Synthetic users are AI-generated research participants. In practice, that means a large language model given a detailed persona, a specific stimulus such as an ad or a product concept, and a structured set of questions, then asked to answer as that persona would. The output looks like an interview transcript. It is not one. It is a model's plausible reconstruction of what such a person might say. Used with that framing, synthetic users are a genuinely useful early-stage research tool. Used without it, they are a fast way to feel confident about things you have not actually learned. In advertising, they belong to the category of pre-spend creative intelligence: tools that improve decisions before budget is committed, rather than measuring results after it is spent.
This article covers where synthetic users help, where they fail, how grounding and validation work, and the practices that separate useful synthetic research from expensive self-deception.
What synthetic users are, in plain English
A synthetic user is a role-played research respondent. You define who the respondent is: their job, budget, buying context, prior beliefs, and objections. You show them something concrete. You ask structured questions. The language model generates answers consistent with the persona and the stimulus.
Three terms are worth separating on first use:
- Synthetic users (sometimes synthetic consumers or synthetic respondents): AI personas answering research questions.
- Digital twins: personas built from a specific real individual's data. Most marketing use cases do not need this and should avoid the privacy implications.
- Scenario simulation: the honest description of what synthetic interviews actually are. The model simulates a scenario. It does not sample a population.
That last distinction carries most of the weight in this article. A real panel draws respondents from a population, so its results estimate something about that population. A synthetic panel draws responses from a model, so its results estimate what the model finds plausible. Those are different objects, and the entire discipline of using synthetic users well comes down to not confusing them.
Where synthetic users genuinely help
Speed. A synthetic interview round takes minutes. Recruiting, scheduling, and running a real qualitative round takes weeks. When a team needs a first read on Monday, not a validated finding in six weeks, synthetic users fill a gap that previously got filled by nothing at all, or by the loudest opinion in the room.
Cost. Real panels cost real money per respondent. Synthetic interviews cost close to nothing per run, which changes behavior: teams test more variants, ask more follow-ups, and explore stranger hypotheses because a failed run costs minutes, not budget.
Hypothesis generation. This is the strongest use case. Synthetic users are good at surfacing candidate objections, confusions, and reactions you had not considered. "Would a price-sensitive procurement lead read this claim as an unverifiable promise?" is exactly the kind of question a well-built persona can raise before any real customer sees the work. The synthetic answer does not tell you whether real procurement leads object. It tells you the objection is plausible enough to check.
Early exploration and question sharpening. Running structured synthetic interviews before real research improves the real research. Teams discover which questions are ambiguous, which stimuli are underspecified, and which segments react so differently that they need separate discussion guides. Synthetic rounds are a cheap rehearsal for expensive fieldwork.
Coverage of segments you struggle to recruit. Some segments are slow or costly to reach in small studies. A synthetic pass gives you a provisional read while you organize access to the real thing. Provisional is the operative word.
Where synthetic users fail
The failure modes are well documented across the research literature and practitioner reports, and they are structural, not bugs to be patched.
Novel products. Language models reconstruct plausible responses from patterns in training data. For a genuinely new category, there is no pattern to reconstruct. The model will still answer fluently, which is worse than refusing: you get confident-sounding reactions to something no real human has reacted to yet. The more novel the concept, the less a synthetic response is worth.
The distribution of real opinion. Real populations are messy. They contain contradictions, minority positions, and long tails of unexpected reactions. Model outputs tend toward the central, plausible, articulate answer. Ask 36 synthetic personas and you may get 36 well-reasoned variations on the consensus view, when the commercially decisive fact is that 15 percent of the real market holds an intense objection the model smooths away. Synthetic panels compress variance. Real markets are made of variance.
Emotional depth. A model can name an emotion. It does not have one. Reactions driven by identity, status anxiety, nostalgia, or humor that lands in the body rather than the intellect are precisely where synthetic responses are most articulate and least trustworthy.
