:05) >> I'm trying to think about what are the best use cases within marketing where you want to see the various segments and audience talking to each other. For me in a from a B2B perspective, it's usually across arc types of a buying team. So I have a naysayer, I have a decision maker, I have a champion, right? And how are those interacting with me or interacting with my message after after they read what they've read and get together and talk about it internally because that's going to get them to the point of purchase. And so

### Validity: benchmarking + back-testing against reports

(17:36) that would be my first use case is to put a buying team into conversation. And we're not we're really not far from that. >> Well, is uh a great question coming up because you you started to talk about validity and and the Oasis model and and other ways that you're doing that. Uh but but we've got uh David asking in the chat.

### Bias extraction: why synthetic differs from live respondents

(18:00) So does your does your approach methodology involve a sideby-side test with real people, actual focus groups, panels, etc. uh uh can contextualize or validate the efficacy the efficacy for uh using synthetic audiences. >> So what we've done apart from benchmarking against the Harvard model is we've back tested when I was at my last firm we had annual reports on different segments.

### Disparity level and explanation

(18:25) We back tested them to see what the overlap would look like. We've also done that with some of the largest consult consulting firms that make their research publicly available. And what we're finding is there's really only a fourpoint disparity between live audience data that comes in through these surveys and the synthetic audiences.

### Why the gap can exist: social desirability bias

(18:44) And the difference here is, and I've worked with GFK, I've worked with Ipsos, so some of the largest firms doing market research. And the data is the data is the data. They don't really tell me what sort of bias is going to be implicit in the responses that I'm getting from my audience. What mava does is it actually pulls out the bias and it helps me to understand okay I might have gotten this answer from a live audience and this answer from the synthetic but this is why because they're saying what they think they should say and it's going to pull out what that bias is and

### Qual vs quant: mixed methods at scale

(19:16) it becomes the reason that we're seeing a difference between live audience data and synthetic audience data >> if that makes sense. >> Yeah. Yeah. So, while there there isn't a sidebyside panel, what we're doing is we're actually testing the same study across the two audiences. >> Mhm. Nice. Um I I I I want to already start open this up because this is a great uh conversation here and and uh let's see.

### Audience questions: qual vs quant and “focus groups” endpoint

(19:46) Oh uh uh Dan was just asking um do you think synthetic research is more powerful in qual or quant and uh and other folks who want to chime in uh on camera happy to just feel free to just raise your hands and stuff or call something out after this. Yeah. Um, so interesting you should say that we do have an endpoint called focus groups which is actually I think a misnomer because what it really does you speak with our chat you're talking to an aggregate of the population.

### Thousand-persona view + endpoints from trends to NPS

(20:17) When you bring it into focus group you're looking at a thousand or 10,000 versions of that aggregate. So now we're increasing the amount of entropy across the audience and we're able to produce mixed methods research. So it's both quantitative and qualitative and it can be something as um top offunnel or topical as a trend analysis or um channel preference and it comes all the way down to NPS scores.

### Example: qualitative brand associations at scale

(20:44) So with that, I mean, I could even show you um I was asking I think Volkswagen customers across four segments to provide what comes to mind when you think of Volkswagen and there were a thousand responses across four segments that were qualitative and very rich um in terms of what they were providing. So I would say it's both.

### Tool name and scope: Mava for B2B and B2C

(21:05) >> What's the tool you're using for your synthetic audience? >> Yep. It's called Mava. >> Ma. Okay. >> Yeah. like Maverick era. That's how I think of it. >> And and Michelle's wondering uh uh if the focus is more for consumer or can also be used for B2B audiences. >> Yeah. So I'm I'm B2B first. I'm kind of just becoming equating with um acquainted with direct to consumer.

### Persona creation approaches: AI assisted to advanced

(21:31) It's built for both. So I I can pop it up if you're interested in looking, but there are three ways to develop those custom personas. Some are AI assisted. Some go really deep for marketers that are wizards that can wield the wand, let's say. And so you're able to provide inputs across B2B and direct to consumer inputs because there's a little bit of difference between that and that will help you to spin up the custom personas within the platform.

