SINUS Milieu-Minds in a Reality Test Against a SINUS Survey

Many research questions never get answered because a dedicated survey is not worth the time or budget. How does a new message land in the Traditional milieu, and how with the Expeditives? Which three of ten campaign motifs should go into the pretest? SINUS Milieu-Minds are built for that stage: synthetic people from all ten Sinus-Milieus, each with an own life situation, values and biography, who can be asked questions like real participants and answer within minutes.

This study checked how close they come to real people. 200 SINUS Milieu-Minds, 20 per milieu, answered 24 questions on personal concerns. SINUS had asked the same questions to 2,005 people in June 2026. The Minds did not see the results of that survey.

Per milieu, the Minds' answer distributions reached 91.7% closeness to the survey. Two real survey waves with the same questionnaire reached 96.3% against each other. A general AI model given a simple milieu instruction reached about 72%.

The result in one view

Distribution closeness is 100 minus the mean absolute difference between the synthetic and the real answer shares, in percentage points. 100% would mean that every answer distribution is identical to the survey. The score measures how close two distributions are. It is not a rate of correct answers and does not describe individual respondents.

Research question and reference data

A survey gives each milieu a set of answer shares: how many people have very great concern, some concern, little concern or no concern at all about a topic. The question was how closely synthetic Milieu-Minds reproduce all four shares together, milieu by milieu.

The reference came from SINUS-Institut: a survey from June 2026 with 2,005 respondents as the primary comparison and a survey from March 2026 with 1,047 respondents with the same questionnaire. The survey results were withheld from the Minds at runtime. The crosstabs are proprietary SINUS data and are not published with this article.

Per milieu: 91.7% closeness

On average across the ten milieus, the Minds differed from the June survey by 8.3 percentage points per answer level. That corresponds to 91.7% distribution closeness. The milieus range from 87% for the Performer milieu to 94% for the Nostalgisch-Bürgerliche milieu. The Performer milieu and the Adaptiv-Pragmatische Mitte are the weakest fits.

Individual answer shares can deviate clearly. For 645 of the 960 answer shares (67.2%) the absolute difference is less than 10 percentage points; the largest single difference is 47.2 percentage points. The most frequent answer of the survey was also the most frequent answer of the Minds in 65.4% of the 240 combinations of question and milieu.

The results therefore give indications of topics and answer distributions within a milieu. For a realistic estimate of actual shares, a survey is still needed.

Against the earlier March wave, the Minds reached 91.4% closeness. The most frequent answer matched in 58.8% of combinations, slightly below the 60% the study had set as a target for the June comparison.

How to read the survey waves

Two real survey waves with the same questionnaire reached 96.3% closeness to each other, per milieu. The differences between them can reflect both sampling variation and real changes in society between March and June. The value is therefore a reference point, not a fixed upper limit. It also shows that two real surveys are still clearly closer to each other than the Minds are to a real survey.

Across all milieus: 95.2% closeness

For the comparison with all 2,005 respondents, the answers of the 200 Minds were weighted by the milieu composition of the June survey, that is, the way the milieus are represented in the survey. This weighting reproduces the official total column of the SINUS crosstabs.

At this level, the Minds differ by 4.8 percentage points per answer level on average; without weighting the figure is 4.7. The two real survey waves reach 98.9% on the same level.

A typical example, not a best case, is concern about one's own health:

The rounded values differ by 3, 7, 6 and 4 percentage points, 5.0 on average. That is close to the average of 4.8 percentage points across all 24 questions on the whole-sample level, calculated from unrounded data.

Why not simply ask a general AI model?

A separate method control tested exactly that. Gemini 3.7 Flash and Claude Sonnet 5 received the same 24 questions with a single instruction to answer as a member of a named Sinus-Milieu. Both reached about 72% closeness per milieu, about 20 points less than the Milieu-Minds. Adding the official milieu portraits from the SINUS website did not improve Gemini's result.

The reason is visible in the answers. A general AI model speaks for a whole milieu with a single voice. Asked the same question 20 times, it usually gave the same answer 20 times: Gemini in 194 and Claude in 211 of 240 combinations of question and milieu. Opinions in real milieus differ, and so do those of the Milieu-Minds: they used almost all four answer levels on average and never answered unanimously.

This control ran separately and at a different time from the Minds run, and it was added after the Minds result was known. It shows the gap to a simple instruction. It is not a matched experiment and does not measure how much any single component of the Milieu-Minds contributes.

Do the Minds get the differences between milieus right?

For milieu research, the differences between milieus matter as much as the individual shares. A follow-up analysis of the same data checked this. It was calculated after the main result was known and uses the mean concern level per question and milieu.

A difference counts as clear when two milieus differ in the survey by more than two standard errors, which applies to 586 pairs of milieus across all questions. Where real milieus clearly differ, the Minds ranked them correctly in almost nine out of ten cases, the AI model with only a milieu name in two out of three. The Minds tend to make the differences between milieus larger than they are.

The analysis also puts the exact shares in context. About three quarters of the remaining squared deviation is what random sampling with 20 respondents per milieu would produce on its own. At the same time, a simple reference value that ignores topic and milieu, a fixed answer mix from the March survey, deviates by 7.5 percentage points per answer level and is therefore slightly more accurate than the Minds on exact shares per milieu. The strength of the Minds lies in ranking topics and milieus, not in exact percentages.

What SINUS Milieu-Minds are for

Milieu-Minds are a fast pre-check and do not replace primary research. They help to choose the right questions and the most promising variants. A field study with SINUS can then deliver robust figures.

• Pre-select campaigns and messages: check motifs, claims and arguments for each milieu and send only the most promising ones into the pretest.
• Set up studies more precisely: test hypotheses and sharpen questionnaires before fieldwork starts.
• Understand reasons: the Minds explain their answers. That gives indications for open questions and helps to explain differences between milieus. Whether real respondents give the same reasons is shown by the field study.

The reality test covered attitude questions. For campaign decisions, the recommendation is therefore both: the Minds for the pre-selection, the field study for the decision.

SINUS Milieu-Minds are available to SINUS clients through SINUS and the Minds Marketplace. Within Minds, the same Audience can answer scale and choice questions, open-ended follow-ups and stimulus tests in one workflow, where enabled. Contact at SINUS-Institut: Jan Hecht, [email protected].

Scope and limitations

• The Minds came from existing production groups, not from a freshly drawn sample, and the study was not preregistered.
• With 20 Minds per milieu and question, the resolution of each milieu distribution is coarse.
• The general AI model controls were added after the Minds result was known, ran at different times and covered two providers. The internal model of the earlier Minds run is not verified, so the gap does not isolate a pure architecture effect.
• Only short public milieu portraits were tested as added context; detailed licensed milieu portraits were not.
• One questionnaire, one country and attitude questions only. The comparison is exploratory, not confirmatory.
• The analysis of milieu differences and reference values was calculated on the same data after the main result was known.

Data and approval

The comparison is based on 200 SINUS Milieu-Minds with 4,800 answers to 24 concern questions, compared with SINUS surveys from June 2026 (n = 2,005) and March 2026 (n = 1,047). Values per milieu compare each milieu on its own; the averages across the ten milieus are simple means. On the whole-sample level, the Minds are weighted by the milieu composition of the June survey. The results are not projected to the population. Tested AI models: Gemini 3.7 Flash and Claude Sonnet 5, runs in August and September 2026.

SINUS-Institut supplied the survey reference and reviewed and approved the case content. Details on method and limitations are available on request.