Concept Test
In a 2025 craft-beer concept test, 54.7% of beer-buying consumers were extremely or very likely to purchase X beer. What was most closely associated with that intent — and how does the picture change for a different income bracket, region, or level of brand awareness? Seven interactive tools let you explore.
Built in partnership with EMI Research Solutions. The data on this page come from a craft-beer concept test conducted by EMI Research Solutions, an established commercial-research firm. EMI delivers Mirror engagements to its clients under its own brand, in partnership with Electric Insights, using the same analytical engine demonstrated across this site.
Ask the data
Ask a question in plain English — the right tool opens below and runs the analysis for you.
Follow-ups refer to: “”
Every number is computed by the statistical engine — never by AI.
- Your question becomes an analysis specification, shown to you before anything runs — change any part of it in the tool below.
- Ambiguous questions get choices; questions this data can’t answer get a plain refusal naming what it can answer.
- Asked what the model might be missing, Ask tests a synthetic variable and points you to the Variable-Ranking Tool for the measured variables; if none would add to the model, the AI proposes what an unmeasured factor might be. Once you’ve run a few analyses you can summarize them on one page.
- AI writes the routing, these short messages, and clearly labeled drafts. The engine computes every number and runs every check.
Summarize this session
One page covering the 2 analyses with captured results (marked above); replay any other recent question to include it. Each number links to the tool that computed it; AI writes only the connecting text.
Results run below in the tool itself; you can change any of the specification there. Everything is computed by the platform’s statistical engine, not by AI.
Analysis tool · the tool itself, running here
Running…
Hypotheses to measure
Written by AI for a future survey wave, within this survey's era and inventory — candidate constructs to consider measuring, not findings from this data. The human decides; only a new wave can measure them.
Summary report
Plain-language version, drafted by AI from the same runs and held to the same check as the standard one.
Composed by AI from the runs listed under each section. A server check keeps only figures that appear in those runs' questions or results; open the linked tool to confirm what each figure measures.
Seven Ways to Engage with the Data
Two simulators and a guided walkthrough open up the published 6-variable model. A nonresponse stress test challenges the headline, while Model Checks probe the result through a structurally different observational estimate and a Bayesian network representation of conditional dependencies among measured variables. An explorer and a model builder let anyone go further — investigate any of six outcomes in the data and build competing models, whether to challenge the published purchase-intent account or to study outcomes the published model doesn't address.
All-or-Nothing Simulator
Pin every respondent to one chosen response level per package factor and see how the 54.7% top-2 purchase intent shifts. Runs on the published model.
Try itFine-Tuning Simulator
Same published model, but redistribute response shares gradually rather than pinning. Watch how small distributional changes move the purchase-intent needle.
Try itInside the Model
A guided walkthrough of how the published model turns one respondent's six answers into a model-implied likely-to-buy probability — and how those individual probabilities aggregate into the headline top-2-box percentage.
Try itNon-Response Test
Specify a nonresponse pattern and see whether the 54.7% top-2 purchase intent would survive it. A vulnerability check on the released score, separate from the explanatory model.
Try itModel Checks
Two exploratory checks on the model behind the headline. A triangulation check estimates the same scenario through a structurally different observational route, helping assess dependence on the published model form; a Bayesian network analysis maps conditional dependencies among measured variables under a different structural representation. Neither check establishes causal direction.
Try itSurvey Explorer
Browse the raw responses. See how consumers reacted to each package factor, which questions moved together, and how one group compared to another. Frequencies, cross-tabs, and correlations — against any of the six outcomes.
Try itModel Builder
Pick any of six outcomes — purchase intent (a top-2-box cut or the full ordered 5-point scale), package appeal, brand fit, high-quality, premium, or most-often-purchased brand (a seven-way multinomial) — and refit with whichever predictors you choose. Add or remove items, refit on a subgroup, or let Auto-Build find the best-performing combination under its selected rule. See where the published purchase-intent account holds up — or model an outcome it doesn't cover.
Try itAbout Concept Testing
How a concept test works
1. Show the concept
Respondents see a package design, ad mockup, or product description.
2. Capture reactions
A battery of questions probes appeal, fit, relevance, and intent — typically on 5-point scales.
3. Model what matters — then test it
Statistical modeling shows which measured reactions remain associated with purchase intent when considered together. This example lets you go further — refit on subgroups, swap predictors, and stress-test the headline against nonresponse.
Package design
The visual concept shown to respondents
Purchase intent
5-point scale, top-2 box modeled
Attribute reactions
Twenty package-perception predictors
Subgroup cuts
Eight subgroup variables for refits
The Concept Test
Beer-category respondents were shown a new package concept and asked a battery of reaction questions. Purchase intent was captured on a 5-point scale; the top-2 box (Extremely / Very likely) is the modeled outcome.
The published logistic model was cross-checked against Stata 18 to four decimal places.
What the survey measured:
Building on a concept test of your own?
The platform supports survey outcomes including binary cuts, ordered rating scales, multi-category choices, and staged funnels, with categorical or continuous predictors. Read the overview for brand and insights teams, or see how engagements are structured.