01Data
What the models are built from
Some data records what people say, some records who they are, and some records what they do. Most providers use a mix, and the providers page lists which each uses.
| Data | Records | Strength | Limit | Amount in one published study |
|---|---|---|---|---|
| Surveys and panels | What people say | Answers to known questions from known people. | Covers only the questions that were asked. | 500 questions from each of 2,058 people, about 2.4 hours each Toubia et al., 2025 |
| Interviews | What people say | A long conversation with one person, in their own words. | Slow and costly to collect from many people. | A two-hour interview with each of 1,052 people, about 6,500 words each Park et al. (Simile), 2024 (v1) |
| Public text | What people say | Reviews, posts and forum threads, written in people’s own words without a researcher present. | Covers only what people choose to post. The writer’s age, location and so on usually have to be estimated. | Not published |
| Census and official statistics | Who people are | The make-up of a population by age, income, place and household. | Says little about attitudes. | Published for whole populations |
| Behavioural records | What people do | Purchases, store visits, app use and clicks. These show what people do, not what they say. | Usually has to be licensed. | 58,466 people’s Facebook Likes, a median of 68 each, used to predict traits such as age and politics Kosinski et al., 2013 |
| Experiment responses | What people do | What people did in controlled experiments, including how much each change affected them. | Covers only the treatments that were tested. | 2.9 million responses from 400,491 people in 210 experiments Kolluri et al. (Simile), 2025 |
| A customer’s own data | Any of these | Research, support conversations or panels the organisation already has. It comes from the people the question is about. | Often a small sample. | Not published |
| Only what a language model learnt in training | Nothing collected | Any audience can be described in a prompt. | The answers reflect the model’s stereotypes of that audience. Wang et al., 2025 | None |
More data about each person does not always help. In one mega-study, digital twins given each person’s answers to 500 questions scored 0.748, against 0.746 for twins given only demographics. Peng et al., 2025 The Accuracy chapter covers the published critiques of each method.