### Parent and child personas from spread-of-opinion

(21:58) I would also say these personas serve as parent personas and very easily like I was talking about, you could start asking something and see that there's a pretty wide spread of opinion within your audience. And when you see that spread and understand the substance and it tells you where that line is drawn, you can then create child personas.

### Drill-down by market, archetype, or psychographics

(22:17) So within that higher order persona, now you're breaking it down and that could be by market or region. That could be and we have a level of specificity all the way down to the zip code all the way up to the global markets. Um or it could be by archetype within a buying team as I was saying before.

### Expert personas for niche domains (oncology example)

(22:34) Um, but there it could be psychographic and demographic, but it just kind of depends on where that spread of opinion is happening >> and like like what happens if if there's an area that's like uh like say someone's got like an oncology product and you don't have like oncologists built in, you know? Uh I mean does that kind of situation come up? Do you then like build those? Do you just hold off? Do you do some more general like what what do you do in that situation? >> So currently we do have the audience interaction but we also have an expert

### “Go-to-market team in a box” and custom experts

(23:09) interaction and we can spin up specific experts. Right now that expert I would call it like a marketing team in a box or a go to market team in a box because we have an SEO expert who is uh fully aware of age refs and somerush data and all of that and keeps its finger on the pulse of what's happening in the conversations on AEO and GEO um which is maybe anyone's guess right now but it it's staying very current relative to an SEO expert where I worked at an enterprise it was very hard to get on her calendar because she was triaging so

### AI data scientist agents and integration direction

(23:41) many things and I actually don't know how much professional development she was doing on the day-to-day basis because she was busy. So we construct these experts whether it's within the go to market >> you know realm or we can look to build a custom expert that is specific to a vertical. >> Awesome.

### Integrations: drive docs → posting → feedback loop

(24:04) Uh there's a question coming in about uh from Dave. Can you use AI data scientist agents to analyze the data from the synthetic audiences? >> I don't know how to answer that question because I have not tried it. But I I think >> it's kind of like a meta layer there, right? >> Yeah. I mean, I'm I'm certain that you could I think it would I would question what is the fidelity of the data science agent first and then Yeah.

### Platform direction: dashboarding and operationalization

(24:32) And then I would I think that's a really interesting idea. Mhm. I mean, well, and I guess a a related question would be that like as this develops, do you like do you see would you want this to be more of a platform that then plugs in to other kinds of dashboards and and tools? >> So, yeah, right now we have I think over 15 different integrations.

### “Why not just prompt a frontier model?” question

(25:02) So, um, it can start with plugging in your SharePoint or your Google Drive so it has access to your own strategic documents and then those stay secure and could just be referenced. Um, but we're pushing it all the way through actually taking the action of posting on LinkedIn for you and then analyzing what those responses look like and then feeding that back into the system.

### Core differentiation: prebuilt context + evidence + cognition/emotion fidelity

(25:23) So, um, yes, there will be a a tremendous amount of integration with your marketing tech stack. Um again from research and planning and strategy all the way through execution and optimization. >> Gotcha. Uh Adam, you want to come back in? >> Yeah, sure. >> Um so I I started typing out but I think it's better to explain it.

### Frontier models still useful; Mava’s value: relational RAG + confidence checkpoints

(25:46) I think >> Jill, this is you've clearly thought about this a lot. I always love uh being the contrarian and putting it back in front of you like why shouldn't we just use one of the frontier frontier platforms ourselves have chatbt construct the layers of prompts and blah blah blah is is it the simple answer of mo is just more efficient to get all that started up because you've already thought it through and then you have a more clearly defined taxonomy on the back end to break it down or like I I feel like there are aspects of this

### Deep business context across history and competitive landscape

(26:18) especially with those three buying roles you just talked about like you could do some of these things yourself and sort of trick yourself into thinking >> yeah it's a great question so to be super clear I use the frontier models too I don't just to use ma to the exclusion of the other I think there are really like specific times where ma becomes much more um useful in terms of having a lens into the audience and that comes from a couple different things is that synthetic layer of data that we talked about. So for every business I

### Swarm, governors, and live context minute-to-minute

(26:52) can create a deep business context. It's a relational rag database. Um instead of being tabular, it's creating a relationship across everything within my business and surrounding my business from um FPNA data, all all historic relationships between content I've put out into the world and engagement products and services, my entire competitive landscape.

