Synthetic audiences: the perks and dangers
To what extent should agencies be relying on synthetic research?
Synthetic audiences: the perks and dangers
To what extent should agencies be relying on synthetic research?
By Conor Nichols, Special Reporter
There is no doubt that AI-powered synthetic audiences are saving agencies and marketers time and money. Imagine telling Don Draper he could skip the market research entirely. No smoke-filled focus group room, no clipboard-wielding moderator, no waiting a fortnight to find out whether the Kodak pitch would land. Just type the question and have Middle America’s opinion back before the coffee’s gone cold. He would probably have lit a cigarette and asked where the catch was. (He would have been right to ask… the Kodak pitch worked because of a human truth no model could have handed him - but please accept this analogy as an explanation.)
Synthetic research has moved from novelty to live decision-making tool astonishingly fast, and it is easy to see why. Electric Twin, a specialist in the field and one of the contributors below, says its synthetic respondents score 92% against a strict accuracy measure, against a 94% ceiling for real humans (who tellingly do not always give the same answer twice either). Figures like this make synthetic research sound less like a shortcut and more like a genuine substitute.
But “synthetic” is a spectrum.
At one end sits synthetic data:
- AI-driven modelling that scales up real, existing datasets - mobility data, panel data, subscriber data - to answer questions the underlying evidence can actually support.
At the other end sits synthetic audiences, or synthetic respondents:
- AI personas built from real data, then let loose to speculate on entirely new questions nobody in the original dataset was ever actually asked.
One is a model. The other is essentially a model’s guess about what another model might say. Both now get sold under the same word, and increasingly, to the same planner, under the same pitch.
With Kantar, Ipsos, Qualtrics and a wave of specialist platforms all racing into this space, agencies are being asked to put more and more faith in synthetic research, in live creative and media decisions, with less and less time to interrogate where any given tool actually sits on that spectrum. So, we asked people building and buying this technology, on both sides of the table: where does synthetic research genuinely earn its place, where should agencies be nervous, and how far should any of us be trusting it right now?
First, the view from the research and technology companies building these tools. Then, the agency perspective on putting them to work.
The people building the tools
Sushmita Jain
Product Director Data Science
Fifty5Blue
Synthetic data is a powerful addition to the modern data science toolkit, enabling teams to accelerate experimentation, solution development, test reactions, add scale, and do so cost-effectively without exposing real individuals.
But it only delivers real value under certain conditions. It must be grounded in continuous supply of real, high-quality data. It must be validated against real-world outcomes. And it needs to be used within the specific purpose it was built for – a model designed to estimate cross-platform reach may not reliably profile audiences or predict engagement. Outside those conditions, synthetic data risks adding noise rather than insight.
Before relying on it, agencies should ask: What real data underpins the model? How has it been validated? Does it preserve genuine behavioural relationships? What is the use case? It's also worth asking how the synthetic data was created, how much is real versus generated. The right spectrum of syntheticity depends on the use case and requirements.
Our Digital Twins solution builds synthetic personas by combining AI with high-quality, people-based data from TGI – long-running study of people's attitudes, behaviours and media consumption. The solution provides an always-on focus group that can be queried on demand to understand likely audience reactions, motivations, and behaviours. The most important takeaway from a data science perspective is that these digital personas are grounded in real, directly observed, high quality data from actual people.
AI will keep improving. That alone won't build trust – real data and rigorous validation will.
As regulation tightens, that's the bar we hold ourselves to: governance, provenance, accountability, built in from the start.
Euan Mackay
CEO
Route Research
At Route, the out-of-home measurement body for Great Britain, we're firmly embracing synthetic data. We're currently building a new generation of audience modelling that helps us understand how people move, where they go, how frequently they travel, and ultimately their opportunity to see out of home advertising.
What's important, however, is that our synthetic models are grounded in robust, real-world data. They are underpinned by high-quality first-party mobility and travel data, complemented by large-scale mobile datasets, ensuring that the modelling remains anchored in observed behaviour rather than assumption. We're also committed to transparency, documenting the methodology carefully so that it remains explainable and open to scrutiny rather than becoming an impenetrable black box.
For us, the value of synthetic data lies in its ability to responsibly scale high-quality evidence. It allows us to produce more granular audience insights, respond more quickly to market needs, and deliver more timely and reliable measurement for advertisers, agencies and media owners.
Synthetic data is based on mathematics, not magic.
As with any model, the quality of the output is determined by the quality of the inputs. The most sophisticated AI or statistical model can't compensate for weak source data.
When evaluating synthetic research, agencies and marketers should ask what real-world data sits underneath the model, how representative that data is, and how the outputs have been validated against observed behaviour. It remains just as important to invest in high-quality data collection as it ever has been. Synthetic methods can amplify good data effectively, but they can just as easily amplify bad data if the foundations are shoddy. Route and other industry JICs are uniquely placed to ensure due diligence here.
