Articles in this section

Resonate Digital Twins: FAQ

Avatar
Mallory Mast
  • Updated
Follow

General Questions:

License and Packaging:

Delivery:

Have more questions? Contact your Customer Success Manager or resonatesupport@resonate.com. 

 

What are Digital Twins and Synthetic Data?

A digital twin is an AI representation of a real person, built to predict how that person thinks, feels, and responds. You'll also see this called a synthetic persona, synthetic audience, AI twin, or digital twin of a customer (DTOC). We survey digital twins the way you'd survey real people: we ask a question, and each twin answers in its own words, just as the person it represents would.

What is a question? A question is one thing we ask digital twins about. We survey them similar to how we’d survey a real person to generate responses like a digital focus group.

Synthetic data is the output a digital twin’s question responses produce. When you ask a digital twin a question, its answer becomes a synthetic data point, sometimes called a synthetic attribute or augmented data, that you can use like any other attribute we have in Ignite or Cortex.

How Resonate Digital Twins are different: Resonate Digital Twins starts with real people and is the only solution that converts answers into targetable audience lists you can act on in about five days. Our digital twins are grounded in an identity graph of 250M ID-matched people, more than 100 million human responses to our USCS, 30 billion daily behaviors and 15K+ attributes for every US adult and we calibrate them to your own data and insights. We tailor them to your audience, survey them, validate every answer and scale the results across the US adult population.

The result is a targetable, activatable audience grounded in real, validated behavior.

 

What can Resonate Digital Twins be Used For?

Resonate Digital Twins supports use cases across the buying cycle, from winning new audiences to growing and protecting the relationships you already have:

Forecast scenarios: Predict how a market will respond to a message, offer or scenario before you act. Data from Resonate’s Digital Twins was used to run a $5M dollar campaign in the NY-12 congressional primary, in which the Resonate candidate outperformed projections by more than 5%.

Guide campaign decisions early: Preview how an audience is likely to respond before you finalize the plan or launch a product, so you adjust based on a real signal instead of a guess. 

Explore new products, markets or segments: Understand a market or segment's preferences before you invest in it. 

Reach hard-to-recruit audiences: Surface niche segments that a traditional panel can't reach quickly or affordably. In a health audience prototype, we're identifying adults likely to have undiagnosed type 2 diabetes, people that claims data and panels can't find.

Test before you launch: Test a new offer, price or message before committing spend. 

Understand the why: Mine open-ended responses to understand not just what your audience thinks, but why. 

 

How accurate are the results?

Synthetic research is growing fast, but trust hasn't caught up, and that skepticism is fair. No one's synthetic data is a perfect stand-in for real people, including ours.

What we can say: when we've measured Resonate Digital Twins head-to-head success against real data, we’ve had an average of 93% accuracy. And when back checked against our own real studies, it's landed within about 5 to 10 points.

And we’ve seen measurable success in the results:

  • We’ve successfully forecast multiple elections, where we’ve correctly called the outcome.
  • Data from Resonate’s Digital Twins was used to run a $5M dollar campaign in the NY-12 congressional primary, in which the Resonate candidate outperformed projections by more than 5%.

That's a solid track record, not a guarantee.

 

What is your validation process?

The market is right to be skeptical of synthetic data. So instead of asking for trust, we hold every build to a standard you can check.

It starts with the data. As Forrester analyst Jay Pattisall put it, "Synthetic audiences are only as good as the data that compiles them." Our digital twins are grounded in more than 250 million ID-matched US adults, more than 100 million US Consumer Study responses, more than 30 billion daily observed behaviors and more than 15,000 attributes per person.

Why validate against evidence the model never saw? Validation works like checking a weather model against last week's weather, without letting it see last week's weather first. For every build, we hold back real answers we already have, from our fielded studies, a customer's research or public benchmarks. Then we ask the digital twins the same questions. If the results hold up on questions we can check, you can trust them on new questions that fielded research would take months and six figures to answer.

