# iMario vs Simile: Choosing How to Model an Audience's Response
> A buyer's guide for product, insights and research leaders: what iMario and Simile let a team see when it models an audience's response, and what the team can responsibly conclude from it.
Source: https://imario.ai/blog/imario-vs-simile
Published: 2026-10-12
Author: iMario Team
Picture a team at a beverage company preparing to launch a ready-to-drink tea. It has two product descriptions and two possible price points, and it wants to know how 18–34-year-old urban tea drinkers will react before committing. The scenario is invented for this article and neither product has been run on it, but it is the kind of decision that puts synthetic research on a shortlist.
iMario and Simile both build synthetic audiences from data about real people, and both can work with a company's own customer data. Where they put the emphasis differs.
With iMario, the work starts with the audience. When a company has relevant customer evidence, such as records, surveys or interviews, we anchor the synthetic audience in it, and our population data fills gaps or extends coverage beyond current customers.[^fromdata][^im20] Before a study runs, the team can see which sources the audience rests on and what they cover. After it runs, the team can see which synthetic respondents chose an answer, read their individual answers, compare subgroups and ask follow-up questions.[^fromimario][^explorer][^followup] The practical advantage is traceability: a team can inspect the real evidence its audience is built on and examine which synthetic respondents' answers support a finding.
Simile's public materials emphasize modeling behavior under specified conditions. A query names a population, a situation and the available actions, and the model returns a distribution of likely responses.[^conf] Simile builds its populations from large-scale studies and proprietary behavioral data, and customers can supply their own behavioral data to train a custom model.[^sim] Each result carries a confidence label based on a predicted error.[^conf][^sim]
## Starting with the audience
For the tea team, the first question is what "18–34-year-old urban tea drinkers" means in each product, and how much of that audience the team can inspect before trusting an answer.
With iMario, a company's customer records, or another usable roster, establish who is in the audience. Surveys, interviews and support conversations add evidence about those people when they can be reliably linked to them, for example through a stable customer ID or email; a transcript on its own does not necessarily establish who belongs in the audience.[^fromdata] The sources can be connected read-only or uploaded, and before the audience is created, the team checks how they match up, such as which column holds the customer ID and which transcript belongs to whom.[^im20] We then build synthetic individuals anchored to the matched people. Once the audience exists, the team can compare its definition and distribution with the source data, and inspect source coverage and example people.[^fromdata] Where a company has no relevant customer evidence, it can start from our population data instead, which is built from official statistics and calibrated against real survey data;[^im20] that audience represents the population it is drawn from, not the company's own customers, and the team reviews its definition, source coverage and any criteria the population can't support before using it.[^fromimario]
The tea team has no customers for this new product yet, but it may still hold relevant evidence: earlier tea surveys, research with buyers of its other drinks, or support conversations about adjacent products. Where that material fits, it anchors the audience, and our population data extends coverage to tea drinkers who aren't yet its customers. If none of it fits, the audience describes 18–34-year-old urban tea drinkers in general, not this company's future buyers. In both cases, the team can see in iMario how "urban tea drinker" is established, which criteria are supported and where the evidence is thin.
Simile also says "every Simile population starts with real people." It runs large-scale studies, combines the results with proprietary data on human behavior, tracks how fresh each type of data is, and re-collects data when it detects drift. Customers can supply their own behavioral data, such as loyalty data, balance histories and telemetry; according to Simile, that data trains a custom model, which is then checked against how those customers actually behave.[^sim] A case study with CVS Health, published by Simile, describes agents built on 2.9 million consented responses from more than 400,000 participants across more than 200 behavioral scenarios.[^cvs] Simile's public pages don't show whether a buyer can inspect a population's composition or see which criteria it couldn't support, so that is worth asking to see in a demo. A company with substantial behavioral data, such as transaction histories or usage logs, should ask both vendors how that data enters the model and what it leaves unsupported.
## Working with the result
The first answer is rarely the end of a study, and the next steps differ between the two products.
Simile's public examples show the shape of its output: a question put to a population, the modeled distribution of answers, a one-line takeaway and a confidence label. One retail banking example asks who would leave if switching banks took ten minutes and shows 31% who would switch, 44% satisfied but still switchable and 25% likely to stay, labeled "High confidence." Another, about electric-vehicle considerers, ranks barriers led by charging access and carries a "Moderately high confidence" label.[^sim] For the tea team, these suggest a demo question: can Simile model the four description-and-price combinations and return a distribution and confidence label for each? Simile's site also quotes Wealthfront's head of user research, who says that speaking with its simulated customers helps the team find areas for further research and that Simile expanded its qualitative research scope 15-fold.[^sim] The pages we reviewed don't describe how a user moves from a distribution to the simulated respondents within it, or what can be exported.
