Stated, observed or synthetic: which consumer data should you trust in the AI era?

Stated, observed or synthetic: which consumer data should you trust in the AI era?
A field guide to the three kinds of consumer data, and how to combine them
Written by: Mark Nguyen, ND Research
Originally Published: November 21, 2026
Labels: Essential Reading, AI-assisted, Continuously Maintained
Topics: Synthetic data, Observed data, Survey research, Consumer insights, Research design, AI
TLDR: Stated data tells you what people say. Observed data tells you what actually happened. Synthetic data tells you what a model predicts people would say. All three are useful. Only observed data can anchor the other two.
Key Ideas:
- Almost every consumer dataset falls into one of three types: stated (people tell you), observed (you record what happened) or synthetic (a model generates it).
- Each type answers a different question. Stated data is best for the why, observed data for the what, synthetic data for fast exploration.
- AI has made synthetic data cheap and stated data faster to collect. It hasn't made observed data any easier to get, which is why it is becoming the scarce input.
- A sound research design in 2026 uses observed data as the anchor, synthetic data as a multiplier and stated data to explain motive.
- Personalized feeds are a growing blind spot for all three types, unless you observe them directly.
Preamble
In our conversations with research buyers this year, one confusion kept coming up: people would ask whether a new tool was "real data" without a shared definition of what that meant. A synthetic panel, a survey and a purchase log all show up in a deck as numbers on a chart. They are not the same kind of evidence.
This is the simple framework we use internally. It's not original (methodologists have drawn these lines for decades), but AI has made it more urgent.
What is stated data?
Stated data is anything people tell you: surveys, interviews, focus groups, diaries, reviews. Its strength is access to motive, emotion and meaning. Its weakness is that people are unreliable narrators of their own behaviour. They forget, round off, say what sounds reasonable or simply don't know why they did something.
Best for: the why, attitudes, language, unmet needs.
Weak at: exact behaviour, frequency, anything people don't consciously notice (like what their feed showed them).
What is observed data?
Observed data records what actually happened: purchases, clicks, app usage, location visits, ethnographic observation and, in our case, the posts a personalized feed actually served someone. Its strength is that it doesn't depend on memory. Its weakness is that it shows the what without the why, and it is usually hard and expensive to collect ethically.
Best for: real behaviour and exposure, ground truth, validating other data.
Weak at: motive, and anything outside what was recorded.
What is synthetic data?
Synthetic data is generated by a model: AI personas, synthetic respondents and digital twins that simulate how a person or audience might answer [1]. Its strength is speed and cost. You can test fifty concepts before lunch. Its weakness is that it can only recombine patterns from the data it learned from, so it can't tell you something genuinely new about people.
Best for: early exploration, hypothesis generation, boosting small segments inside a large enough real sample.
Weak at: novel behaviour, fast-changing culture, anything nobody has observed yet.
The vendors we spoke with in 2026 were candid about this. They drew their own lines: use it to boost small segments, don't use it to model what drives choices, and don't invent what nobody observed [2].
Why does the difference matter more now?
Because AI changed the cost of each type unevenly.
- Synthetic data went from near-impossible to nearly free.
- Stated data got faster and cheaper to collect and analyse (AI moderators, automated coding).
- Observed data didn't get much easier. You still need consent, access and real people.
When two inputs get cheap and one doesn't, the expensive one sets the quality ceiling. That's the same pattern we describe in our calculator thesis: when answers get cheap, real understanding becomes the scarce thing [3].
How should you combine them?
Our rule of thumb:
- Anchor on observed. Start from what actually happened, even if the sample is small.
- Explain with stated. Ask people about the behaviour you observed, not about behaviour in the abstract.
- Multiply with synthetic. Use models to extend and stress-test patterns that are grounded in the first two, never to replace them.
Think: you want to know why frequent buyers of a snack brand respond to certain creators. Observed feed data shows which creators actually appear in those buyers' feeds. Interviews tell you what those creators mean to them. A synthetic panel, calibrated on both, lets you test new creator briefs quickly.
Where do social feeds fit?
Feeds are the biggest blind spot in this framework. People can't reliably state what their feed showed them. Social listening observes what people publicly post, not what they're privately served. And synthetic models have no access to an individual's personalized recommendations.
That's the gap our product, Tapestry, is built for. Consented participants share screenshots of their own signed-in feeds, and AI classifies each post so the feed itself becomes observed data, linked to who the person is. We call this feed exposure intelligence [4].
A quick checklist before you trust a dataset
- Is this stated, observed or synthetic? (If it's a mix, which parts are which?)
- If synthetic: what real data was it calibrated on, and how recent is it?
- If stated: are people describing something they could actually remember?
- If observed: who consented, what was missed, and how was it sampled?
- Is the finding tied to a decision you actually need to make?
FAQ
Is synthetic data "real data"?
It's real model output, not real observation of people. It can be useful, but it's only as good as the observed and stated data underneath it.
Can synthetic respondents replace surveys?
Not on their own, in our view, and not according to most of the synthetic vendors we spoke with in 2026 [2]. They're best used alongside real samples.
What is the most reliable type of consumer data?
For behaviour, observed data. For motive, stated data. Reliability depends on the question you're asking.
Who is ND Research?
A Toronto consumer research company building Tapestry, a feed exposure intelligence platform, and running qualitative and quantitative research for brands, agencies and technology companies.
How to cite this
Nguyen, M. (2026, November 21). Stated, observed or synthetic: which consumer data should you trust in the AI era? ND Research. https://www.nguyendigital.co/blog-posts/stated-observed-synthetic-consumer-data
List of References
(Continuously Maintained)
[1] Greenbook. How AI Is Changing Market Research in 2026: What Practitioners Need to Know. https://www.greenbook.org/insights/artificial-intelligence-and-machine-learning/how-ai-is-changing-market-research-in-2026-what-practitioners-need-to-know
[2] ND Research. Nguyen, M. (2026, October 23). Will synthetic respondents replace real consumer research? https://www.nguyendigital.co/blog-posts/will-synthetic-respondents-replace-consumer-research
[3] ND Research. Nguyen, M. (2026, November 21). What comes after AI in consumer research? The calculator thesis. https://www.nguyendigital.co/blog-posts/what-comes-after-ai-consumer-research
[4] ND Research. Nguyen, M. (2026, October 23). What is feed exposure intelligence? https://www.nguyendigital.co/blog-posts/what-is-feed-exposure-intelligence