Sampling validity. There is no sampling frame. A synthetic panel is not drawn from anything, so no statistic computed over it, no percentage, no ranking, no "7 out of 10 personas preferred B", inherits any inferential meaning about a real population. Counts of synthetic responses are a description of model behavior, not an estimate of market behavior.
Sycophancy and plausibility bias. Language models are trained to be helpful, and helpfulness leaks into role-play. Synthetic personas are systematically more polite, more coherent, and more persuadable than real buyers. They also prefer answers that sound plausible over answers that are true of any actual person. If your synthetic panel loves your concept, you have learned almost nothing. If it surfaces a specific, concrete objection, you have learned something worth checking.
How grounding and validation work
The difference between a toy and a tool is grounding on the way in and validation on the way out.
Grounding means constraining the simulation with real information:
- Real stimulus, not a description. Reactions to the actual ad differ from reactions to a summary of the ad. Show the model the real creative.
- Specific personas, not demographics. "Female, 34, urban" grounds nothing. Buying context, budget authority, current alternatives, and prior frustrations give the model something to be consistent with.
- Observable anchors. Where you can attach the simulation to measured properties of the stimulus, do it. If an ad's visual attention pattern or clarity has already been assessed, personas reacting to that assessed creative are constrained by evidence rather than free-associating.
- Structured elicitation. Fixed questions, fixed order, comprehension separated from emotion separated from action. Unstructured synthetic chat produces unstructured plausibility.
Validation means checking synthetic output against reality before trusting it:
- Compare synthetic reactions to past real results you already hold: previous campaign learnings, past panel research, support tickets, sales objections.
- Where the synthetic round shaped a decision, follow the decision into the real world. Did the objection the personas raised show up in real comments, real replies, real sales calls?
- Calibrate over time. Calibration, in plain English, is tracking how often a tool's signals turn out to match reality, and adjusting how much weight you give them accordingly. A team that has run twenty synthetic rounds and checked ten of them against outcomes knows what its synthetic signal is worth. A team on round one does not.
For a fuller treatment of how synthetic panels compare to the real thing, see synthetic consumers versus real consumer panels.
A worked example: the early hypothesis stage
Here is what a grounded synthetic workflow looks like in practice. In SaliencyLab, BuyerLens runs structured synthetic buyer interviews with 36 buyer personas, and it only opens from a scored RoastIQ result. The sequence matters. A marketer uploads an ad, RoastIQ scores it against five fixed KPIs and returns a verdict, and only then can BuyerLens interview personas about that specific scored creative. Every interview is anchored to a real stimulus with a known score profile, completes in under 2 minutes, and returns structured hypotheses about buyer resistance: what a segment might misread, what claim might trigger skepticism, where interest might stall before action.
The framing is explicit by design: BuyerLens is scenario simulation, not a consumer panel. Its 36 personas are not real people and its output is directional hypotheses about resistance, intended to be checked against real behavior, not quoted as customer evidence. The underlying creative scores it anchors to are model predictions, validated against public engagement and click-intent outcomes across 1,200+ ads with held-out Spearman correlations of +0.30–0.32 as of May 2026 (TikTok engagement n=700, TikTok CTR n=691, YouTube views n=403); they do not predict sales, ROAS, or brand recall, and the full validation is documented in the methodology.
That combination, a real scored stimulus plus structured interviews plus honest labeling, is the pattern to look for in any synthetic tool, whatever you end up using. A comparison of the current options is in best synthetic user platforms.
Best practices
These are recommendations drawn from the failure modes above, not universal facts.
- Define segments before generating personas. Persona quality is decided by your segment definitions, not by the model.
- Anchor every round to a concrete stimulus. A real ad, a real landing page, a real price. Never a paraphrase.
- Ask the same structured questions across personas and variants. Consistency is what makes synthetic comparisons meaningful at all.
- Read for patterns and objections, not quotes. A specific objection appearing across several personas is a hypothesis. A single eloquent quote is noise.
- Use comparative reads, not absolute ones. "Variant B drew fewer comprehension failures than A" is defensible. "Buyers love B" is not.
