AI can summarise what has been written about a market. It cannot measure what nobody has recorded yet.
Can AI do a market analysis?
Yes, for the parts that rest on existing text and data. AI market research works for reading, sorting and drafting. A market analysis also needs current measurements, and a language model does not collect those by itself.
Think of it as two jobs. The first is synthesis: what has already been said about this market? The second is measurement: what are buyers doing now? AI helps directly with the first. The second needs a source outside the model.
A Harvard field study followed 758 consultants on tasks from their daily work. For tasks within AI's capability, GPT-4 raised speed by over 25%.
Human-rated performance rose by over 40% on those same tasks. The authors call the boundary a jagged frontier, because AI falls short on other tasks.
This was consulting work and not market research as such. It supports one narrower point: you need to know which side of the line a task is on.
| Research job | Role for AI | What you still need |
|---|---|---|
| Desk research | Find, sort and summarise public material | The original documents |
| Questionnaire draft | Propose questions and answer options | A researcher's review and a pilot |
| Market size | Structure the calculation | Sourced inputs for every number |
| Demand direction | Explain data you supply | A dated behavioural source |
| Consumer reaction | Generate early ideas and hypotheses | Real respondents or market tests |
Can ChatGPT do market research?
ChatGPT can do desk research and draft research material, but it does not replace measurement. Treat its answer as a first briefing that still needs checking.
The tasks where ChatGPT market research tends to help are practical ones:
- Orientation. Explain a new category, its segments and its common terms.
- Competitor overview. Build a first list, then verify it against company websites and filings.
- Reading at scale. Summarise reports, reviews or interview notes that you supply.
- Research design. Draft hypotheses, interview guides and survey questions.
Supply your own documents where you can. A summary of a report you can open is easier to check than a claim from an unnamed source. For the wider setup, read the guide to market intelligence.
Where does AI fail in market research?
It fails most often on specifics: exact figures, recent events and data nobody has published. These are often the details a market decision depends on, so they deserve the most checking.
Invented numbers. A model can state a market size or growth rate in a confident tone. OpenAI researchers argue that training and evaluation reward guessing over admitting uncertainty. A clean number is therefore not evidence.
The Tow Center at Columbia ran 1,600 queries across eight generative search tools in 2025. The tools gave incorrect answers to more than 60% of them.
The test asked tools to identify the source of news excerpts. It did not test market research, so the figure does not transfer directly.
The practical lesson does transfer. A link in an answer needs opening, because the page may not exist or say what the answer claims.
Stale data. The Harvard authors note that models train on static data. They may miss current conditions without regular updates. OpenAI's help pages also say answers exclude events after the training cutoff, unless a tool such as search is used.
No proprietary measurement. A general chatbot has not scanned sales, run a panel or asked your customers. It can only work from text it was given or can retrieve. If a figure comes from a paid dataset, you may see a second-hand copy of it.
A confident answer shows how the model writes. It does not show what the market did.
Can synthetic respondents replace real respondents?
Not for decisions that need precise answers. Synthetic respondents are AI personas that answer survey questions as if they were consumers. Current evidence shows they can reproduce broad patterns while distorting levels, variation and differences between groups.
Bisbee and colleagues compared ChatGPT personas with a major US election survey. The averages matched closely. Responses varied less, regression results often differed, and the same prompt gave significantly different results over three months.
Kaiser and colleagues compared digital twins built with GPT-4o against 461 US soft drink buyers. The twins matched real brand choices about 79% of the time.
They also overestimated selection of well-known brands and rated brands more positively. Attitude scores differed by an average of 1.2 points on a seven-point scale.
The authors see a possible fit for early concept tests and lower-stakes uses, not for precise decisions.
Market research studies point the same way. Brand, Israeli and Ngwe report that willingness-to-pay estimates were sometimes comparable to human studies, but often inaccurate. Some even had the wrong sign.
An earlier abstract of the paper reports that tuning with demographic data helped predict willingness to pay for new features. It used conjoint studies on household products and computers.
The authors later wrote that out-of-the-box models rated novelty higher than humans did. Tuning on real study data corrected this and brought results closer to human responses.
Their conclusion is clear: "For anything beyond early-stage high-level trend detection, human research remains essential."
Rare views are another weak spot. Nielsen Norman Group found synthetic respondents reached the same overall conclusion as humans. They collapsed onto a dozen labels, where humans used more than 50. We cover this in our article on synthetic panels.
Consistency does not rescue the method. Jahn and colleagues found that silicon samples reproduced their own outputs more faithfully than human data. A review in Psychology and Marketing sees most promise upstream, in qualitative pretesting and pilot studies.
How do you check AI market research?
Trace each important claim to a source, then test the conclusion against a measurement. Run these checks before a claim reaches a slide.
- Open every source. Confirm that the page exists and says what the answer claims.
- Find the original. Replace blog copies with the report, filing or dataset behind them.
