Every marketer knows the dilemma. You have three candidate headlines for a campaign that will run in four countries, and the honest way to choose between them — recruit panels, run fieldwork, wait for the field institute — costs more than some of the campaigns themselves and takes weeks you don't have. So most messages ship on instinct, and the market runs the test for you, with real budget as the entry fee.

Samplio is our answer to that dilemma: a synthetic audience research platform for EU markets. You paste your message — or 2–5 variants, or a conditional survey — pick an audience, and in about a minute you get a segment-by-segment read: how strongly each segment leans toward appeal, trust, skepticism, indifference or concern, plus a short "voice of the segment" quote that tells you why.

A Synthetic Population, Not a Focus Group Prompt

The obvious way to build something like this is to ask a language model to "pretend to be a German consumer". We deliberately didn't. Samplio's audiences start from real public statistics: NUTS-2 regional population shares from Eurostat and household income quintiles from EU-SILC, with the data provenance recorded per axis. A Dirichlet + IPF step expands those marginals into a joint population matrix — 10,000+ synthetic agents per run — and K-Means clustering groups them into the segments you see in a report. The segments emerge from the data; nobody hand-picks a persona.

Only then does the LLM enter: one call per segment, conditioned on that segment's region, income band and value orientation. Country packs are simulated in the audience's own language — German copy is judged in German, Polish copy in Polish — across 27 EU country packs and 7 simulation languages. Ask for a macro region or the EU-27 average and Samplio doesn't ask the model to imagine "Europe": it runs every member country separately and combines the results population-weighted, in plain code the LLM never touches.

Honest by Mechanism

Synthetic research has a credibility problem, and it deserves one: it is easy to dress up model output as fieldwork. We took the opposite bet — that the way to sell synthetic data is to make it structurally incapable of overclaiming. In Samplio this is not a policy document; it is code:

Every result carries a SYNTHETIC SIMULATION watermark, on screen and in print. No accuracy figure appears anywhere until it has been measured against a sourced corpus of real-world cases — the backtest tooling ships with the product, and until that corpus exists, the honest number of accuracy claims we publish is zero. There is no ±% margin of error on any report, because sampling error is a statement about random samples of real people and would be fake precision on synthetic agents; what we report instead is re-run stability. And where an axis of the population model is not yet calibrated against real survey data, the engine mechanically stamps every affected output — the stamp cannot be configured away, by us or by anyone.

The Red Line

One decision predates the first line of code: Samplio is for commercial research only. Election and political-campaign use is prohibited on every surface — the profile schema cannot even express a political domain, and every customer-supplied text passes a two-layer screen at the engine boundary: an eight-language keyword filter plus an LLM classifier that fails closed on doubt. This is not a configuration option, and it is written into the terms of service.

What It Looks Like in Practice

The homepage at samplio.talivio.com shows a real run, reproduced verbatim from the engine: a German oat-milk message ("same taste, 30% less packaging") tested on the German consumer pack. The progressive segment responds with 55% appeal and talks about waste reduction; the price-conscious segment mostly shrugs — 35% indifference — and asks whether less packaging will mean a lower price. That split is exactly the kind of thing that decides where a media budget goes, and it surfaced in about a minute.

Samplio is live now, with free registration and one-time pilot packages from €390 — no subscription during the pilot phase. It runs on the same engine discipline as the rest of Talivio's AI products: real data provenance, per-call audit trails designed with the EU AI Act in mind, and a hard preference for saying "we don't know yet" over inventing a number.