Glossary

A single reference for the terms used across this documentation and in the product.

Core objects

Seed data — the high-quality survey data an audience is built from. Each respondent in the seed data carries demographic, cultural and behavioural characteristics, for example location, income, education, household composition, values, attitudes, lifestyle preferences, media consumption, or purchase history. Seed data is the ground truth for everything Merlin predicts.

Persona — a model of one individual respondent from the seed data. Each persona has its own coherent identity and response patterns derived from that person’s data profile, and holds that identity consistently across questions and conversations. Sometimes also called an agent.

Audience — a group of synthetic personas, defined by seed data, you can query on demand. An audience may be your whole population, or a subset defined by specific criteria — “under 25 and living in London”, “price-sensitive shoppers”, “concerned parents”. Audiences are the unit you survey, compare and analyse. (Some parts of the product still label this a segment; the two mean the same thing.)

Working in the product

Project — an optional grouping for studies you want to keep together or share as a set. You don't need one to start a study. See Concepts.

Study — the workspace where analysis happens. Results — plots, surveys, focus groups — accumulate in a feed under the Results tab, with an Insights tab alongside it. See Your first study.

Tool — any of the research tools you run from the toolbar at the bottom of a study: Plot, Survey, Focus Group.

Result Card — a single result in a study feed. Cards carry their own configuration, so the … menu can re-run a variant with Ask again, change the chart type, or export that one result.

Stimulus — an image or piece of text attached to a question, either as context for the question or as an answer option to choose between. Optional.

Breakdown — an optional second question layered onto a plot, so you can cut one answer by another. See Plots.

Template — a saved survey configuration, stored against your account or shared across your organisation, so a frequently-run question set can be reapplied to new stimuli.

Agent — Merlin’s research partner. Rather than putting your question straight to personas, the agent translates your input into well-framed research questions first.

Research outputs

Plot — a chart of how an audience answered a question already in the seed data. The numbers are counts, not predictions. See Plots.

Survey — a new question, not present in the seed data, put to one or more audiences. Merlin predicts how the audience the survey question is directed to would answer.

Focus group — a simulated conversation with a small panel of personas drawn from an audience, used to explore why rather than what.

Debate — the focus group mode where an AI moderator runs the session and pushes personas to surface the full range of views on a topic.

Audience analysis — what an audience page shows once you run Analyse audience: a written audience profile, its top influences, and the same characteristics grouped by theme. See Audience analysis.

Top influences — the traits that most strongly separate an audience from everyone else, ranked by statistical significance and derived from regression analysis against the rest of the population.

Audience explorer — the panel that opens when you click answers on a result, showing what you’ve selected, how many respondents match, and a Save audience button.

Confidence and quality

Confidence score — the A / B / C grade shown on a survey result, indicating how much weight to put on it. See Confidence grades.

Source alignment — one of the two signals behind the confidence score: how close the question you asked is to the seed data your audience was built from. Sometimes referred to as coverage.

Consistency — the other signal: whether the model gave broadly the same answer when the question was re-run.

Evaluation run — a validation test comparing Merlin’s output against known real-world data. More than 30,000 have informed the current architecture.

Hold-out test — a validation method where real research data is withheld, Merlin is asked the same question, and the two sets of responses are compared. See How we know it works.

NDAM — the metric behind the headline hold-out figure, measuring how closely a predicted response distribution matches the real one.

Human noise level — how closely real people agree with themselves: the score you get by asking the same person the same question twice in one survey, which sits at 94%. It is the practical ceiling for any method, Merlin included.