### Hallucination checks and re-framing on high spread

(27:15) So that's all prepopulated and I don't have to set that context which is something you have to do with a frontier model. In addition to that firstparty data, there's that synthetic layer that I was talking about with the ice cream before. So as you start to get into billions of data points, you're going to get to see again that entropy across the data set a little bit more and get a better approximation of what uh the predictive aspect of this will be.

### Longitudinal company data + augmentation with proprietary inputs

(27:42) And you don't have that um hallucinate. You don't have that sense of hallucination where it always feels it's right. Ma is built with a way to have checkpoints against confidence. Again, um we're also looking at hallucination and and it will refra the entire audience if there's too big of a spread of opinion.

### Empathy as core of marketing

(28:02) Again, this is not something you're seeing with a frontier model. And then the last piece I would say is that fidelity to emotion and cognition. Because for me and as someone who studied this and very much feels empathy is kind of the core of all marketing, I would rather be looking at an audience that's modeled around what people are thinking and feeling than what they sound like.

### Timeliness: refresh rate and “news through the eyes of your audience”

(28:24) And that's what those frontier models are doing is they're taking all that first party data and they're just mimicking the tone of the audience. So it sounds really convincing, but it's really not connected to what they might actually do. >> And what they might actually do feeds off of primarily the first party data that is put into ma or is there secret sauce that's ignoring the sycopantric stuff and really >> it's it's less that the adversarial model is part of it.

### Validating outputs and reliability vs. frontier variability

(28:58) um uh which is that that GAN but I would say that even more than that it's that there are a swarm of models that are looking to model the population from different aspects of its thinking. So you've got how they think, how they feel. Those are modeled. You have a governor that's calling upon different models to then bring to the front or bring to the four what this audience is made of.

### Market research industry response and TMRE context

(29:22) And then finally, it's all built in context. So frontier models often have updates that will uh train the model but it will be delayed. Whereas we're in live context it updates minuteto minute. All of the data is stored temporally. So for the company I was at which was 40 plus years old. I have all that data from 40 years back all built into the system again without giving it context. And that's not to say that I didn't give it context. I gave it all of our strategic plans across seven P&Ls. I gave it um 40 transcripts from interviews with some of our top stakeholders on what they thought the brand meant to them. I mean things that you can't find out in in the ether. So you can always augment it with what it is you know to be true about your organization as well.

### Practical demo elements: evidence traces, confidence, and hallucination risk

(30:11) >> And there's a a great question that came in from Lisa. How timely are these personas? For example, a sample quir on the site. How do millennials feel about sustainability? Is this is the synthetic persona refreshing taking in new data and inputs from the wild or is it more of a snapshot in time >> which is what I was just saying.

### Market research validation: practitioners first

(30:33) So on those frontier models you do have kind of this delay on when the model was completely updated and pushed out right um >> for ma and these rag databases that I was talking about these uh deep business context they're updated minute to minute >> so every time I would go in I can I can show you every time I I refresh that model or go and take a look at those relationships it's it takes a second for it to load because it's always taking in context today.

### Market research practitioners vs vendors (GFK/Ipsos)

(31:03) In fact, we even have an endpoint that looks at the news through the eyes of your target audience and it will analyze given a story um and all the articles related to that story and how the story is evolving how that's affecting your audience today. >> Thanks D. >> Uh hi Joe. >> Uh just follow follow-up question. Um I think you had mentioned the major kind of market research players like the Ipsoses of the world etc etc.

### Adoption topic at TMRE + builder vs buy

(31:29) Um how have are you working directly with them? Have you have they have you approached them and are they like responding with excitement or do they have major concerns because I feel like this is could be very very pioneering for them. Um >> absolutely. >> I'm just curious to hear is is there like validation from that side. >> So interesting you should ask.

### Story: synthetic audience surfaced real-world security concern

(31:48) I'm because I was always on the client side. Most of my contacts are the you know heads of market research at firms. So, head of market research at Morning Star, at Coinbase, at Pepsi, those are the people that I've been talking to a little bit more. Um, at Capital Group, and you know, at TMRE, which is the number one market research event of the year, um, they all go, it's in Las Vegas. Um, that was back in October.

