Synthetic panel:what it is, and where it stops

    A synthetic panel is a simulated cross-section of a population. No real respondents answer, and no profile in it corresponds to a living person. What makes one usable rather than decorative is entirely a question of how it is weighted — and that is most of what this page is about.

    The definition

    In conventional research you find people, ask them something, and count the answers. In a synthetic panel the last two steps happen inside a language model: a question is put to a set of described profiles, and each profile produces a response in character.

    The profiles are not people. They are combinations of attributes — age band, education, household income, region, and a value orientation — assembled so that the combinations occur in the panel at roughly the rate they occur in the population.

    So the output is not a measurement. It is a structured prediction: how a cross-section of this country would plausibly react, and which argument each part of it would reach for. That is a genuinely useful thing to have before a decision, and it is a different thing from what people actually said. Holding on to that distinction is the whole discipline of using one well.

    The US panel, as the worked example

    Reynard's American panel is built in two layers. The first is the demographic distribution taken from U.S. Census Bureau data: age, educational attainment, household income band and geographic spread, in the proportions they actually occur. That layer is what stops the panel from consisting of whoever is easiest to imagine.

    The second layer is eight value segments. It groups people by what they want and what they distrust rather than by what they earn or how old they are. This layer does the work demographics cannot: two 42-year-olds in the same income band, in the same state, can react in opposite directions to the same offer, and the reason is almost never their age.

    The Dutch panel is built the same way, on CBS (Statistics Netherlands) population data plus eight value groups. It came first, which is why it appears throughout the site as the proof market, but the structure is identical in both.

    Value-based segmentation of the Dutch population was popularised by Motivaction with its Mentality model. Reynard's grouping is built on publicly documented descriptions of Dutch value orientations and is not the Mentality assignment itself, which requires Motivaction's own instrument and database. Reynard is not affiliated with, licensed by or endorsed by Motivaction.

    Why weighting changes the answer

    An unweighted synthetic panel gives you an average of the profiles that happened to be generated. If the generation over-represents the kind of person a model finds easy to write — engaged, articulate, broadly urban and mildly progressive — then the average is an average of that, and it will read as plausible while being wrong in a consistent direction.

    Weighting to population share fixes the arithmetic: each segment's contribution to the overall result is scaled to how much of the country it actually is. A segment that is a fifth of the population moves the number five times as much as one that is a twentieth.

    The practical consequence is that results are returned per segment, never only as an average. A national 'yes' with one large segment firmly against it is a different situation from a national 'yes' everyone is lukewarm about, and the average is identical in both cases.

    /census-weighting works this through with numbers, including what the same run looks like unweighted.

    What a synthetic panel can tell you

    • Direction: whether an idea, price or line reads as acceptable or not, before you have spent anything on finding out.
    • Structure: which parts of a population split, and roughly how far apart they sit.
    • Reasoning: the argument each segment reaches for, in its own register — usually more useful than the score itself.
    • Ranking: which of several options a cross-section reads more favourably, and which segments disagree with that ranking.
    • Blind spots: the objection nobody in the room raised, because nobody in the room is in that segment.

    What it cannot tell you

    Reynard is a simulation. It is not fieldwork, it carries no statistical significance, and it cannot replace talking to a customer. Treat it as a fast, structured second opinion that tells you where to look — not as evidence.

    • How many people will buy. There is no purchase intent in a simulation, and a favourable reading is not a forecast of revenue.
    • Anything with statistical significance. There is no sample, no margin of error, and no confidence interval, because nothing was sampled.
    • What your existing customers think. The panel models a population. Your customers are a self-selected slice of it that already chose you, and they are frequently unlike the average in exactly the way that matters.
    • How a niche behaves. The narrower and more specialist the audience, the thinner the public statistics behind it, and the more the model is filling in from general priors.
    • Anything a human notices and a model does not: the phrasing that reads as tone-deaf this month, the reference that has aged badly, the detail only someone living that life would flag.

    The honest limit: a population is not your customers

    This is the sentence to keep. A synthetic panel models a population, not your market and not your buyers. Every weighting decision on this page makes it a better model of a country, and none of them makes it a model of the people who would give you money.

    For a broad consumer question — a mainstream product, a price that a general audience has to find reasonable, a public communication — that gap is small enough to be useful. For a specialist B2B tool, a genre readership or an established customer base, the gap is the whole answer, and a run will read confidently past it.

    Used well, the output tells you where to look and what to ask. It is not the answer, and no amount of weighting turns it into one.

    Keep reading

    Run the free check first

    One check costs nothing and needs no account. You see the format of the output before you decide whether it is worth paying for a full run.

    Start the free check