Commentary
Dentsu Begins Using Bots To Synthesize People For Planning & Buying
- by Joe Mandese @mp_joemandese, 9 hours ago
Just as the ad industry begins grappling with the concept of targeting ads at bots as a proxy for reaching people -- so-called agent-to-agent marketing -- a major agency holding company this morning announced it will now use bots as surrogates for how it plans and buys media to reach humans.
Characterizing it as a "next gen planning capability tailored to the evolving demands of the Algoritmic Era," Dentsu this morning announced a deal with "synthetic research" platform Evidenza to combine AI-generated audience "respondents" -- facsimiles of human research respondents -- into Dentsu's own proprietary consume research panel.
That panel, Carat's so-called Consumer Connection System (CCS), has long been the pride and a point of differentiation for Carat, and ultimately the rest of Dentsu's media services and consumer insights organization for more than a quarter century.
advertisement
advertisement
Dentsu did not disclose how Evidenza's synthetic respondent data would be integrated with or incorporated into its human-based media planning and buying research, but it characterized the initiative as "fostering a new culture of experimentation -- one that delivers greater depth, speed, and agility in planning and activation."
Here's how Dentsu described its new process in this morning's announcement from the Cannes Lions Festival in France:
"Synthetic audiences are AI-generated digital profiles that reflect real-world audience behaviors and preferences – a fast, cost-effective and flexible alternative to traditional methods. They will allow strategy and planning teams to quickly get closer to their core audiences, unlocking deeper insights and discovering new segments for clients."
While not new, the methodology still is considered novel, even among bleeding edge consumer researchers, raising numerous questions of the ad industry's "consumer insights," especially what impact synthetic audiences have on a vexing long-term consumer research issue: "respondent bias."
That said, Dentsu Chief Data & Technology Officer Shirli Zelcer disclosed the agency's initial results "show a 0.87 correlation with traditional research
— proving we can match
the rigor of legacy methods while accelerating time to value and unlocking new growth opportunities.”
To understand the process and rationale utilized by Evidenza, you can watch a brief explainer video on their site here.
Joe Mandese is the Editor in Chief of MediaPost. You can reach Joe at joe@mediapost.com.
Sounds like a multiple correlation simulation model designed to predict actual audience attainment and, maybe, ad reach/frequency, etc--all supposedly producing outcomes. What I don't get is that media plans for a future year are usually not specific as to what TV or other platforms---what TV network or social media website, etc--- are to be used nor what vehicles--TV shows, magazines, etc---wil carry a brand's ad messages. So what data is being used? Or is the model creating synthetic data for its machinations?
@Ed Papazian: It's not a model. Synthetic audiences are AI-generated respondents that participate in panel-based research just like human respondents. Basically, they're asking the synthetic respodents questions to generate responses as if they were humans responding to the questions.
But Joe, when you create a systhetic respondent don't you have to give that synthetic person characteristics, mindset and behavior patterns which, I assume, would be based on some sort of complex profiling. For example, say you ask one of your synthetic respondents ----who is a man, 36 years old, married with two children, working at a markting company who is a light viewer of TV, etc. etc.---- questions about what media he uses or how he might respond to excess ad frequency, etc. To get that far, don't you have to define each synthetic respondent's activities, interests, attitudes, etc. so "he" and others like him can supply useful answers? Or is everything purely random?
@Ed Papazian: That is actually how it works. I'm no expert on how you do that, but I asked one -- ChatGPT -- and here's what it said:
Creating a synthetic respondent for a consumer survey involves generating a simulated persona or dataset that mirrors the characteristics, preferences, and behaviors of a real consumer—without actually surveying a real individual. This can be useful for modeling, testing, hypothesis development, or scaling insights when real data is limited. Here's a step-by-step overview of how to do it:
1. Define the Purpose
Clarify why you're using synthetic respondents. Common use cases include:
Testing survey design
Modeling market segments
Simulating future scenarios
Running pre-research experiments
2. Build or Acquire a Foundational Dataset
You need a basis for realism. Sources can include:
Actual survey data
Public datasets (e.g., census, Pew Research, Nielsen)
CRM or digital behavior data
Market research reports
3. Identify Key Variables
Decide which attributes your synthetic respondent needs to emulate. These typically include:
Demographics: age, gender, income, education, location
Psychographics: values, attitudes, interests
Behavioral: purchase habits, media usage, brand preferences
Situational: employment, family status, device ownership
4. Generate the Respondents (Methods)
There are multiple approaches depending on the fidelity you want:
A. Rule-Based Generation
Use if you want control or simplicity.
Create profiles manually or with logic rules
E.g., "A 35-year-old urban mom likely shops online weekly and values convenience."
B. Statistical Modeling / Simulation
Use probabilistic sampling from real distributions.
Techniques: Monte Carlo simulation, Bayesian modeling, bootstrapping
Helps generate statistically valid synthetic populations
C. Machine Learning-Based Generation
Use when you have enough real data.
Clustering (e.g., K-means): to generate personas from real data clusters
Generative models (e.g., GANs, VAEs): to synthesize data matching real distributions
Can be refined with techniques like SMOTE for balancing attributes
D. Language Models (LLMs)
Use when you're simulating open-ended responses or behavior narratives.
Prompt an LLM (like GPT-4) with a persona and a question:
"You are a 42-year-old suburban dad who drives a hybrid and cares about sustainability. How do you decide which cereal to buy?"
This is useful for qualitative survey pre-testing or ethnographic simulation
Agreed, Joe.
Long ago, when I was at BBDO, I created a "model" that used fake--or "synthetic"--- groups of viewers and their demo characteristics plus TV viewing habits and used it to estimate the ratings for network TV shows for an upcoming season--including, of course, new programs with no rating track record. It was much simpler than what Dentsu is doing and much more focused on one aspect, but we got interesting results. Typically, we predicted about 65-70% of the new show ratings within a few share points of what actually happened--which was a little bit better than the human experts were doing. Then we refined the "model" and improved its performance slightly so that its predictions for TV's second season--that time in the first quarter when an additional slew of new programs replaced those that bombed out in the fall----and our "accuracy" level rose about 5 points on the newbies and, of course over 90% for those shows that weren't cancelled. It was an interesting experiment--but as upfront buying moved away from show-specific buys to multiple show bundles there was no need for our model as the rating ups and downs for a mass of TV show ad placements more or less cancelled out any errors.
@Ed Papazian: You were ahead of your times. I think the difference is it's not a model, but synthetic respondents responding as if they were actual people. I'm not sure it's necessarily better, but from everything I've ever heard of -- and reported on -- vis a vis human respondent bias, I'm not sure it's any worse either.