Social Rocket
Goodway Group uses synthetic audiences to run through media plans before real campaign tests. According to a report from 12 October 2026, the agency is working with the data provider Arima for this. Its offering also includes a model of the German population with 80 million virtual people and around 20.000 modelled variables. The figure describes artificially generated datasets – not a survey of 80 million real people.
One media plan, many simulated scenarios
According to its own information, Arima combines various datasets into a "Synthetic Society". Goodway wants to use it to check in advance how different audiences or budget allocations are likely to respond. The hypotheses are then to be tested in limited real ad campaigns. The advantage in agency practice: scenarios can be compared without spending test budgets on every idea right away.
This case is not merely the announcement of a new AI tool. It shows that synthetic customer groups are increasingly feeding directly into decisions about media budgets and creative strategy. According to Digiday, brands in the beauty and consumer goods sectors, among others, are using similar methods to examine product ideas and audience assumptions early.
80 million model personas are not 80 million voices
This is the most important limitation. The dataset is meant to replicate distributions and correlations. It does not follow from this that its forecast is reliable for a specific brand, region or ad. Moreover, biases in source data or model assumptions can be carried forward.
A research paper by Netzer and Sambandam explicitly distinguishes between generated LLM responses, segment personas and individual digital twins. It shows that high agreement on average can conceal weak predictions for individual people. What matters decisively is for which question a synthetic model was previously validated against real data.
Even more concrete is a study of 399 statistically meaningful headline A/B tests from the Upworthy archive. There, an LLM without elaborated personas predicted the ranking of the variants better than the tested persona panel. The study is a preprint with a limited scope, so it does not refute all synthetic data models – but it contradicts the assumption that more persona detail automatically means better forecasts.
Three checks for German marketing teams
First: Ask for an error report against independent real surveys and campaign data from the target market, Germany. A matching overall demographic distribution is not enough.
Second: Distinguish the simulation of a budget effect from an effect measured causally. An artificial control market can generate ideas, but it does not replace a properly designed geo holdout or another real incrementality test. Our overview of retail media and incrementality explains how real causal measurement is used in retail.
Third: Define which decision the model is supposed to improve at all: audience segments, messaging, budget level or channel selection. For automated delivery, additional control questions apply, which we explain in our piece on agentic media buying and in our assessment of supposedly fivefold ROAS values.
Assessment: Synthetic panels can speed up research and pre-selection. Their economic benefit is only proven once decisions based on them outperform previous methods against real results.
Sources: Digiday, 12.10.2026; Arima, synthetic population Germany; Netzer/Sambandam, Synthetic Data in Marketing Research, September 2026; Maiorano, Sim-to-Real study, preprint 2026.
Illustrative image: Luke Chesser / Unsplash. Not an image of the Arima software.