### Demo: agent trace (plan → research → execution → validation)

(32:13) This was one of the biggest topics, and it's builder buy, you know, and how do we how do we get to the fidelity? So, yes, I think there's a lot of conversation about this. I've not contacted um GFK about it, although it's a great idea. I'm kind of wanting to know what are the practitioners on the ground at the organizations who are trying to get that market research doing and what is their take on it.

### Concern: speed without receipts; data provenance

(32:37) So, that's really where I've started um since I joined Mava and it's been let's call it two months since I've taken on my role. So, more soon. More soon. >> Great. Thank you. Yeah, >> Ailen, I I'm actually going to call on you because you're one of the folks I was thinking of who would have so much to think and say about this.

### Case: security/transparency concern predicted; validated on LinkedIn

(33:01) [laughter] Anything you wanted to make sure we cover here? >> I I I want to share an experience with synthetic audience that maybe relates to it. So, I created a synthetic audience from uh this group. So, I have all of the video. I have a rack system. created the persona audience and one of the concerns that this audience always has is security and transparency.

### Importance of blind spots + proof

(33:31) >> Mhm. >> So I just want to show you how accurate it is. So recently I shared my tool on LinkedIn and then I asked people to comment on it and I think Paul Greenberg responded and his question was exactly that. He said how does your tool deal with security and how is my data protected? So as far as synthetic audiences I think you will get signals.

### Showing the agent output: evidence + reasoning + confidence flags

(34:04) I mean one of the values that it provides for you is your blind spots. If you're focusing on an audience and you want to know what their concerns map, then that is really a great way to test and learn about. So that's my own personal experience with synthetic audiences and how effective it may be. >> Very cool.

### Demo: confidence 40, hallucination risk high

(34:30) And I I mean I think data and security super important, but I mean what I've been kind of putting out into the world is just speed. Yes. And I understand that. And I went with our CTO to dev day and saw them construct an agent that moved all the lights in the auditorium in less than 8 minutes. So speed, yes, there's a lot of speed and impetus behind that.

### Persona creation workflow + assets/integrations

(34:50) But I think without proof or receipts is what I've been calling them, then it doesn't really mean a lot. And you have to be able to explain where did the model come up with this and how is it bearing out as it rep uh you know as it relates to my audience. So what level of confidence do we have before we put something in market um rather than just a gut feel.

### Tools called, personas used, and why

(35:10) And so that's I think what we've been really concentrated on is that data and security um that feeling that people can access any part of their stack and still have it feel like a closed system and then also understand how is this set of models working and where are they pulling that data from. So data providence I think is is a really important piece.

### Video analysis use case: speech rehearsal, differentiation, and delivery

(35:35) So just you know taking a look we're not going to do like a a demo but I just wanted to show you this one endpoint which is called the mave agent where I'm just saying I want to understand how to best position this business and how to approach comprehensive marketing strategy for them. So it's giving me that output that you would expect of any model but beyond that it's also pointing to the evidence of where it pulled from to get that and then also how it approached that.

### News endpoint: story impact through persona lenses

(36:02) So it detected the strategic query. It's restating what it is that I asked it to do. Then it's doing some planning. So what tools did it use in terms of web search and SEO analysis, but also what personas did it call upon >> to really provide the output? And it gives me a reasoning of why it did that, you know. So now I'm getting a why.

### SEO expert question: importing an expert video

(36:25) It also then shows the research phase and what it actually did. And then what does execution look like to actually bring this into fruition? And then finally validation phase. And it's telling me here that I have a confidence level of 40 and a hallucination risk that's high. So it's flagged that for me.

### Live updates: minute-to-minute refresh

(36:47) Now I have the ability to work with it to get to a different level of confidence before I'm going to even put it into any sort of application across strategy or execution. And so I haven't really seen any other platform that proposes, you know, these personas to have that kind of guess and check in place for me. And so that becomes a really important trace.