- Label synthetic output as synthetic everywhere it travels. In decks, in tickets, in stakeholder conversations. Undisclosed synthetic findings become fake customer evidence the moment they leave your desk.
- Close the loop. Every synthetic hypothesis that shaped a decision earns a real-world check: a live variant test, a set of real customer interviews, a look at actual campaign comments.
- Never let a synthetic round cancel planned real research. It can reshape it, sharpen it, or reprioritize it. It cannot replace it.
For how this looks specifically applied to ad creative, step by step, see the practical guide to synthetic users for ad testing.
Honest limitations
A synthetic-users article that does not state its limits is part of the problem, so, plainly:
- Synthetic users do not sample any population. No count, percentage, or ranking over synthetic personas estimates market behavior. This is the hard limit everything else follows from.
- They are weakest exactly where research is most valuable: novel concepts, emotionally driven categories, and minority reactions with outsized commercial impact.
- They inherit the biases of their training data, including underrepresentation of some markets, languages, and buying cultures. A persona from a poorly represented segment is a thinner simulation and gives no warning that it is.
- They are systematically agreeable. Positive synthetic reactions carry very little information. The signal, such as it is, lives in specific, repeated objections.
- Validation evidence for the field is early and mixed. Published results range from encouraging alignment on structured tasks to poor alignment on open-ended and subgroup questions. Any vendor claiming a general accuracy figure for synthetic respondents is ahead of the evidence.
- They do not replace real panels, real interviews, or live tests. Real customers and live campaign data remain the ground truth. Synthetic users are a way to arrive at that ground truth with sharper questions and fewer wasted rounds.
The bottom line
Synthetic users are a hypothesis engine, not an oracle. Their value is concentrated at the start of the research process: exploring cheaply, surfacing objections, rehearsing real fieldwork, and comparing variants before anything expensive happens. Their danger is concentrated in the temptation to treat fluent output as evidence. Ground every round in a real stimulus, keep the questions structured, label the output honestly, and validate against reality before the findings shape spend.
If you want to see the grounded version of this workflow on one of your own ads, upload it to SaliencyLab, get the scored RoastIQ verdict, and open a BuyerLens interview round from the result. The hypotheses arrive in minutes. What you do to verify them is still the research.
Frequently asked questions
What are synthetic users? Synthetic users are AI personas, built on large language models, that answer research questions as a defined buyer segment plausibly would. They simulate interview responses; they are not real respondents and do not represent a sampled population.
Are synthetic users accurate? Sometimes, for structured comparative tasks, and the published evidence is early and mixed. They are least reliable for novel products, emotional reactions, and minority opinions. Treat outputs as hypotheses and validate against real customer behavior before acting on them.
What are synthetic users best used for? Early exploration: generating candidate objections, comparing variants, rehearsing discussion guides, and sharpening questions before real research. They are strongest where speed matters and the cost of a wrong hypothesis is low.
Can synthetic users replace real consumer panels? No. Synthetic panels have no sampling frame, so they cannot estimate what a market actually thinks. They complement real panels by making the expensive rounds better targeted. See the full comparison at /blog/synthetic-consumers-vs-real-consumer-panels.
What is grounding in synthetic user research? Grounding means constraining the simulation with real information: a real stimulus rather than a description, specific persona detail rather than demographics, and, where possible, measured properties of the creative being tested.
How do you validate synthetic research findings? Check synthetic hypotheses against evidence you already hold (past research, sales objections, campaign comments), then follow decisions into the real world with live tests or real interviews. Track over time how often synthetic signals matched reality.
What is sycophancy bias in synthetic users? Language models are trained to be agreeable, and role-played personas inherit that. Synthetic respondents are more polite and persuadable than real buyers, so positive synthetic reactions carry little information. Specific repeated objections are the more trustworthy signal.
Is BuyerLens a consumer panel? No. BuyerLens is scenario simulation: structured synthetic interviews with 36 buyer personas, anchored to an ad already scored by RoastIQ. Its output is directional hypotheses about buyer resistance, not statements from real people.