The replication
One study, run twice.
We took a published consumer study that ran for six months in the field, and ran the same study on a simulated panel — overnight.
We describe the study rather than name it. The point isn't whose study it was; it's whether a simulation can land where months of fieldwork landed, on the same questions, against the same population definition.
The setup
Same questions. Same population. No peeking.
Rebuilt, not borrowed
We rebuilt the study's population definition as a simulated sample — the same market, segments, and quotas the fieldwork used.
Thousands of agents, in parallel
The simulated panel took the study the way Murmur runs every study — agents acting on habit, price sensitivity, and social influence, not filling in a questionnaire.
Results first, then the answer key
The simulation was run and locked before its results were placed next to the published findings.
Results
Murmur's simulated panel matched the study's findings within two points.
Same questions. Same population definition. A fraction of the time. A full comparison chart — including where the simulation diverged and why — is being prepared alongside the sample report.
Limitations
What simulated panels aren't good at. Yet.
Sensory research. Murmur can't taste, smell, or feel. Flavor panels, fragrance work, texture testing — that's real product in real hands, and we'll tell you so before you spend anything.
Categories with no behavioral history. Agents are built from how people have actually behaved. If a product is genuinely unlike anything consumers have encountered, there's little behavior to model — and fieldwork still wins.
Candor is the deal: when a study fits either case, we say so at the brief stage, not after the invoice.
Hard questions
The questions you should be asking.
How can simulated respondents represent real consumers?
Fair question — it's the question. Murmur's agents aren't invented characters; they're constructed from population-level behavioral data and calibrated against real-world benchmarks before any study runs.
The proof standard we hold ourselves to is replication — the study above — not promises. And where the behavioral data for a segment is thin, representation gets harder, which is why sample design happens with you, in the open.
What stops the model from just making answers up?
Nothing stops a single model output from being wrong — which is exactly why a Murmur result is never one model's opinion. Results are distributions across thousands of agents; outliers are inspectable rather than hidden inside an average.
Every study reports its confidence honestly. When the panel is genuinely split, the report says so instead of manufacturing a winner.
How do you handle bias in the underlying models?
The models Murmur builds on carry biases — pretending otherwise would be the real risk. We measure agent behavior against known population benchmarks and correct where the panel drifts from reality.
Residual bias exists. The honest claim is that we measure and reduce it, not that we've eliminated it.
How is the simulated sample matched to my market?
Sample design is the first step of every study, done with you at the brief stage: market, category, segments, quotas. You sign off on the population definition before a single agent runs — the same discipline you'd apply to a fieldwork sample.
When should we NOT use Murmur?
When the question is sensory — taste, smell, texture — put real product in real hands. When a category is so new there's no behavioral history to build agents from, fieldwork still wins.
If your study fits either case, we'll say so before you spend anything.
How does this fit alongside our existing research program?
As a complement and a pre-filter, not a rip-and-replace. Run Murmur first: kill the weak concepts, tighten the price ladder, find the segment worth recruiting for.
Then field only what still needs fielding — with sharper questions and a smaller bill.
Run the comparison on your own study.
Bring us a study you've already fielded — or one you're about to. We'll run it on the simulated panel, side by side, and you judge the gap.
Send us a brief