Leanne Tomasevic
Insights and Strategy
Electric Twin
Electric Twin works on the synthetic audience side. We start with real respondent-level survey data, often a client’s own, and use it to build synthetic audiences that answer new questions. The real data keeps answers grounded. The synthetic audience makes it dynamic and more useful, because clients rarely need to re-ask what they already know. We are giving a masterclass on the subject of synthetic data and audiences early November so will have more we can lift from that then too.
That grounding makes the work testable. We hold back part of each dataset, build the audience without it, then compare its answers with real respondents. On a strict measure that checks the full spread of answers, not just the most popular one, we score 92% (NDAM), against a human ceiling of 94% (because humans don't always answer the same way twice). In a pre-registered study with the London School of Economics, we commissioned an independent polling firm to run research across seven UK online panels, our results fell within the range of the human panels on most questions.
Synthetic audiences with Electric Twin perform well with concept, message and creative testing, naming, audience understanding and fast exploratory work before fieldwork. The Times uses audiences built on its own subscriber data, and makes business decisions related to product, marketing and commercial. The most stand out was improving subscriber retention by 7%. Wingstop used it to shape a lunchtime bundle, which lifted sales and improved margins.
Synthetic audiences shouldn’t be used for brand tracking, post-campaign lift, sensory testing, reactions to breaking news, or regulator-grade evidence.
Before using any provider, it is worth asking three questions:
- What real data is this built on?
- How was it tested on questions it hadn’t seen?
- Does each result tell you how confident to be?
Over the next 12–24 months, audiences will increasingly draw on a mix of survey and behavioural data (narrowing the gap between stated and actual behaviour), and start using autonomous research agents that can conduct deeper research on a topic for you.
The people putting them to work
Daniel Hulme
Chief AI Officer
WPP
The answer to this isn't straightforward, because it depends on the audience. Take CEOs, CFOs or CMOs, for example - ironically, those audiences are easier to model than others. CFOs tend to think quite homogeneously: they care about their P&Ls and the like. CEOs access similar information sources, like Gartner and so on. So there are some audiences you can model well, and what you can do is A/B test, or create focus groups, to then validate whether the actual audience responds in the ways your synthetic audience predicts.
The nuanced answer, unfortunately, is: it depends on the audience, and it depends on the data.
There's a deeper reality here too - when you can predict things, acting on that prediction changes the prediction. If you imagine you could create a magic oracle that could predict human behaviour, what you'd then do is influence that behaviour and shape it. So your model is now already out of date.
The question becomes: what new data do you need to retrain it, to make sure it's still representing that audience?
What we're doing at WPP is, instead of assuming age, gender and postcode represent how you perceive things, we're trying to understand audiences at the behavioural level. There are general data sources you can use to understand general human behaviour, but for specific audiences, you have to understand what data essentially influences them – and then how you compare that against focus groups.
You'll never be perfect, but the question is: is it better than the alternative? Even very crude synthetic audiences are great for testing content against very, very quickly, and for ideating. They're a powerful tool – not only to see how your target audience might feel and think about content, but also whether it might trigger a community you hadn't thought of. One of the things we're looking at is building representative audiences we can show content to and see - this might cause harm, or it might trigger an audience - so we can mitigate that risk early in the supply chain.
Venya Wijegoonewardene
Chief Strategy Officer
Spark Foundry UK
For me, the real value of synthetic audiences is speed and scale. They give planners an incredibly easy way to interrogate large datasets, test hypotheses and test hunches throughout the planning process. I find they help explore a problem quickly and refine the questions we might then take on to dig into. Of course, their usefulness is only ever as good as the data behind them – which makes the quality, depth and freshness of the underlying data really important.
I don’t see them as a replacement for qualitative research – quite the opposite. I think qual and synthetic audiences work best hand in hand because they play fundamentally different roles. Synthetic audiences can help you navigate breadth at pace; qual gives you the depth, texture and real human context.
The danger is treating a synthetic response as a proxy for a person, rather than as another source of evidence.
My rule of thumb would be to use synthetic audiences to explore and iterate, but qual is key to maintain in the loop for when you need to understand the ‘why’.
Scott G. Thompson
Associate Insight Director
Publicis Media
The important question to ask about data usually isn't "is it good?", but "what is it useful for?"
The numbers we get for people watching something on TV is a model - grounded in real data from real people, through panels, surveys etc., turning it into something useful for media planning and buying.
"Synthesising" data from different parts is nothing new. We use the best methods available to create the best data, to help us make the best informed decisions.
'Synthetic respondents' are something quite different. Rather than a model based on and grounded in ‘real’ data, it's a model (the interviewing 'agent'), creating more models (the respondents), and guessing what their answers should look like.
Often, they are good guesses! Sometimes, "right shaped" answers can be useful.
But what really matters is what they are useful for.
If you can re-run a simulation until you get the output you want, then what it’s really useful for, as the old saying goes, is "support, rather than illumination". Data that is useful to validate a decision is fundamentally different to data that helps you make the right decision in the first place.
So beware of getting what you want, instead of what you need. Because the most critical part of the aphorism "All models are wrong, some models are useful” is truly understanding what “useful” means.
