Synthetic Sandbox
soonTest the campaign on your customers before you spend the media budget.
Pressure-test concepts, emails and landing copy against OCEAN-profiled versions of your own audience — built from your customer data and your compiled identity, so the personas can actually disagree with each other instead of converging on one bland consensus.
Brand Identity
v14 bundle
voice · principles · persona · colors (hex + RGB) · logo system
Synthetic Sandbox reads
Capabilities
What it does
Profiles, not demographics
Segments are built from your own customer research and scored on the Big Five, the most validated model of human personality there is. A profile is a temperament, not an age bracket.
Reactions that differ
Generic persona tools assemble people from public data, which is why frontier models keep producing three phrasings of the same consensus. Profiles built from different real people produce different answers — including the awkward ones.
The layer before research
Run forty concepts synthetically in an afternoon, kill the thirty-seven that land badly, and take the survivors to a small real sample. Cheaper filtering, not cheaper truth.
Tested against your brand, not a blank model
The concept is evaluated as your brand saying it. A line that works for a plain-spoken brand fails for a formal one, and the test knows the difference because it reads your compiled identity.
Segment disagreement, surfaced
The most useful output is rarely the score. It is the segment that reacted differently from the rest, and why.
Brand identity inside
It reads your identity. It does not redefine it.
A test is only as good as the profile it runs against. Sandbox reads the brand's compiled identity — personality, voice, principles, the words the brand avoids — and pairs it with the audience profiles, because the product is not the brand or the customer alone. It is the match.
See the contractMethod
How the method works, and what it cannot do
The evidence for synthetic respondents is real but bounded. A 2024 Stanford study simulated a thousand real individuals as language-model personas and reproduced their attitudes more accurately than demographic baselines; Google researchers in 2026 covered over 80% of the range of human opinion on unseen topics. A 2026 controlled comparison against 117 real interviews found the opposite failure: three frontier models converging on identical modal answers, giving you one consensus in three phrasings rather than genuine disagreement. Our answer to that is conditioning — profiles built from a brand's own customers, paired with that brand's compiled identity, rather than a population average.
The two-sided modelWhat it does not do
- This is not a replacement for research. It is the filter that runs before it.
- Every study should be validated against a small real sample. We will tell you when a result is too close to call.
- Synthetic respondents under-represent outliers. Treat strong consensus with suspicion, and treat segment disagreement as the signal.
- The quality of the output is capped by the quality of the customer data going in. No profile data, no credible test.
Early access
Want Synthetic Sandbox on your brand?
We onboard the first brands by hand, identity first. Tell us what you need and we will reply personally.