The future is synthetic. Here's a plain-English guide for independent agencies: what it is, which kind to trust with which decision, and how to put it to work.
Arima's team supplied the framework, the validation research and the checklist at the heart of this Blueprint.
Read this firstThis Blueprint is for educational and illustrative purposes only. It is not legal, financial, technical or purchasing advice, and it is no substitute for your own due diligence. Arima, an IAN partner, sells synthetic data products and contributed the three-kind framework, the checklist and a summary of its validation study with a third-party research firm, which is an abstract of a longer paper Arima plans to release. This Blueprint is educational, and Indie Agency News does not recommend any vendor. Views belong to the people and companies named, and inclusion of any company implies no endorsement by Indie Agency News.
Imagine putting a research question to an entire market without fielding a survey. That is the promise of synthetic data.
Synthetic data is changing how research gets done. Think of a synthetic population as a population simulator, or a marketing sandbox: a place to test an idea on a market before you spend money on the real one.
Arima's team uses an analogy that keeps both the promise and the limits in view.
"Think of it like a flight simulator. The early ones had trees that looked like broccoli. We're well past that, but it isn't hyper-realistic yet. A simulator lets you train faster and more often, but you still have to fly a real plane to get your license."
That's the frame for this Blueprint. Use the simulator to practice and move faster, then check it against the real thing before the decisions that count.
Arima's Chief Innovation Officer, Sabaa Quao, takes the analogy a step further. Fighter pilots go back to the simulator a couple of times a week to rehearse the emergencies that are too dangerous or too expensive to stage in a real aircraft, like an engine failure. A marketing sandbox works the same way. You can run the what-ifs, including three things going wrong at once, before any of them happen for real.
The respondents are realistic stand-ins for people and markets, generated from real data sources using statistics. They come with demographics, locations and habits. They answer fast, and none of them will ever file a privacy complaint. They don't copy any living person. They reproduce the patterns.
The appeal is easy to see. Surveys are getting harder to run. Privacy rules are tighter, hard-to-reach audiences answer less often, bots pollute the results and the people who do respond are tired of being asked. Traditional research is also slow. By the time the findings are gathered, analyzed and acted on, the market may have moved.
The words "synthetic data" are covering three very different things: personas, panels and populations. A persona works like a focus group, a panel like a research panel and a population like a census. They share a name, and they are built for very different decisions. Ask a persona to size a market and you'll get a number it was never built to give.
Our expert partner on this Blueprint is Arima, which builds synthetic populations for a living. The spine of the report is a test Arima ran with an established third-party research firm, putting a synthetic population up against a human panel and recording where it matched and where it missed.
The misses are the most useful part. They tell you where to trust synthetic data, where to check it and how to ask it better questions.
Used well, synthetic data puts research behind your pitches, makes you the adviser when clients get pitched it and can open a new service. Used without care, it can put a number in front of a client that the data can't back up. This guide is about getting it right.

Arima is one platform across the full marketing workflow: research, strategy, optimization, measurement and attribution, with no cookies, no PII and no data-science team required. Its recognitions include Digiday's Best Marketing Analytics Platform and Gartner's Coolest Innovation in Synthetic Data. Exclusively for IAN members, Arima partners with agencies on pitches to help them win business. Its team supplied the three-kind framework, the validation research and the eight-question checklist at the core of this guide.
Talk to Arima →The first thing anyone asks about synthetic data is the right thing to ask: how do you know it's accurate?
Arima's team hears it more than any other question, and the honest answer starts with an awkward admission. No complete outside dataset exists to check a synthetic population against. Nobody is holding the answer key.
So Arima runs two checks. The first is housekeeping. The synthetic population is continuously rebalanced against the source data it was built from. If the source says one share of men wear glasses and a different share of women do, the synthetic version has to land on those same shares every time.
The second is a stress test, and it's the heart of this section. Arima went to an established third-party research firm and set up a side-by-side.
A large, representative online panel of real people, surveyed the usual way.
Built by Arima and asked the same questionnaire through an AI polling tool.
The two sets of answers were compared on three levels. The formal names are in the small type. The plain versions are worth memorizing.
The third one is the keeper. You're rarely paying for a number. You're paying for a decision, and two datasets can sit a few points apart and still send you to the same place.
