Description
Acme Inc is a grocery retailer deciding how hard to push its Acme Own store brand. The simulation tests how categories, claims, and price points affect trial and shopper trust.
How a simulation works.From start to finish.
See what you define, how Aaru builds and trains a simulated population, and how each result traces to the question behind it.
Set the conditions.Simulate what comes after.
Start with the decision and what success means. Choose the populations who can affect it, then write the questions that test each option.
Objectives
Curiosity has never been the barrier — it was knowing where to begin. Objectives translate this into simulations, framing the analysis and configuration: which options are under review, what success means, and which constraints the report must respect.
Acme Own store-brand simulation
Objectives 3/5
- 1Find the claim most likely to earn a first purchase
- 2Find the claim that gives loyal shoppers a reason to switch
- 3Find the price where trial starts to fall
Audiences 3
- A1Experience seekersShops 2–3 times a week across at least two stores, skews 25–44, and readily tries unfamiliar store-brand products.
- A2Value optimizersWeekly shoppers with household incomes under $80k who compare unit prices and want proof before switching.
- A3Habit loyalistsBuys 3+ times a week, two children at home, 10+ name-brand purchases this year, and no reason yet to change routine.
Questions 3
- Q1Which claims would make you try Acme Own on your next grocery trip?
- Q2How well does each Acme Own claim convey these qualities?
- Q3At which prices would you still try Acme Own instead of your usual brand?
Audiences
Each audience becomes a population of agents, incredibly broad or niche depending on use case. Its description can carry hard parameters — purchase frequency, income, children at home — alongside behavior, and selects plausible profiles where those traits connect. Agents treat each hypothetical as a condition of the simulation.
Acme Own store-brand simulation
Experience seekers
Description
Shops 2–3 times a week across at least two stores, skews 25–44, and readily tries unfamiliar store-brand products.Agent count
10,000Value optimizers
Description
Weekly shoppers with household incomes under $80k who compare unit prices and want proof before switching.Agent count
10,000Habit loyalists
Description
Buys 3+ times a week, two children at home, 10+ name-brand purchases this year, and no reason yet to change routine.Agent count
10,000Questions
Questions run against selected audiences. Use familiar formats such as single choice, multiple select, ranking, and free response. Add stimuli, hypotheticals, branching logic, and more.
Acme Own store-brand simulation
Which claims would make you try Acme Own on your next grocery trip?
How well does each Acme Own claim convey these qualities?
At which prices would you still try Acme Own instead of your usual brand?
In your own words, what would make you hesitate to try Acme Own?
Where do you first notice a grocery product you later buy?
Two similar grocery products are on the same shelf. What decides between them?
The most valuable populations to understandare always the most difficult to access.
HNW households, policymakers, decision makers, even GLP-1 users — these groups shape markets, society, and culture, but are rarely observable at scale. Through our ground-truth training, we represent every population and audience on the globe — no matter how difficult to reach.
Aaru starts with signals.
Aaru builds each world from public and licensed records of how populations live, work, and spend. Public sources describe populations, work, health, prices, and place. Licensed data adds anonymized transaction records, point-of-interest visit patterns, search demand, and media use. Customer data narrows that picture to the decision at hand. No source is used raw or alone: each is weighted against the others, so agreement strengthens a signal and disagreement flags it.
Public Data
The explorer above is a sample of the sources Aaru draws on — examples, not the full catalog. Aaru keeps each source at its available grain and cadence rather than smoothing away missing detail. Coverage gaps, irregular updates, and bias remain part of the model's limits. Each source is assessed before use, and those limits carry into the simulation and its report.
Traits move together.
A population is not a spreadsheet of independent columns. Age and work status are not independent; neither are sex and industry. Aaru models those traits jointly. Without those relationships, a population can match every one-way total and still fill classrooms with retirees or assign executive salaries to children.
Move from d = 2 to d = 6 below. Each step adds one trait and shows what the larger joint distribution has to preserve.