The standard every build meets:

  • Written first. We write down benchmarks and pass/fail rules before the model is fit, so the goalposts can't move.
  • Held out. Every build holds back evidence the model never saw. It ships only if it beats that holdout.
  • Made to be checked. Every prediction is locked and timestamped when it ships, so it can be tested against whatever comes later: a lift test, a past campaign or your own survey. We report the result either way.
  • Traceable. Every score traces back to the individual and the evidence behind it. It's not a black box.

How we counter the known weak points of AI-generated data:

  • AI models tend to tell you what you want to hear. We ground every twin in real, measured attributes instead of letting the model improvise. The evidence is applied to the twins' answers from outside the AI, so the model can't argue its way past the data.
  • Question wording and order can skew results, just as in a live survey. We rotate option order, test wording variations and interview each twin more than once.
  • Real panels carry their own bias. We project results across every adult in Resonate's national database and weight to Census demographics.

Where we're careful: Synthetic research works best on established categories and on questions about values and preferences. We're more cautious with brand-new products, awareness questions and fast-breaking news, and we'll tell you when a question is a better fit for fielded research.

Validation process and results:

To test how well our synthetic pipeline performs against survey research topics it has never seen, we used Resonate’s vast survey research library as a ground truth, and asked our pipeline to predict how people responded to those questions.

The test set is 35 questions chosen from Resonate’s US Consumer Study. We chose 25 spanning beliefs, values, behaviors, media, purchase intent, and brands, and chose an additional 10 at random from the catalog. We scored every respondent as if the question had never been asked.

Across all 35 questions, our average similarity to fielded research was 93%, ranging from 84.3% on a question about the importance of cultural heritage, to 99.6% on customers for a major insurer. Top-decile lift, a measure of how much more likely the highest-scored tenth of an audience is to hold the attribute than a random adult, ranged from 1.2x on live-tv streaming to 4.9x on mortgage intent, averaging 2.8x. Cohen’s kappa, a measure of inter-rater reliability, ranged from 0.1 to 0.52. Area under the ROC curve (AUC), a measure of how well statistical models order their target populations, ranged from 0.59 to 0.88. In all cases, test targets beat robust controls, including a supervised demographics-only model, as well as a “LLM planner” mode that resembles what a sophisticated user could get from Claude or ChatGPT.  

 

What should I consider when evaluating Synthetic Data or Digital Twin providers? 

Not all synthetic data is built the same way, and the differences matter. Here's what to ask any vendor, including us:

  • What's actually grounding the responses? Some tools generate answers from a general AI model with no real person behind them. Others are trained on a research panel's past responses. Ask exactly what real data the synthetic responses trace back to, and how individual-level that grounding really is
  • Can they show you validation, not just tell you about it? A credible vendor should have a documented process for checking synthetic results against real research, and should be willing to back-test against your own holdout data rather than asking you to take their word for it
  • How do they handle known AI and research pitfalls? Ask specifically how they address AI models defaulting to agreeable answers, how question wording and order get managed, and how they correct for bias instead of just claiming they do
  • Is the audience representative by construction, or adjusted after the fact? Panel-based tools inherit whatever skew exists in who joined and stayed active in that panel. Ask how the vendor corrects for that, and whether their numbers are calibrated to real population benchmarks from the start
  • Does the output resolve to something you can activate, or does it stop at insight? Many synthetic data tools hand you a report or a chart. Ask whether the result can become a targetable, addressable audience, or whether you'll need a separate step, and a separate vendor, to act on it
  • Are they honest about what it's not good for? Any vendor claiming synthetic data works for every use case, including breaking news or genuinely new-to-world questions, should raise a flag. The credible answer names real limits

 

How is this different from asking ChatGPT?

Ask ChatGPT to imagine how a rural Pennsylvania gun owner feels about background checks, and it generates a plausible-sounding answer from general patterns in its training data. There's no real person behind that answer, and no way to trace the response back to anything concrete. It's a guess dressed up as an opinion.

