Methodology: how we design, check and report research

Standards

We work to the principles of the ICC/ESOMAR International Code on market, opinion and social research. In practice this means that research is kept separate from selling, that respondents take part voluntarily and are told who is collecting the data and why, that children and other vulnerable people receive extra protection, and that findings are reported honestly, including results the client may not welcome.

We use ISO 20252, the international standard for market, opinion and social research processes, as a reference framework for how a project is planned, documented and checked. We do not claim certification against it; we use its structure as a checklist so that every stage, from briefing to archiving, leaves a record.

Personal data is processed under the GDPR. Respondents give informed consent, we collect only what the study needs, and identifying details are separated from answers or removed before data files are delivered.

Sampling and statistics

Every sample starts with a definition of the population: who exactly the results should describe, in which country and at what date. We then choose the sampling frame, the list or source from which respondents are drawn, and state where it falls short of the population.

Most consumer surveys use quota sampling from online panels or telephone samples, with quotas on age, gender and region that reflect official population statistics. Where a probability sample is feasible, for example from a complete customer list, we use random or stratified sampling. After fieldwork, small imbalances are corrected by weighting, and we report the effective sample size so that the cost of weighting in precision is visible.

Sample size follows from the precision needed in the smallest subgroup that matters, not from a round number. At 95% confidence, around 385 interviews give a margin of error of about ±5 percentage points for a result near 50%; our sample size calculator shows the arithmetic. For non-probability samples, we present margins of error as a guide rather than a strict bound.

Quality control

Fieldwork control

Telephone interviews are monitored live or from recordings, and a share of each interviewer's work is checked. Face-to-face and mystery shopping visits are verified against time stamps, locations and receipts where available. Online fieldwork is monitored daily for quota fill, dropout points and unusual answer patterns, and we run a soft launch before full fieldwork to catch routing errors.

Data cleaning

Before analysis, every survey dataset goes through documented data cleaning:

• Speeders: respondents who finish far faster than the median time for their route through the questionnaire are removed.
• Straightlining: identical answers down whole grids, especially where items point in opposite directions, are flagged and reviewed.
• Attention checks: instructed-response items reveal respondents who are not reading the questions.
• Open answers: nonsense, copied text and answers unrelated to the question are flagged, together with duplicate respondents.
• Consistency: contradictions, such as a non-user rating a product in detail, are checked case by case.

The report states how many interviews were removed and why, so you can judge the cleaned sample for yourself.

Qualitative methods

Qualitative research starts with recruitment. A screener checks that each participant fits the profile, has not taken part in similar research recently and has no professional link to the category. We recruit a mix of profiles within each group, so that one type of view does not dominate.

Moderation follows a discussion guide agreed with you in advance, but the moderator is free to follow unexpected leads. Focus groups and in-depth interviews are recorded with consent and transcribed; observers can watch live or online where the format allows.

Analysis is thematic: transcripts are coded against the research questions and against themes that emerge from the material, and conclusions are supported by verbatim quotes. We are explicit about what qualitative research can show, namely reasons, language, barriers and ideas, and what it cannot, such as shares or market sizes, which need a survey.

Open data on this site

The open data on this site come from three primary publishers: Eurostat for EU-wide statistics on households, businesses and industries, the World Bank's World Development Indicators for population and GDP, and CBS StatLine for Dutch consumer confidence. Each chart names the publisher, the dataset code and the latest year shown, so any figure can be checked at source.

The figures are retrieved directly from the publishers' databases rather than copied from secondary websites, and they are refreshed when a new release appears. We do not adjust or estimate official values; where a series has gaps, such as years in which a Eurostat survey module was not run, the chart shows the gap rather than filling it. In reports, every desk-research figure carries its source and year in the text. See the market data section.

How we use AI

We use AI tools where they save time without replacing judgement: a first pass at coding open answers, draft summaries of interview transcripts, first drafts of translations and the search phase of desk research. Every AI output is checked by a researcher before it is used, codes and themes are verified against the raw answers, and translations are checked by a native speaker.

We do not use synthetic respondents as a substitute for real ones. Language models can help to draft hypotheses or test a questionnaire for clarity, but their answers are not data about your market. Personal data and confidential client material are not entered into AI tools without a contractual basis that protects them. Our article on market research with AI discusses the limits in more detail.

Frequently asked questions

Do you use probability or non-probability samples?

Both, depending on the frame available. Surveys of your own customers or members can often use random samples from a complete list. Consumer surveys usually rely on online panels or telephone samples with quotas, which are non-probability samples. We say clearly which type was used, and we interpret margins of error for quota samples as indicative rather than exact.

How many interviews do you remove during data cleaning?

It depends on the sample source, the length of the questionnaire and the target group, so we do not work to a fixed quota of removals. Every report states how many completed interviews were removed, for which reason and whether replacements were recruited, so that you can see the effect of cleaning on the final sample.

Can we observe focus groups or listen to interviews?

Usually, yes. Focus groups can be observed live or through a video link, and you can listen to selected telephone interviews or recordings, provided the participants have agreed. Observers do not take part in the discussion and are bound by the same confidentiality rules as the research team.

Do you weight survey data?

When the achieved sample differs from the population on characteristics that affect the results, we weight it to known population figures, typically age, gender and region. We keep weights within sensible limits, report the effective sample size and show unweighted bases in the tables, so that you can see how much the weighting matters.

Do you use ChatGPT to write reports?

No report is written by an AI tool. Researchers write the analysis and conclusions. We may use AI to speed up specific tasks, such as a first coding of open answers or a draft transcript summary, but a researcher checks every output against the underlying data before anything reaches a report.