What AI does in market research, and what it doesn’t
Market research answers three questions: who are my customers? What do they really need? And why do they choose me or not? AI doesn’t change these questions, but it changes how fast you analyse the material you already have.
| Task | AI helps well | AI helps badly |
|---|---|---|
| Analysing interviews | Summarising, finding themes, collecting quotes | Deciding which statement matters |
| Analysing reviews | Sorting hundreds of texts into praise, criticism and wishes | Spotting which reviews are bought or fake |
| Surveys | Coding free-text answers, drafting questions | Replacing representative samples |
| Personas | Condensing existing data into profiles | Inventing new, real insight about customers |
| Competition | Comparing websites, prices, messages | Knowing competitors’ internal strategies |
Synthetic personas: useful for preparation, dangerous as a replacement
Synthetic personas or synthetic users are AI-generated profiles that imitate a target group and answer questions as if they were customers. They come from language models, not from conversations with real people.
In 2024 the Nielsen Norman Group re-ran three studies it had previously done with real participants, using synthetic users. The result, per the article: answers were often shallow or overly favourable. One synthetic user claimed to have completed every course mentioned, while real learners often dropped out. The AI users rated discussion forums positively; most real participants found them contrived.
| Use | Recommendation |
|---|---|
| Desk research on a target group | Suitable for pooling publicly available knowledge |
| Preparing and testing an interview guide | Suitable |
| Forming hypotheses and proto-personas | Suitable if they are tested with real people afterwards |
| Having concepts or offers rated | Unsuitable: AI tends to agree with everything |
| Niche target groups | Unsuitable: data is patchy |
| Presenting results as real customer opinion | Unacceptable and misleading |
Based on the Nielsen Norman Group recommendations, plus how they apply to marketing.
For marketing that means: an AI persona can help you ask better questions. You get the answers from real people. And if your ads say “our customers say …”, you need real customers, not simulated ones.
Whom to interview and how to find them
The best analysis is useless if you talk to the wrong people. Three groups bring the most insight.
| Group | Why | How to reach them |
|---|---|---|
| New customers from the last three months | They still remember why they bought | Personal email asking for 20 minutes |
| Prospects who dropped out | They show what’s missing or off-putting | Short message after a quote without a deal |
| Non-customers in the target group | They show how they solve the problem today | Network, communities, referrals |
A small thank-you, such as a voucher, raises acceptance without much bias, as long as it isn’t tied to the answers. Ask permission for recording and analysis beforehand and say what you’ll use the results for.
Analysing interviews with AI
Five to eight honest conversations with customers or prospects often bring more than any database. The tedious part used to be the analysis. That’s where AI saves the most time.
- Get consent: ask before recording whether you may record and which service will transcribe.
- Transcribe: with a tool you have a data processing agreement with.
- Anonymise: remove names, companies and identifying details before the AI analysis.
- Summarise each conversation: problem, trigger, alternatives, buying criteria, verbatim quotes.
- Condense across conversations: which themes recur? Where do statements contradict each other?
- Check yourself: match every key statement against the original transcript. AI sometimes summarises things that weren’t said that way.
One key trick: have the AI give the verbatim quote with its location for every statement. Quotes without a location are a warning sign.
Better questions: situations, not opinions
The most common weakness in customer conversations is opinion questions: “Would you buy this?” People answer politely and vaguely. Questions about the last concrete situation are more productive.
| Instead of | Better |
|---|---|
| Would you use a tool like this? | When did you last have the problem? What did you do then? |
| What matters to you? | What tipped the balance in your last decision? |
| How much would you pay? | What do you spend on solving it today, in money and time? |
| Do you like our offer? | What almost stopped you from getting in touch? |
AI helps rephrase guides in this direction. And in the analysis it shows you where participants talk about real experiences and where about wishes. Real experiences weigh more.
Mining reviews and forums
Public reviews on Amazon, G2, Trustpilot or in forums like Reddit are a treasure: customers write in their own words what bothers and delights them, including about your competitors.
- Collect: copy a manageable selection by hand or use official exports and APIs. Check the platform’s terms before fetching content automatically.
- Sort: AI groups by theme (price, quality, service, usability), sentiment and frequency.
- Borrow the language: the words customers use are often the best words for your website and ads.