The company that built the synthetic population is not the right judge of the result. The read that matters came from the statistics team at the firm running the real panel, and its conclusion was that the two lined up in most instances.
"Most instances" is the honest part of that verdict. A claim of a perfect match would deserve less of your trust. Section 04 covers the rest: the five conditions where the two crowds stopped agreeing.
The scores. This section draws on an abstract of a longer paper Arima plans to release. The question-by-question results aren't in the abstract, and the list of 20 questions wasn't included with it. When the full study is published, both belong on this page.
Strip away the vocabulary and the method is simple enough for any agency to borrow. Take questions you can put to real people. Put the same questions to the synthetic version. Compare the headline, the subgroups and the decision. Then write down where they split.
You don't need a giant research panel to try a small version. If you already hold survey results for a client, you have a benchmark. Run the same questions through the synthetic tool and see how close it lands on a subject you know cold.
Before you use synthetic data, know which of three kinds you're working with.
Arima's team explains each one by its cousin in old-fashioned research. A persona works like a focus group, a panel like a research panel and a population like a census. They are not interchangeable, and the trouble starts when one is asked to do another's job.
Exploring what customers want, how they behave and what they prefer. Ideas, messaging and early concepts.
Forecasting or sizing a market. It isn't statistically representative.
Segment-level analysis and directional testing. Comparing audiences and choosing targets, with bigger samples than personas.
Very small or niche audiences. The makeup of the panel also has to be managed over time.
Market sizing, predictive modeling, location intelligence, business cases and simulation. Broad, granular by geography and privacy-safe.
It takes sophisticated modeling across many variables. Niche behaviors are only as good as the source data.
The sorting rule is the size of the bet. If you're exploring an idea, testing a concept or looking for a rough steer, personas and panels are fast and cheap, and they help you sharpen a strategy before real money moves.
When the decision carries financial or strategic weight, the bar goes up. A population represents the full market instead of a slice of it. That supports the heavier work: statistically sound analysis, fine geographic detail and simulations that smaller synthetic datasets can't carry.
The more money riding on the answer, the closer to a full population you want to be.
| If the job is to… | Reach for |
|---|---|
| Find fresh angles, explore motivations or draft messaging | Persona |
| Get a quick read on a concept before you commit resources | Persona or panel |
| Compare audience segments or choose a target | Panel |
| Size a market or forecast demand | Population |
| Build a propensity model that predicts who is likely to act | Population |
| Judge a media investment or build a business case | Population |
| Plan a retail network or read a hyper-local market | Population |
| Reach a niche audience too small or too costly to survey | Population |
The use cases run across the whole workflow, from strategy to execution and from understanding an audience to measuring what worked. In practice, that means a team can do four jobs in one place:
Two further uses are aimed at bigger bets. One is rehearsal: test what might happen if millions of dollars move into a channel before a dollar moves. The other is a sandbox, a safe place to watch how an autonomous AI campaign behaves before it meets a real consumer.
Knowing where a tool is weak is as useful as knowing where it's strong. Arima's side-by-side turned up five conditions where the synthetic answers pulled away from the human ones.
Arima shared the misses along with the matches, which is what makes the test so useful. Each one comes with a workaround.
The models were more accurate on the past and present than on projections, and the gap widened the further out the question reached. Answers about financial health today versus six months ago matched the human panel more closely than answers about financial health six months from now.
The moveAsk about what is and what was. Treat anything forward-looking as a first read.
Subgroup analysis was weaker unless the subgroups were built into the question from the start. Asking how many people wear glasses and then filtering by gender worked less well than asking about men and women directly.
The moveDecide which cuts you need before you ask, and write them into the question.
The more detailed the answer choices, the wider the gap. The two methods diverged more on 10-point scales than on 5-point scales. The same thing happened with narrow numeric ranges, such as 1% to 2% blocks, compared with a broader 1% to 5% range.
The moveKeep the answer choices coarse. Five options beat 10.
Real people shrug. Simulated respondents don't. The polling tool in the test currently prompts a definite answer from every synthetic respondent and doesn't account for non-response, so questions with a "Don't know" option skewed the comparison with the human panel.
The moveFlag any question where a real person might reasonably have no opinion, and read those results with care.
Large language models are more mature in English, and the polling tool in the test translates non-English questions before running them. Questions asked in a language other than English showed wider swings.