### Onboarding personas: assumptions, research, and intent

(37:14) Um, and it again, it'll flag specific things that it will want me to review before I put them into anything that I'm doing. >> So how how do you go about creating your personas again? So we do have um custom personas where again you have an AI assisted persona. This would be for like a less sophisticated business that maybe doesn't have a marketing lead.

### Uploading static persona decks and other assets

(37:36) You know uh it could be a startup founder who you know is more in the technological domain but needs to understand product market fit. Um, again, there's an intermediate step and then I think for probably most people on this call, there's an advanced pipeline where you're starting to establish what your assumptions are about your audience um, across who's the persona, what are the psychoraphics for B2B specifically, how am I talking about role type, industry, um, their tech stack, etc.

### Video analysis example: CEO speech coaching and outcomes

(38:08) For direct to consumer, it gets more into um, some of these psychoraphics. I can provide all the market context research that I either have on my own or that I've used another tool to derive. And then finally, I can provide, you know, what is the purpose of my developing this persona? Am I building a content strategy? Am I generating research? Is this for product marketing? Am I developing sales assets? And then I can start to upload any of the research I've done on my own because we all have those static persona decks with like Wendy wealth manager and her

### News analysis example: audience risk assessments

(38:38) picture and like what are her her pains and motivations are. So all of those could also be uploaded as well. And from there the system will generate it and give you a full look of pain points, motivations, channel preferences and then you can select what you want to integrate into your workspace. And I you know I can point out you also have the ability to upload your own assets or connect them through integrations.

### Comparing Mava vs frontier models: consistency and retest reliability

(39:05) Um like I had mentioned there is the expert panel. So for me I even have a type form agent expert who can help me program a survey if I wanted to or I can work specifically with you know an earned media expert. Um because sometimes marketing and comms are separated and I don't have access to a comm's lead at any given time. I had mentioned always a pain point for me at my last firm was getting time on the SEO experts calendar and so I love having this here and it seems so important because it seems to be like such a massive change for our you know for our

### Crisis PR potential application

(39:40) domain right now. So, >> so for SEO expert, could you for example just import a video of some SEO expert talking about the topic? >> Yeah. So, I can I can actually show you um something I did. So, I you know I was at Morning Star that was my last experience and um I used to work with our CEO to co-author speeches for him.

### Credit usage: rough magnitudes and how to control cost

(40:09) And so Morning Star Investment Conference was his biggest flagship appearance. And so as we were practicing, I would upload the video of what he was doing and it would analyze the substance of what he was saying, the creative that I put on the screen and his performance. Um, and I gave it, you know, the communication intent for the speech was primarily, secondarily, and tertiary.

### Political advertising and fast-changing audiences

(40:32) You know, these are the intents. Here's my goal. You know, I'm looking at building trust for the brand. And I'm thinking that, you know, we want people to think of us and our research as um something that is like their first stop on the bus and then I want them to go look at our products in the exhibition hall.

### Political use case: constituencies and committees

(40:49) And so based on that goal, it's going to analyze the video at 2C chunks and then it's going to put forward an analysis again of emotion, cognition, and behavior and then provide me with an opportunity to talk to my audience. In this case, I was talking to the advisor audience and I wanted to understand, you know, tell me how Kunol was resonating with you and also I had mentioned you have that rag database and I don't want to make you guys carick by trying to scroll up to the right spot, but I had noticed that we had a low differentiation score and so I started

### “How often do you update?” answer: minute-to-minute

(41:24) to ask the audience, you know, what is it that doesn't sound differentiated about the message here? I'm trying to find that because this was a little while ago that I did it. Um, tighten the speech with a differentiation sound bite. So, it started telling me, you know, where is it that we were falling down and what should he have said instead so that this sounded uniquely Morning Star.

### Closing: contact info and community wrap

(41:49) It was also looking in the topic I know Yan, but it was about the convergence of public and private markets. So, it started pulling from that rag database. Who else in the competitive set is talking about the convergence of public and private markets? What is it that they're saying and how can Kunal oneup them or how can we make it sound uniquely Morning Star when we're talking about this topic? It also gave me timestamps of where he should pause for greater emotional effect.