Panel-trained synthetic data is a step up. It's grounded in a real survey panel’s response, so it reflects patterns from actual people who were surveyed. But it's still bounded by whatever that panel looked like and doesn’t always reflect the true marketplace. If the underlying panel has coverage gaps, fraud, and quality problems that plague traditional research, the synthetic layer inherits them.

 

Here's how Resonate Digital Twins are different:

  • Grounded in real people's responses, not averages or invented profiles: Each digital twin is grounded in more than 250 million ID-matched US adults, more than 100 million US Consumer Study responses, more than 30 billion daily observed behaviors and more than 15,000 attributes per person, not a demographic label or an invented profile.
  • Individual, not aggregate: A chatbot or public dataset can only hand you an aggregate guess. Resonate Digital Twins predict a response for every individual in the population.
  • Corrected for AI bias, anchored to real research: General AI models carry measurable bias when asked to role-play a person with different views. We correct for that and anchor results to existing responses, so they don't contradict your own prior research.
  • Representative from the start: Every audience is built to match census demographics from the outset, not adjusted after the fact the way panel-based tools often are.
  • Validated, not just plausible: Resonate’s synthetic pipeline has been tested on numerous survey response distributions, including ones from our United States Consumer Study, Pew research, and others. It produces results with 93% similarity to published research. 
  • Built to activate: The result isn't just an insight you read and file away. It resolves to a real, addressable audience you can put to work.

How is this different from focus groups or traditional fielded research?

Focus groups and fielded research, including Resonate's own custom attribute models, give you a real human baseline. But they can take weeks to six-plus months. Resonate Digital Twins works differently:

  • Faster: Results in days, not weeks or months.
  • Scaled: Results extend to every U.S. adult in Resonate's database, not a small sample.
  • The why, on demand: Every twin answers in its own words, so you can revisit the reasoning anytime, like a focus group you can reopen.
  • Activatable: You get a targetable audience, not just a report.

It complements fielded research rather than replacing it. Use it for a fast, directional read, or to test a question before you field a full study. For well-understood, repeatable work like pitch and RFP research, it can stand in for fielded research. For categories that are new to the world, or for breaking news, a real human baseline is still the better choice.

 

If a Digital Twin represents a real person, why can’t I just survey or ask a question of that real person the Digital Twin represents?

There are times  when traditional primary research is the best or even the only answer, but that is costly and time-consuming and must be part of the decision process.

For privacy reasons, the actual identity of any of the respondents to Resonate’s surveys are kept confidential and not shared, therefore there is no mechanism to create a specific panel of real people and make them available over time to respond to questions.

 

What counts as a question?

A question is one thing we ask respondents (in this case Digital Twins) about. It counts as one question no matter how many answer choices it has, whether that's a 5-point scale or a 'select all that apply' list. If we ask about multiple things, like rating five different banks, each one counts as its own question."

Examples:

  • "How likely are you to bank with Wells Fargo?" with five scale options = 1 question
  • "Which of these banks have you heard of? Select all that apply." = 1 question
  • "How likely are you to bank with each of these 5 banks?" = 5 questions

If I license a subscription and have up to 4 digital twin panels and 20 questions a month to ask, can I ask a different set of questions every month?

It all depends on your goals and what you’re trying to accomplish.  Technically, you can ask a different 20 questions every month as long as you’re not activating a given audience that needs to refresh.

What this basically means is that you can ask the 20 questions in any combination of the four panels.  One month, you could ask 10 questions of 1 panel and 5 questions each of two other panels.  The next month, you could ask 5 questions of each panel.

The only restriction here is that if you are activating an audience, because it needs to refresh on a regular basis, the attributes and therefore the questions you are asking of a panel must still be asked as long as the campaign is activating and Resonate is sending refreshes.