With iMario, that next step is where most of the work happens. A researcher picks a format and gets a result card for each question. The format can be a survey, an interview with follow-up probes, a focus group, a preference test, a creative test, feedback on a live web page or a PR test.[^methods] Clicking an answer's bar in a result card opens Explorer, which shows the synthetic respondents who chose it. From there the researcher can:
- combine answers into a group
- break that group down by another question or a demographic
- read each synthetic respondent's answers, or send one of them a follow-up message
- save the group as a segment, or download its rows as a CSV[^explorer]
A new question can then go to just that group.[^followup] Reports pull several runs together, with tabs for each question's analysis, the individual synthetic responses and the code book showing how answers were classified. They export to PDF, Markdown or PowerPoint.[^reports][^share]
Suppose the tea team finds that one description wins overall, but a sizable group of synthetic respondents worries that the drink tastes artificial. With iMario, the team would select that group, check whether it skews toward lighter tea drinkers and ask it what would change its mind, ending with a coded report a stakeholder can review. A per-result confidence label serves a different purpose: it helps a team decide which of several comparable answers to check first.
## Interpreting the evidence
A synthetic study involves three distinct kinds of evidence, and confusing them leads to wrong conclusions.
**The evidence an audience is built on.** With iMario, this is the company's existing customer evidence, where it has some, together with our population data, which is modeled from official statistics and calibrated against survey data.[^fromdata][^im20] The customer evidence is real information about real people, collected before the study.
**The answers in a study.** Answers to a new study are generated by synthetic individuals. Existing evidence about the real customer behind a synthetic individual may inform its answer, but that answer is not a new reply from the customer to the study's question. Our evidence guide calls a modelled answer "a simulated response to the supplied question and context," not an observed action or a guaranteed outcome, and our guidance asks readers to keep observation, inference and simulation apart and to note where evidence is limited.[^evid][^overview] In an audience built from a company's own data, the team can open the audience and read what its real customers said in the interviews behind it, so a synthetic answer can be set against the customer evidence it was built on.[^fromdata][^im20]
**Optional human validation.** A team may also choose to compare a particular study with newly collected responses from real people when a finding needs checking. This is a separate step, not something that happens automatically in every study.[^validate] Our pricing page lists a "validation loop with real people interviews,"[^pricing] and our validation guide describes how to run a comparison:
- keep the decision, audience definition, question or stimulus and outcome measure comparable to the modelled study
- decide in advance what result would support, revise or reject the working hypothesis
- read the modelled result and the observed responses side by side, recording where the audiences, context, wording, sample source or outcome measure differ[^validate]
The guide treats a disagreement as a possible coverage or modelling limitation rather than proof that either source is correct, and suggests using it to refine the audience, material or question before a new run. It leaves the choice of method and provider to the team, from interviews and surveys to a usability session or a controlled product or message test.[^validate] A team that wants interviews arranged for a validation should ask us how that would work for its study.
Simile reports reliability at the level of the model. Each result's confidence label comes from a confidence model trained to predict how far the simulated distribution is likely to sit from the real one. Simile treats a result as decision-quality when that distance, measured as total variation distance (TVD), is under 0.16, a threshold drawn from an internal rating exercise that Simile says it expects to revise. On about 8,600 held-out questions, Simile estimates that a result labeled High is 95% likely to meet the threshold and a result labeled Low 38% likely.[^conf] It also reports more than 7,000 evaluations against real people each week, which it uses to train the confidence model.[^sim] These are ongoing vendor evaluations across many questions, not a buyer checking one study, and we did not find the underlying evaluation data published. Simile's published CVS Health case describes simulated results being used to prioritize and validate downstream pilots and experiments;[^cvs] the pages we reviewed don't say whether Simile offers a step for comparing a particular study with a team's own real-world responses.
Our model-level evidence is a public benchmark that anyone can rerun. It compares synthetic and real answer distributions on 922 questions about politics, values and social attitudes, across 11 populations and about 47,000 respondents, and scores every question blind. We score 88.9% on 1-TV, a measure of how much the two distributions overlap, against an estimated ceiling of about 93% for how closely real samples reproduce one another. The data and scoring code are public.[^bench]
The two figures share a scale, since 1-TV is one minus the total variation distance: our 88.9% is an average distance of about 0.11, under the 0.16 that Simile uses as its bar for a decision-quality result. They still can't be ranked directly. One is a pass mark for a single result, the other an average over a fixed set of questions, and the question sets differ. Neither figure says how close either product will come to real tea drinkers reacting to a new drink at a new price, which is what the pilot below is for.
| | What it shows | What it cannot establish |
|---|---|---|
| Simile confidence label | How likely one result is to fall within the 0.16 TVD threshold, based on Simile's own evaluation | Calibration on the buyer's category without further testing |
| iMario audience sources | For audiences built from company data, which customer records and interviews the audience rests on, how much of it they cover, and what the real customers said | How close a new synthetic answer is to what a real person would say |
| iMario benchmark | Average agreement with real distributions on a fixed set of attitude questions | Accuracy for a new product, stimulus or audience |
| Optional study validation with iMario | How one modelled result compares with newly collected responses gathered under comparable conditions | Accuracy for other studies or categories; a disagreement alone doesn't show which source is right |
## Testing before you commit
A pilot can give a bounded comparison on one decision, provided the real-consumer sample is large enough for the comparison being made.