### Outro: gratitude and end of session

(42:15) So, we shared all of this with him from one practice to the next and watch his scores go up. And when we got to the actual conference, we saw greater traffic coming into the booth. Um we we heard more laughs because we actually do like chronicle how many times the audience is laughing and we did see more shares on engagement when it came to the digital version.

### Additional examples: webinars to atomized content

(42:37) Now could all of this be attributed to the work we did with the ma hard to make that attribution but it was a material difference from the last time over the course of seven years that I had been doing this with him. So um that's one you know video use case. We also took their top performing uh webinars which was I think the most performative channel for Morning Star's product marketing team and we used this to analyze what within the webinar was most engaging so that we could atomize that contact content and make you know different form factors for different

### News endpoint recap

(43:09) channels. So that's something that we also did here in in the video analysis tool. So again, everything that you're doing here is going to be put in front of your audience. And as I was saying before, the news is no different. So when I'm looking at the news, I can look at news in general, but I can also look at news through the eyes of my audience.

### Host closing + future conversations

(43:32) And so that's so important because we want to be highly topical. You can see I've set three personas because I had a brand campaign in market for those three personas. It looks like asset managers and heads of fintech were very concerned with the Bondi beach mass shooting. From here I can understand, you know, what is the story history obviously a summary and sentiment analysis of that.

### Audience impact and communications response

(43:56) But I think almost more importantly is that for any given news story I can run an analysis from the perspective of my audience and it's going to and this goes back to it being minute-to- minute. It's going to take a look at what is the audience's reaction overall. How do they feel about it? What do they think about it? What are they most likely to do? How are they assessing risk as a consequence to what it is that they do and their jobs to be done? And so, this is now going to give me a sense of how can I take this response and put it into action in my

### Wrap-up

(44:28) own communications with them. >> That makes sense. So, just kind of a a whistle stop tour through some of the different things that we have. Um, but I I mean it's become indispensable as a marketer for me and how I'm I'm doing different things on the day-to-day, but like I said, I still use chatbt and claude for this.

(44:51) >> You know what? And one thing also I think like for some folks who are used to just having their LLM of choice uh and uh and and it seems like you could debate this but it see it seems like the trend is they're getting smarter or getting richer information right and so so for for some of these interactions and you know we talked about like validity and comparing things like do you ever compare outputs for say like telling chat GBT or Gemini you're an SEO expert versus you know versus your own more custom audience. Yeah.

(45:31) >> Yes. And I would say and this is me being a PH dork but I think like the biggest issue here apart from my questioning how predictive those engines are would be the like the test and retest reliability. Okay. For example, um our CTO was making fun of me because I'm in my 40s and I've never made caramelized onions.

(45:53) And so he started asking and I know well I love I love eating them. I've just shame before shame. >> And so he was asking the models about you know how many people who are 40ome years old like have or have not made this and going from one to the next. Ma had a level of consistency and a score on response stability around both the quantitative and qualitative output.

(46:18) Whereas working with chat GBT in particular from minute to minute he was getting very different outputs. And that's something that you're going to see there whether you're saying you're an expert or you're just asking it a query it's not going to have that level of consistency. Mhm.

(46:45) >> I just want to say I I thought an interesting application for this could actually be crisis PR because you you have to go with something and you've got to react to something and it just seemed like when you did that there could be

(47:06) something you know I immediately thought of the news but I also just think of other things that happen in corporate life. So kind of interesting. >> Absolutely. Yes.

(47:06) >> I had a question um that uh example you just ran through. What's the rough order magnitude of the number of credits? I noticed that ma prices based on >> credit.

(47:06) So it it changes because it's generative. So it depends on what your query is. Um how heavy it is like in terms of how many documents are you loading into the query and how robust is that outbook and how long is the conversation. So I've had conversations where they've been like500 credits because I'm carrying it forward. My advice then to keep it smaller is to break things up into really discreet projects so I can be very goal oriented.

(47:31) This is the conversation I want to have. this is the output I want to get and then I can keep my credit allocation down. So, I've had conversations that have been 50 credits and like I said, I've had conversations that have been almost 2,000 credits. It it depends.

(48:01) >> No, but that's that's a helpful order of magnitude like you know, you've there's that starter plan 300 a month and um like you sounds like you could still do some very easy simple analyses multiple times over there.