- Find gaps: what do customers criticise across all providers? That’s often where the opportunity lies.
- Careful with numbers: reviews aren’t a representative sample. The very unhappy and very happy write more often.
How customer language feeds into ads and UGC videos is shown in AI avatars and UGC ads.
Planning and analysing surveys
With surveys, AI helps in two places: writing good questions and analysing open answers.
- Check questions: AI finds leading, double-barrelled or unclear questions. One question per idea, neutral wording.
- Code open answers: AI suggests categories and assigns answers. Check a sample by hand.
- Calculate, don’t estimate: compute percentages and cross-tabs in a spreadsheet or an analysis tool that shows its steps, not by asking a chat to estimate.
- Name the sample honestly: 40 answers from your newsletter are a mood check, not a market study.
For analyses with numbers, see also AI marketing analytics.
Building personas from real data
A good persona sums up what you know about a customer group and makes visible what you don’t know yet. AI helps condense it, as long as the material is real.
| Building block | From | Example |
|---|---|---|
| Trigger | Interviews, enquiries | “I noticed I spend three hours a week on meeting notes.” |
| Goal | Interviews | More time for client work |
| Barriers | Reviews, interviews | Fear of data protection problems |
| Alternatives | Interviews, reviews | Excel, asking a colleague, doing nothing |
| Buying criteria | Interviews, sales calls | Price per month, German-language support |
| Open questions | Gaps in the material | Who else decides in the team? |
Mark the source for every item. A persona where every line comes from real data is a tool. One where everything sounds plausible but nothing is backed up is a story.
How personas fit into the overall plan is covered in AI marketing strategy.
From insight to action
Research that ends up in a folder was expensive. Turn every important insight into a concrete action with an owner and a date.
| Insight (example) | Action |
|---|---|
| Customers name data protection as the biggest barrier | Move the data protection section up on the offer page, extend the FAQ |
| Prospects compare with Excel, not competitors | A “Excel vs our tool” comparison instead of a competitor table |
| Almost all new customers came via referrals | Actively ask for referrals before raising the ad budget |
| Onboarding criticism keeps recurring in reviews | Redesign the first week after purchase |
After implementing, measure whether anything moved. How is covered in AI marketing analytics.
Watching competitors and AI answers
Market research today also means: what do AI assistants say about you and your competitors when customers ask? Those answers shape whom customers consider at all.
- Ask five real customer questions in ChatGPT, Claude, Gemini and Perplexity without naming yourself.
- Note which providers are named and why.
- Repeat regularly, because answers vary from run to run.
That’s exactly what our AI duel does: it asks several AI assistants typical customer questions and shows whom they recommend. How to show up in such answers is covered in GEO.
Tools: what you really need
Little is needed to start: a recording and transcription tool with a data processing agreement, a chat assistant with file upload and a spreadsheet for numbers. Specialised research platforms only pay off once you run many conversations regularly.
- Transcription: many meeting tools create transcripts automatically. Check where the data is stored and whether you can sign a processing agreement.
- Analysis: a chat assistant handling large files, instructed to always give quotes with locations.
- Numbers: a spreadsheet or analysis tool with visible calculations.
- Filing: one folder per study with guide, consents, transcripts and results, so in six months you still know what a decision was based on.
Data protection with customer data
| Topic | What applies | Source |
|---|---|---|
| Legal basis | Art. 6 GDPR: analyse interview recordings, surveys and customer data only with consent or another legal basis | GDPR |
| Notice | Art. 13 GDPR: participants learn beforehand what is analysed, for what and with which services | GDPR |
| Processing on your behalf | Art. 28 GDPR: a contract with every AI service processing personal data | GDPR |
| Data minimisation | Art. 5 GDPR: process only what’s needed, remove names and details beforehand | GDPR |
Public reviews are public, but not free for any use: authors’ names don’t belong in your analysis, and the platform’s terms apply. For advertising, German competition law (§ 5 UWG) applies: if you make claims about customers, you must be able to prove them.