- Check the date and scope. Note the year, the country and how the market is defined.
- Ask a second way. Rephrase the question or use another model, then compare the answers.
- Test the arithmetic. Rebuild the calculation yourself and check each input.
- Compare with behaviour. Hold the claim against sales, panel or search data.
- Label what is left. Mark each remaining statement as verified, estimated or hypothesis.
The same logic applies to synthetic data. Sarstedt and colleagues advise benchmarking against human data, using more than one model and comparing distributions as well as averages. They also expect validation to remain necessary.
ESOMAR publishes 20 questions to help buyers of AI-based services for market research and insights. Use them as a starting checklist.
Add your own test. Ask the vendor to rerun a past study from your category, then compare the output with the human result.
When should you still use real data and real people?
Use them whenever a wrong answer would be expensive or hard to reverse. Budgets, launches, pricing and market entry all belong here. So does any question about a new segment or a change that has only just started.
| Decision | AI can start with | Confirm with |
|---|---|---|
| Exploring a new category | Overview and vocabulary | Primary sources and interviews |
| Choosing concepts to test | Long list and early screening | A concept test with real buyers |
| Setting a price | Hypotheses and competitor list | Human research and market tests |
Real people explain motives, and behavioural data shows what happened. The guide to consumer insights covers the first. Measuring market demand covers the second.
Trendata publishes this guide. It measures 18.5 billion orientation moments a day. These are moments when people look into products, needs and brands, explore alternatives and prepare choices.
They form in Search, Social and AI environments. The count covers aggregated moments, not individual people or purchases. Orientation is not a sale, so the checks in this guide apply here too.
Trendata's demand forecasts run up to 12 months ahead, at category, type and specification level. They do not predict revenue or stock per SKU.
No source type answers everything, so compare them by job. The last column shows what Trendata adds to each.
| Source type | What it does well | What Trendata adds |
|---|---|---|
| Internal BI and performance data | Shows exactly what happens inside the company | The outside view, 2-6 months earlier |
| Market research and panels | In-depth answer to a question set in advance | A continuous view, including developments you had no question about yet |
| Retail and sales data | Realised sales by channel or product | Demand and needs before they become sales, also outside your own assortment |
| Free trend tools | Quick impression of interest | A structured market model, several sources, business context and forecasts up to 12 months |
| Consultancy and ad-hoc analyses | Tailored advice on a specific question | A fixed base that keeps running, which every team can question |
The 2-6 months compares with internal performance data: marketing, sales, CRM and finance. It is no guaranteed lead over competitors. The 12 months applies at category, type and specification level.
A model can only use a measured source it can reach. People question Trendata through the AI Market Analyst. The Trendata MCP server is the connection that lets AI assistants and agents query Trendata directly, in MCP-compatible clients.
Visibility inside AI answers is a separate topic, covered in generative engine optimization.
Frequently asked questions
Can AI replace market researchers?
Not based on current evidence. AI can speed up reading and drafting. People still define the question, judge the sources and decide what the results mean.
Are numbers from ChatGPT reliable?
Treat them as unverified until you open the original source. A model can present an invented or outdated figure in the same confident tone as a correct one.
What are synthetic respondents?
They are AI-generated personas that answer survey questions as if they were consumers. Researchers also call them silicon samples or digital twins.
Who are the respondents in market research?
They are the real people who answer a survey or interview. They are usually chosen to represent a defined group, such as category buyers in one country.
Sources
- Navigating the Jagged Technological Frontier (HBS AI Institute, summary of Dell'Acqua et al.)
- Larger, Faster, Cheaper: The Future of Market Research with AI (HBS AI Institute, 28 Aug 2025)
- Using LLMs for Market Research, abstract (Brand, Israeli and Ngwe, SSRN)
- Using LLMs for Market Research, working paper listing (Marketing Science Institute, 2025)
- Why Language Models Hallucinate (Kalai et al., arXiv, 2025)
- ChatGPT accuracy and limitations (OpenAI Help Center)
- We compared eight AI search engines (Jaźwińska and Chandrasekar, Tow Center, CJR, 6 Mar 2025)
- Synthetic Replacements for Human Survey Data? (Bisbee et al., Political Analysis, 2024)
- Leaving Insight to Digital Twins? (Kaiser et al., NIM Marketing Intelligence Review, 2026)
- Using large language models to generate silicon samples (Sarstedt et al., Psychology and Marketing, 2024)
- Your Next Respondent Might Be an LLM (Sarstedt et al., NIM Marketing Intelligence Review, 2026)
- Re-fielding Replicability of Silicon Sample-based Marketing Research (Jahn et al., 2026)
- Synthetic Users vs. Human Respondents (Banawa, Nielsen Norman Group, 16 Oct 2026)
- ESOMAR launches 20 Questions to Help Buyers of AI-Based Services (ESOMAR)