(48:01) Yeah, I I have a number of firms that start there and try kind of get their feet wet to see like where are we finding the most value here.

(48:21) >> Sure. >> Um >> I would also say it doesn't preclude you from inviting your entire team.

(48:21) So because it's credit based, there's no like seat model like everyone on the team can take a look and have a conversation. it can be much more collaborative and break down silos.

(48:49) >> Mhm. I was just curious about um there seems to be a lot of accability to uh political advertising. Has there been any discussions on that and how you would uh simulate potential political audiences because I think this is where this the potential is.

(48:49) >> Yes. So actually we are working with a member of the House of Representatives right now to reconstruct not just his constituencies but also the committees within which he sits. So I think what he's wanting to understand is he wants a custom news digest that's going to take lots of information and consolidated and then he wants a custom dropdown so that he can start to look at things through the lens of any given audience at any given time and then he's ready with sound bites when he needs them.

(49:19) Um so that's kind of our first foray there. We're also speaking with um a lobbying group and so that will be another opportunity to uh really kind of understand how best to use the tool in this in the way that you're that you're suggesting. >> That's [clears throat] great. >> Uh how often are you guys updating the data set because obviously especially politically the times change three months, six months when not whatnot. I'm curious.

(49:46) How often are you updating your data set for your simulated audiences? >> So, it's it's minute to minute. And I I just put this up here because I wanted to show you. I just went to business knowledge and you saw that it was generating. So, in that span as we were sitting here maybe 5 seconds with latency, it's now updated this entire data set.

(50:12) And this is again like 40 plus years of data that Morning Star has been around, but it's current to the minute. So if there were a controversy that Morning Star was embroiled in like this morning, it would be here. >> Nice.

(50:31) >> So they would be [laughter] fair enough. No, you're covered. You're covered. 7:00, you're covered. But yes, awesome. Thank you so much.

(50:31) >> Yeah, of course. >> Well, well, this is amazing. >> Jill, what's the best way to get in touch with you and keep tabs on what you and Mava are doing? >> Yeah, so I mean I'd love to connect with anyone and everyone um on LinkedIn. Um you can also reach out to me at [jillmava.io](http://jillmava.io). You can visit our website and we don't have any credit card for trial.

(50:55) So, I encourage you to like jump in and kick the tires a little bit. And, you know, if you're running into any questions or want to think about how to best apply it, then just reach out to me because I' I'd love to work with you. I'm I'm more a marketer than I am like a like a CEO or a seller, any of that.

(51:10) I'm I'm one of you, you know. >> Well, well, yeah, this is great. This is a topic that I can tell from all the comments that have been flowing through here that >> we need to be covering more of because there's a lot to understand and and really I mean >> uh I I don't want to necessarily like give you a whole new like uh press kit you can use here based on this but just the way you're describing it even that political example at the end and thanks all for bringing that up.

(51:36) It's like like I feel like one way or another like professionally even personally right for things like cooking or dating or all these kinds of things that'll come up like we might have our own syn you know there's a lot of talk about having our own agents but having our own audiences to tap into and to yeah >> uh go and and use as an initial sounding board.

(52:03) >> Uh we're in some fascinating territory here. appreciate you sharing this. welcome back to participate and >> continue to be involved in the community and everyone here really like so many faces that have seen a lot here >> uh uh and know a few first- timers and folks who don't get to come every week but it's just like I look forward to this every week and I look forward to all the interactions come and I know so many of you interact with each other outside of this and whenever I'm talking to people about the community I'm like the Slack's

(52:33) great newsletters all this stuff Like all this stuff's wonderful, but like we actually get to have conversations and learn from each other here and I learn from all of you who show up every week and I'm grateful for that. >> happy holidays. Keep me posted on ways the community can better serve you in the year ahead.

(52:52) >> Uh it's been really exciting this year being doing this as part of the March family and >> uh lots more we're going to be building on. So Jill, thank you. Everyone, thank you. We hope it's a really fulfilling and wonderful holiday season for you all and >> uh and here's to a wonderful start of 2026 >> uh for each. So, thanks everyone. >> Thank you. Thanks for having me.

(53:15) Bye everyone.