Short answer: “synthetic” isn’t one thing. It’s at least five distinct products sold under a single word, and they’re good for completely different jobs. The confusion is not an accident — it lets a thin product borrow the credibility of a rigorous one. Before you evaluate any vendor, work out which of these they’re actually selling you.
Here are the five, in plain terms, with what each is for and where it falls down.
1. AI-generated personas
What it is: a simulated client you can talk to — an AI configured to respond as, say, a general counsel, which you can question and follow up with.
What it’s for: exploration and reasoning. This is the one most people mean by “synthetic research”, and it’s the right tool for understanding how a category of client thinks and why they see things the way they do.
The catch: quality depends entirely on how the persona is built. A persona grounded in real decision-making data behaves very differently from one conjured from a job title and a couple of demographics. Same label, wildly different reliability.
2. Synthetically-derived insights
What it is: aggregated findings produced by AI — a summary of what a simulated audience “thinks”, delivered as conclusions rather than as individual responses.
What it’s for: a fast directional read on a broad question.
The catch: you get the summary, not the respondents. You can’t interrogate an individual answer, can’t slice by segment, and can’t check the reasoning underneath. If the aggregation is wrong, you have no way to see it.
3. Simulated respondent-level data
What it is: a complete, AI-generated survey dataset — structured exactly like the output of a traditional human panel, one row per “respondent”.
What it’s for: slotting into quantitative workflows that expect panel-shaped data.
The catch: it looks like a panel and isn’t one. It inherits every weakness of AI-generated numbers — including the well-documented unreliability of model-produced ratings — while presenting them in the reassuring format of a real dataset. Clean structure is not the same as trustworthy data.
4. Digital twins
What it is: a model of a specific, named real person, built from that individual’s own data — often a long interview with them.
What it’s for: the highest-fidelity simulation available, because it’s grounded in a real human rather than a stereotype. The evidence backs this: in the largest study of its kind (Park et al., LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals, Stanford and Google DeepMind; arXiv, version 3, June 2026), agents built from a real two-hour interview with a person matched that person’s General Social Survey answers at 83% of the consistency the person showed with themselves two weeks later, against 74% for agents built from demographics alone.
The catch: you need the real person’s data to build one. Digital twins model people you already have access to — they don’t conjure a representative client you’ve never met. For most firm-level positioning questions, that’s a different job from the one you’re trying to do.
5. Simulated conversations and interviews
What it is: AI-moderated or AI-participant interviews and focus-group-style sessions — a conversational format rather than a survey.
What it’s for: qualitative texture, and workflows that are built around interviews rather than questionnaires.
The catch: the value hinges on details the label hides — whether the participants are real people being AI-interviewed or synthetic participants being simulated, and how good the moderation is. Two products described the same way can be completely different things.
Why the ambiguity is the vendor’s friend
All five get sold as “synthetic research”, and that’s exactly why the word is dangerous. A vendor offering aggregated insights off a general-purpose model can stand next to a vendor building interview-grounded personas, use identical language, and let you assume they’re the same product at very different prices. The vagueness does the selling.
It also makes fair comparison impossible. You can’t judge whether a price is reasonable, or whether a method fits your question, until you know which of the five you’re looking at. Two quotes for “synthetic research” might be for two products that share nothing but the adjective.
The one question that cuts through
Ask every vendor, early: “Which of these are you actually selling me — personas I can interrogate, aggregated insights, a respondent-level dataset, digital twins of real people, or simulated interviews? And why is that the right one for my question?”
A serious provider answers immediately and can tell you why their format fits your problem. A vendor who blurs the categories, or bristles at the question, has told you something useful before you’ve seen a single output.
Which one does a firm usually need?
For the questions professional services firms actually care about — why we lose pitches we should win, what our positioning means to a client, how a general counsel chooses between us and a rival — the useful end of this list is usually AI-generated personas and simulated conversations you can interrogate: reasoning you can follow up on and push back against, across a range of client types. Aggregated insights are too shallow to act on, and respondent-level datasets give you tidy numbers without the “why”. And wherever personas are involved, how they’re grounded matters more than anything else on the vendor’s slide.
Knowing which product you’re being sold is the first of the questions worth putting to any vendor. For the full set, see the companion guide, how professional services firms should evaluate synthetic research.
Taleist builds decision intelligence for professional services firms — interrogable reasoning about how your clients actually choose, not a dataset that looks tidy and answers the wrong question.