The tea team would start by choosing a market both products cover. The scenario doesn't name one, so the team should confirm coverage of its target market and audience with both vendors. The Simile pages we reviewed don't state geographic coverage, and its methodology's worked example is "Floridians who order food delivery."[^sim][^conf] The team would then define the closest corresponding audience in each product and record what each audience is built from. Next, it would run the same descriptions, prices, wording, answer options and pack image in both. Testing all four description-and-price combinations separates the effect of the message from the effect of the price.
The real-world check is a compact survey of real target consumers, randomly assigned to the four combinations, whose responses stay hidden from both products until both synthetic outputs exist. The sample size depends on the smallest difference the team wants to detect and the precision it needs; if the sample can't support a reliable four-way comparison, its results are directional, not decisive. If the team uses our validation workflow for this, it should apply the same real sample and outcome criteria to both products' outputs, since running the comparison through one vendor's workflow is not evidence that its predictions are closer.
Start with the comparison. Did each product favor the same description as the real respondents? Did each show the same change in stated interest at the higher price, or the same lack of one? Before a mismatch counts against a product, look at the real sample's margin of error and at how closely the audiences matched. The team can also note the confidence label on each of Simile's answers, and check whether a reviewer can trace iMario's main findings back to the audience's sources and to the synthetic respondents' answers behind them.
A pilot like this supports a narrow conclusion: how useful each product looked for this audience, this material and these decisions. It can't show whether Simile's labels are calibrated, which takes many questions with known answers, or that either product will perform the same way in another category. Stated interest at a higher price is also not a purchase.
## Deciding where to start
Start with iMario when the research depends on inspecting the evidence an audience is built on and examining which synthetic respondents' answers support a finding. That suits a team that wants its own customer evidence to anchor the audience, with our population data, 5.5 billion simulated people across 61 markets, filling the gaps,[^imhome] and wants to review coverage before relying on it. It also suits a team whose deliverable is a research readout: surveys, probing interviews, focus groups, creative or page tests, subgroup drill-down, follow-up questions and a coded, exportable report that includes the individual synthetic responses. We also document an optional workflow for comparing a particular study with newly collected human responses, and publish a benchmark anyone can rerun.[^validate][^bench]
Start with Simile when the decision turns on how a population's behavior changes under different conditions, when a team wants its behavioral data used to train a model checked against observed outcomes, or when a per-result reliability signal is central to how decisions get made. That is where Simile's public materials put their emphasis.
Our pricing is public. Credits cost $0.05 each in one-time purchases, and monthly packs are 20% off. Each research format has its own credit rate. Reports, AI analysis of a selected group, and follow-up messages to individual synthetic respondents are also charged in credits. Simulating a synthetic audience and drafting surveys and discussion guides are free.[^pricing] Simile showed no published pricing on the pages we checked.[^sim]
On the question many buyers care about most, which product will be more accurate for their category and audience, the answer comes from a pilot on the buyer's own decision, not from either company's headline figure. Before committing, ask each vendor for:
1. evaluation results on questions close to your own category, with the methodology behind them
2. a live demonstration of building your target audience and of what happens after the first answer
3. a quote for a pilot designed around a real decision
Whichever product you start with, the pilot is where the choice gets made: a comparison against real people on one of your own decisions shows how useful each product is for that decision.
---
## Sources
All sources accessed October 2026. Vendor sources reflect each company's own claims.
[^sim]: Simile, product site: (vendor; includes example outputs and customer testimonials published by Simile). No pricing appeared on the pages checked; returned "not found".
[^conf]: Simile, confidence methodology post, 25 August 2026: (vendor-published evaluation).
[^cvs]: Simile, "Simile × CVS Health," 24 February 2026: (vendor-published case study).
[^imhome]: iMario, homepage: (vendor).
[^im20]: iMario, "iMario 2.0: The Human Voice Behind Every Decision," 15 September 2026: (vendor).
[^fromimario]: iMario docs, "Create an Audience from iMario": (vendor documentation).
[^fromdata]: iMario docs, "Build an Audience from Your Data": (vendor documentation).
[^methods]: iMario docs, "Choose a Research Method": (vendor documentation).
[^explorer]: iMario docs, "Explore Results": (vendor documentation).
[^followup]: iMario docs, "Run, Review and Follow Up": (vendor documentation).
[^reports]: iMario docs, "Generate and Read a Report": (vendor documentation).
[^share]: iMario docs, "Share and Export": (vendor documentation).
[^evid]: iMario docs, "Interpret Evidence": (vendor documentation).
[^overview]: iMario docs, "From Results to Decisions": (vendor documentation).
[^bench]: iMario Accuracy Benchmark v2: (vendor-published; data and scoring code public; methodology notes state the ~93% human-reproduction ceiling is estimated from Pew 2024→2025 same-question overlap and applied pool-wide).
[^pricing]: iMario, Pricing: (vendor).
[^validate]: iMario docs, "Validate with real world data": (vendor documentation).