The moveIf your audience doesn't live in English, test before you trust.
| Instead of | Ask it this way |
|---|---|
| "How will your finances look in six months?" | "How are your finances today compared with six months ago?" |
| Asking everyone, then filtering by gender | Asking about men and about women directly |
| A 10-point scale | A 5-point scale |
Go back to the admission at the top of the test: nobody is holding the answer key. Arima says plainly that no comprehensive outside dataset exists to benchmark a synthetic population against. Validation is a comparison between two imperfect instruments, a model and a survey.
Assume that's true of any synthetic data you use until someone shows you otherwise. The useful question shifts from "is it accurate?" to "accurate compared with what, on which kinds of questions, and checked by whom?"
Five known weak spots are worth more than a match rate. They're a map of where to be careful.
The three kinds and the five weak spots do most of the work. Here's how they fit together on a real brief.
Arima's team, which builds full synthetic populations, points to four things a population can do that a handful of personas can't.
The trade-offs are real. A population takes sophisticated modeling, niche behaviors are only as good as the source data, and for exploring ideas a persona or panel is faster and cheaper. Use the kind the decision needs.
For a smaller shop, the case fits in a sentence: synthetic data can level the competitive playing field.
The case is practical. Synthetic data promises faster insight, lower research costs, more flexibility and more detail than traditional data. For a shop without a research team down the hall, that changes what you can bring to a pitch. You can test a message, size an audience or model a media plan before the meeting, and show your work.
It won't stay new for long. Gartner estimates that by 2028, 60% of the world's top 100 retailers will use synthetic population data to validate marketing strategies, up from fewer than 20% today (figures as cited by Arima). The more abstract the data gets, the more it rewards creativity upstream, which is where indies do their best work.
The sensible first move is modest. Dip a toe in on a small brief. Find out where synthetic data adds value in your workflow and where it doesn't, while the stakes of learning are still low and before a client asks what you think of it.
Put the sections together and one idea runs through all of them. The technology isn't the hard part. The matching is. A persona pressed into service as a population will fail, and so will a population asked about a future it can't see. Every warning in this Blueprint is a version of the same instruction: fit the tool to the question.
Know which kind you're using. Ask where it breaks. Let the size of the decision set the standard of proof.
Run these before synthetic data goes in front of a client, whether it's a provider's, a client's or your own. The checklist comes from Arima's team. We added the reason behind each question.
Everything else depends on this. If nobody can answer in one word, keep asking.
Synthetic data can't be better than what it was built from. You want names and dates.
The first half tells you what you can claim with the output. The second tells you whether you'll use it.
A good answer covers three kinds of agreement and a list of misses.
Inferred isn't bad. Unlabeled is. You need to know which numbers were measured.
Granularity is a population's big strength. Know where it runs out.
This is the building-block test. It decides whether you end up with something proprietary.
The most useful question on the list. A specific answer tells you exactly where to be careful.
The companion Synthetic Data Planner turns this Blueprint into a one-page plan for your project: which kind to use, what to watch for, what to ask your provider and what to tell your client. It takes about a minute.
Open the planner →Arima, our expert partner on this Blueprint, builds synthetic populations and runs validation tests like the one in section 02. For help choosing the right kind, setting up a side-by-side test or bringing synthetic data into a client project, Indie Agency News can introduce you.
hello@indieagency.news
Six moves. Pick two and start this week.
Build a plan for your project in the Synthetic Data Planner, then send it to Arima's team for a second look before it reaches a client. They'll tell you whether you're tracking: the right kind for the job, the questions likely to drift and the best way to test the results. If it's headed into a pitch, they can work alongside you on it.
Email us with "Arima second look" in the subject, and we'll pass it to Arima's team.
Expert partner: Arima. Arima's team contributed the subject-matter expertise behind this Blueprint: the three-kind framework, the validation research and the eight-question checklist.
Validation study by Arima with a third-party research firm.
Edited and produced by Indie Agency News.
Indie Agency News is the global home for independent agencies. Our mission is to empower indies with visibility, resources and community, creating more opportunities for agencies and more choices for marketers.
Join us: indiehq.is/join · hello@indieagency.news
© 2026 Indie Agency News. The Blueprint Series. This Blueprint is for educational and illustrative purposes only and is not legal, financial, technical or purchasing advice. See the full note at the top of this Blueprint.