AI market research: what ChatGPT can and cannot do for your study

How many European firms already use AI?

Adoption has accelerated sharply. According to Eurostat, 19.9% of EU enterprises with ten or more employees used at least one AI technology in 2025, up from 8.1% in 2023. The share more than doubled in two years.

The averages hide a wide spread between member states. The Netherlands is well ahead of the EU average; Bulgaria is well behind it, and the distance between them keeps growing.

In 2021 the gap between Dutch and Bulgarian firms was 9.8 percentage points; by 2025 it had reached 24.6. Two cautions apply. The indicator counts any AI technology, from text analysis to machine learning, so it does not show how many firms use AI for marketing or research specifically. And it covers firms with ten or more staff, so it says nothing about sole traders. Other indicators of business digitalisation, including social media use and paid online advertising, are in our business data section.

For a research buyer, the practical consequence is that AI tools are now routine in many client and supplier teams. The question is no longer whether to use them, but where they help and where they mislead.

What does a language model do well in a research project?

Language models work best on material you already have, with an instruction precise enough to check the result. Mapped onto the stages of a typical project:

The time saved is real at the analysis stage. Grouping two thousand open-ended comments into themes used to take days of manual coding; a model produces a first pass in minutes, and the researcher's job shifts to correcting the code frame rather than building it from nothing.

Where does ChatGPT fail as a research tool?

The failures are predictable once you remember what a language model is: a system trained on past text that generates the most likely continuation of your prompt.

1. It answers questions it cannot know. Ask what share of British pet owners would pay for a premium food subscription and you will get a confident percentage. No sample stands behind it, so there is no margin of error and no way to judge how wrong it might be.
2. Its knowledge has a cut-off. A price rise last quarter, a new competitor or a change in regulation will be missing unless you supply it.
3. It invents sources. Models sometimes cite reports that do not exist, or attribute real-looking tables to real institutions.
4. It speaks for the people who write online. Training data over-represents frequent internet users. Eurostat reports that 89.5% of individuals in the EU used the internet daily in 2025, against 82.7% in Bulgaria and 98.4% in the Netherlands. The people outside that daily group are under-represented in what models learn from, and they are also the hardest to reach through online panels.

The fourth problem matters most in multi-country work. A model's view of an average consumer is likely to fit Amsterdam better than a small town in north-west Bulgaria.

Synthetic respondents: a test bench, not a sample

Synthetic respondents are AI-generated personas that answer a questionnaire as if they were customers. The output looks like survey data, with percentages, segments and quotes, which is exactly what makes it risky.

They are useful in a narrow role:

• checking whether questions are understood as intended before a pilot;
• estimating how long the questionnaire takes and where routing breaks;
• generating hypotheses and answer options for a later survey of real people.

They are a poor substitute for real respondents because they have no sampling frame, so no margin of error can be calculated (see our sample size guide for what that means in practice). Their answers vary less than human answers do, they reproduce stereotypes present in training data, and they cannot know about events after the model's cut-off.

If a supplier proposes synthetic data for a real decision, ask for a validation study: the same questionnaire fielded to a real sample, with agreement criteria set in advance. Compare the distribution of answers, the ranking of options and, above all, the differences between subgroups, which is where synthetic data most often drifts.

How to validate AI output before it reaches a report

Treat every AI output as a draft from a capable but unreliable junior colleague. Five checks catch most problems:

• Double-code a random subset. A researcher codes a sample of open answers, for instance one in ten, without seeing the machine codes, then compares the two.
• Trace every number. Each statistic must lead to a named table, year and URL. The AI adoption figures above, for example, come from Eurostat table isoc_eb_ai.
• Check quotes against transcripts. Require the line or timestamp for each quote, and open it.
• Re-run the prompt. If a second run on the same material produces different themes, the themes are not robust.
• Log the set-up. Record the model, version, date and prompt, so the analysis can be repeated.

Governance: data protection, the AI Act and research codes

Interview transcripts, recordings and survey files often contain personal data: names, voices, employers, sometimes health or financial details. Uploading them to a public chatbot is a transfer to a third party, which under the GDPR and the UK GDPR requires a lawful basis, information for participants and a processing agreement with the provider.

A workable policy covers four points:

• remove names, contact details and other identifiers before text enters an AI tool;
• use business versions with a data processing agreement and training on your data switched off;
• tell participants in the consent text that their answers may be processed with AI;
• keep the raw data separately, so the analysis can be repeated without AI.

The EU's Artificial Intelligence Act, whose obligations apply in stages, includes transparency duties when people interact with an AI system. If you run AI-moderated interviews, tell participants that the interviewer is not a person. The ICC/ESOMAR International Code also requires honesty towards clients about methods, so a report should state which stages were AI-assisted and how they were checked. Our methodology page describes how we apply these rules.

When do you still need a commissioned study?

AI shortens preparation and analysis. It cannot supply the evidence when the decision depends on how many people will act, at what price, or why:

• to measure shares, awareness or purchase intent, you need a quantitative survey with a defined sample; the sample size calculator shows how many interviews a given precision requires;
• to understand motives that customers do not write about online, in-depth interviews with real buyers remain the reliable route;
• to analyse large volumes of reviews and social media posts, a digital market analysis combines machine classification with human checking;
• for official market statistics, start with our list of free market data sources for Europe.

If you are unsure which applies, describe the decision in the proposal request form and we will suggest a design that uses AI where it saves time and real respondents where the answer depends on them.