Lena Lucas
Vice President Media
Unbound IA
Sourcing CISOs or CFOs for creative pretesting is notoriously slow and expensive. In my role as VP of Media, I view synthetic research as an essential initial sandbox for refining enterprise SaaS campaigns before launch. This technology delivers clear value across four key applications in B2B marketing.
Key Applications in B2B Enterprise SaaS Marketing
- Simulating the Buying Committee: Enterprise deals routinely involve over twenty stakeholders. Agencies must understand the unique job to be done for every committee member to produce effective content. Synthetic panels allow teams to build and test messaging tailored to each role at scale.
- Preflighting High CPM Campaigns: Before committing media dollars to expensive LinkedIn or account based advertising, media strategists can filter copy through role calibrated models to eliminate weak hooks before buying live impressions.
- Validating Dynamic Personalization: Synthetic personas evaluate vertical specific landing pages at scale, ensuring niche positioning sounds authentic across distinct industries.
- Stress Testing Competitive Hooks: Marketers can simulate competitor claims against synthetic buyers to uncover hidden objections and sharpen top of funnel messaging.
While these applications accelerate workflow, synthetic research works best when treated as an initial draft rather than the final word.
Synthetic findings must always be validated against real human surveys before launching full scale campaigns. In addition, agencies should exercise extreme caution when selecting vendors. Choosing platforms built on rigorous research principles and validated baseline data ensures accuracy. Using synthetic research as a preliminary filter ensures that when real human prospects finally see a campaign, the positioning is already battle tested.
Lee Wardell
Chief Creative Officer
Unbound IA
When synthetic research makes it easier than ever to get an answer, our job is to stay interested in what's actually true.
With budgets cut and timelines shrinking, using AI is becoming less of a choice. So agencies need governance: clear rules on what synthetic research can claim and when it must be checked with real people. But guidelines don't enforce themselves.
The real enemy is lethargy.
Nobody needs to be told that synthetic research should be balanced with real people. Every panel, deck and LinkedIn post about AI says it. Knowing it isn't the hard part. With barely any time or budget left, it's all too easy to decide the synthetic findings look convincing enough. Why bother listening to an actual person?
Used well, synthetic research can be a great addition to your process. I use synthetic data early to understand an audience and build a hypothesis, and synthetic audiences later to pressure-test creative. Put three concepts in front of a well-built persona panel and you quickly see what's landing and what's confusing. But as with all AI, you get out what you put in. Poorly built personas give flimsy feedback.
Some things only come from people. Recently, while we were sharing early thoughts with a client, their CEO made an off-the-cuff joke about how his prospects buy. One of the team noted it down. It changed the direction of the work, and the results. That only happens when you take the time to listen and don't pretend to have all the answers.
Laura Bebbington
Managing Partner, Insight
Havas Media Network UK
Synthetic research is about enhancing, not replacing human insight. Havas has integrated synthetic data strategically for years - filling gaps in modular surveys and more recently, powering our Science of Desire study, where predictable data structures let us generate credible estimates when primary data isn't available. And for our healthcare teams, synthetic populations have unlocked access to niche audiences previously impossible to reach at scale.
Both synthetic data and audiences provide greater speed, scale, and efficiency - making research nimbler. In advertising, where speed-to-market and agility define success, synthetic turns research into an enabler, not a constraint.
The caveat is that synthetic models are only as good as the real data training them. Ungrounded AI is unreliable. Quality output depends on bespoke, real-world data, refreshed over time to avoid synthetic outputs drifting from reality.
A further watchout is that synthetic data is a snapshot in time. Feed it a one-off survey, and it will accurately predict the past but miss yesterday's cultural shift, this morning's economic shock, or tomorrow's competitor move.
Crucially, it doesn't replace keeping our ear to the ground. Human insight captures immediacy, context, and cultural nuance no model can replicate.
The future isn't either / or. Synthetic data and human insight each have strengths and the smartest approach leverages both. We should use synthetic research to extend our understanding, while humans remain in the loop to interpret, challenge and validate it.
By combining them we deliver better, faster, smarter insights for our clients.
What is striking across nine fairly different voices is how little real disagreement there is. Nobody argues synthetic research should be avoided, or that it can be trusted blind. The discomfort is with the word “synthetic” doing too much work - flattening a scaled, well-grounded model and a speculative one into the same pitch.
The practical answer seems to be: trust it as far as the real data underneath it goes, and not a step further. Synthetic data can responsibly stand in for slow, expensive fieldwork. Synthetic audiences can sharpen the questions worth asking a real person. What nobody here will vouch for is a synthetic respondent replacing that real person’s answer entirely.
Don Draper never had to ask whether Middle America’s opinion was real. Agencies now do.
Published by the IPA, the IPA AI Magazine helps member agencies make sense of the fast-changing AI landscape, with practical insights, expert perspectives and real-world applications of generative AI.
Managing Editor: Conor Nichols
Design: Jenna Betts
© 2026 The Institute of Practitioners in Advertising. All rights reserved. No part of this site may be reproduced without our permission.