 

I’m an agency, if I get access to up to 4 Digital Twin Panels as part of my subscription license, can I spread the panels over multiple clients?

 

Yes – You can have us build one panel for each of 4 clients or 2 panels each for 2 clients or 4 panels for one client.  However, once a panel is built, it cannot be swapped for a different client.  That would require either using 1 of your unused panels, if you have any, or licensing another panel.

 

What gets delivered?

Some market providers create and deliver digital twins you can interact and chat with— that is not the case for our solution. We are not delivering Digital Twins in a self-serve way for you to interact with directly. However, we are building and asking these twins your questions either a single time or a subscription with a set monthly number of questions you can ask of your Digital Twin panel.

There are two ways to get fast answers you need:

  1. Get an Audience – Ask something we haven’t fielded. We model and scale to the full population so there’s a response associated with every individual. We deliver an audience you can put media against right away via Ignite or Cortex or as a managed service.
  2. Get Insights – Ask something we haven’t fielded. This doesn’t get modeled and scaled to the full population but delivers the answers and responses as a report. It works like an AI-powered focus group: a fast, directional read. Insights are delivered via a summary report that, without activation involved.

This is also not a self-service functionality like it is elsewhere. We manage the customer process similarly to our other DaaS products. Accordingly, there is no fixed package, and the outputs for delivery will vary based on customer need, learn more about the delivery process and outputs in this article. 

 

How long does it take?

Resonate Digital Twins delivers validated, individual-level synthetic attributes in about 5 days. Many research firms can take a few weeks and up to a few months to turn results.

 

What is the refresh schedule? 

Refreshes are six times a year aligned to wave releases. Each refresh delivers a complete, full-replacement set of IDs in to your audience. The audiences are structurally identical to your prior delivery. The selected IDs reflect the most recent wave of Resonate consumer data.

IDs may be added to or removed from the audience between waves as the underlying identity graph evolves. Predictions for the same ID may shift between waves to reflect updated signal. Your downstream pipelines should treat each refresh as a full reload of the audience, not an incremental update against the prior audience.

Some variation between refreshes is expected and reflects the underlying signal — not a problem with the delivery. Here's what's normal:

  • ID counts change between waves. The Resonate consumer universe isn't static. IDs are added as new people are detected in the identity graph; IDs are removed as records age out or no longer meet inclusion criteria. Modest wave-over-wave changes in total count are normal.
  • Predictions for the same ID may shift. Each refresh reflects the most recent wave of survey data and updated modeling. A customer scored as "Frequent Traveler = TRUE" in one wave may be scored differently in the next as the underlying behavioral and motivational signal evolves. This is the value of refreshing — your predictions stay current with consumer reality rather than reflecting a static snapshot.

How are these delivered to Ignite?

These data/audiences will arrive in Ignite similarly to how Flash audiences do.  They will show up in the Taxonomy under Custom Data. They are audiences that you can Look-alike, deliver to a DSP, or add to a campaign in Cortex.

Here’s where it can get a little complicated: These technically exist as attributes within Ignite, but you cannot use them to insight on an existing audience. For instance, if your synthetic audience is “left-handed Coke Zero drinkers” you can insight and better understand that synthetic audience, but you can’t understand other non-synthetic audiences makeup with synthetic attributes, for instance if I have an audience for “Millennial moms in Chicago” I can’t insight on that audience to understand how many of those are “left-handed Coke Zero Drinkers.”

 

How do I access these in Cortex?

Similar to custom attribute modeled data (SPARK data) in Cortex, functionally, yes – they can.

But you first have to build the audience in Ignite. Then, in Cortex, you’ll need to assign the audience you built in Ignite to a campaign. That will enable you to run dynamic segmentation, LAM and activate on the audience in Cortex.


Have more questions? Contact your Customer Success Manager or resonatesupport@resonate.com. 

 

Related articles

Comments

0 comments

Article is closed for comments.