The participation trough
- Age
- Sex
- Economic activity
- Description
- Age and economic activity relate differently for men and women. In some countries, women’s participation falls during the years when childbearing and caregiving are most common, then rises again. The curve for men does not show the same interruption.
- Example
- Take a 32-year-old woman outside paid work while caregiving. Each trait is common on its own. The pattern is their concentration at this age. A model can match Age × Activity and Sex × Activity while spreading the same absences across the wrong ages.
- Why preserve it
- That three-way joint carries a life-course pattern no pairwise table contains. Childcare, labor, benefits, and pension models can match their headline totals and still assign the people underneath them incorrectly.
Each path above is one joint profile. A real population contains hundreds of traits, so the number of possible relationships grows combinatorially before a single agent is sampled. Many candidate interactions add no measurable structure. Aaru tests which ones change the fitted population, then keeps those relationships in the learned network. Adjust the controls below to see how the candidate space grows with trait count and interaction depth.
Traits measured
Interaction depth
0relationships through depth 3
0 cells to model through depth 3, at 8 levels per trait
One-way margins are cheap to test because there are few of them. Higher-order interactions multiply quickly. At each depth, the model has to separate relationships supported by the data from combinations that add noise.
Aaru does not store a finished population as a spreadsheet of agents. It learns a network of the relationships supported by the data, then samples complete profiles from that network. Each sampled person carries a consistent set of traits. The same network can produce populations of different sizes without copying rows or treating source records as the agents themselves.
Training works at two resolutions.
Training at only one resolution leaves a different error at each level. Fit only individual responses and each agent may look plausible while the population totals drift. Fit only aggregates and the totals can match with no coherent people underneath them. Aaru trains both at once: individual effects and population distributions are compared with observed evidence throughout the run. The model must account for the people and the total they produce, which limits how far it can rely on a pattern that works at only one level.
Training within this run0%
Population-level training
How the predicted price-band mix moves toward the observed population.
Individual-level training
How each trait changes the predicted response for one profile.
Profile · Age 32 · woman · not in paid work
A simulation does not claim to predict one person's next act. Individual behavior contains chance and circumstances the profile does not observe. Aaru estimates how responses change across a population under the stated conditions. The result is a distribution of possible behavior, not a promise about any named person.
How we validate
Validation is the accumulated evidence that a system is reliable for a specific purpose under specific conditions, including the limits of that evidence. Aaru validates against outcomes — the decisions people make in the real world, not just what they say — and traditional survey data becomes one more test of final accuracy rather than the benchmark.
Aaru builds that evidence in two places. Internal tests measure the system across datasets and settings. Each engagement then tests whether that evidence applies to the customer's world and decision.
Internal validation
Aaru's internal program withholds evidence from model development. Tests measure whether the system recovers population outcomes, preserves differences between people, and carries learned relationships across time or settings.
When intervention evidence exists, predicted changes are compared with observed changes in direction and size. The tests are rerun as data, worlds, and model components change. Results are recorded with their scope and limits.
Validation in an engagement
Each engagement starts by naming the outcomes that matter and the performance needed for the decision. Those criteria are set before the results are interpreted.
Available first-party data, historical outcomes, prior research, and live tests are compared with the world, audience, and responses. Any gaps remain part of the report rather than being folded into a general claim of accuracy.
EY tested Aaru against its Global Wealth Study.
How EY validates Aaru
EY asked Aaru to recreate its 2025 Global Wealth Study, blinded — 3,600 affluent investors across more than 30 markets. Aaru completed the run in one day, reaching a median Spearman correlation of 0.90.
Read the case studyThe simulation runs.The evidence returns.
Aaru returns interactive artifacts, not static research. Read each result by audience, compare responses, trace every finding to the question that produced it, and take the analysis straight into the decision.
Analysis
Aaru returns interactive artifacts organized by question and audience. Each finding traces back to the question that produced it, with cross-tabs for the differences beneath the topline — analysis built to be used in the decision, not filed.
Acme Own — Final Report
Generating report…
View as