Checklist for AI market research
- Research question in one sentence
- At least five conversations with real customers or prospects planned
- Consent for recording and AI analysis obtained
- Transcripts anonymised before analysis
- Every key statement backed by a verbatim quote with location
- Reviews and forums used as a supplement, not a sample
- Synthetic personas only for preparation, clearly labelled as such
- Numbers calculated in a spreadsheet or analysis tool, not estimated
- Persona with a source per line and a list of open questions
12 prompts for AI market research
I want to find out: [topic]. Turn it into one precise research question and three sub-questions I can answer in customer conversations.
Create a guide for a 30-minute conversation with [target group] about [topic]: open questions, no leading questions, ask about specific recent situations rather than opinions.
Play a person from [target group] and answer my guide. Then mark questions that are unclear, leading or double-barrelled. Note: your answers only test the guide; they aren’t research.
Summarise this anonymised transcript: problem, trigger, alternatives, buying criteria, objections. Give the verbatim quote with line number for each point: [transcript].
Here are summaries of [number] conversations: [text]. Which themes appear at least three times? Where do statements contradict each other? What’s missing?
Sort these reviews by theme, sentiment and frequency. Give two typical customer phrasings per theme: [reviews without names].
Here are reviews of three providers: [text]. What do customers criticise in all three? What is praised for only one?
Check these survey questions for leading, double-barrelled and unclear wording and suggest neutral versions: [questions].
Suggest 5–8 categories for these open answers and assign each answer. Mark answers you can’t assign confidently: [answers].
Condense this real data into a persona with trigger, goal, barriers, alternatives, buying criteria and open questions. Give the source for each line: [data].
Which phrasings from these customer quotes work for headlines and ads? Verbatim or near-verbatim only: [quotes].
Here are my research findings: [list]. Which are only assumptions? What should I test with real customers before spending money?
The most common mistakes
- Personas from a prompt. Plausible isn’t true.
- Having AI rate concepts. It agrees with almost everything.
- Reviews as a sample. Loud voices aren’t the majority.
- Letting AI estimate numbers. Calculations belong in a spreadsheet, not a chat.
- Unfiltered customer data in tools. Anonymise first, then analyse.
- Never talking to customers. No tool replaces five honest conversations.
Frequently asked questions
What’s the difference between a persona and a target group?
A target group describes attributes such as industry, age or region. A persona condenses what you know about a typical person in that group: triggers, goals, barriers and buying criteria. It’s only as good as the real data behind it.
How long does a small customer study with AI take?
With five to eight 30-minute conversations and AI-assisted analysis, often one to two weeks, depending on how fast you get appointments. The analysis itself is much faster with AI than by hand.
Can I use synthetic personas for ad copy?
As an idea generator for variants, yes. But claims about customers in ads must rest on real customers, or they’re misleading.
Which questions should I ask in customer interviews?
Questions about the last concrete situation rather than opinions: when did you last have the problem, what did you do, what tipped the decision? Opinion questions like “Would you buy this?” get polite but vague answers.
Whom should I interview?
Most productive are new customers from recent months, prospects who dropped out and non-customers in the target group. That shows reasons for, against and around your offer.
Can AI personas replace real customer interviews?
No. In a comparison with real studies, the Nielsen Norman Group found that synthetic users often give shallow or overly positive answers. They help with preparation, not as a replacement.
What are synthetic personas useful for?
Pooling publicly available knowledge about a target group, preparing an interview guide and forming hypotheses that you then test with real people.
May I analyse interview recordings with AI?
With participants’ consent, prior notice and a data processing agreement with the AI service. Remove names and identifying details first. Not legal advice.
How many interviews do I need?
For a first picture, often five to eight conversations with people from the same target group. When themes start repeating, you have enough for the next decision.
May I analyse reviews from Amazon or Trustpilot?
You can read public reviews and analyse them for yourself; check the platform’s terms before fetching automatically and leave out authors’ names.
Can AI calculate survey results?
AI helps code open answers. Percentages and cross-tabs should be calculated in a spreadsheet or an analysis tool that shows its steps.
How do I find out what AI assistants say about my company?
Ask five real customer questions in several assistants without naming yourself, and note who gets recommended. Repeat regularly, because answers vary.
Sources
- Nielsen Norman Group: Synthetic Users – Rosala, Moran, Jun 21, 2024
- GDPR (Regulation (EU) 2016/679) – Art. 5, 6, 13, 28
- § 5 UWG (Germany) – Misleading claims






