“I might try Mango Burst once if I see it on quick commerce, but I'd compare it with what I already buy before deciding.”
Shrota puts your concept in front of a synthetic Indian cohort, lets them answer in their own words, and measures what they said. A run takes an afternoon. You will know which of your forty ideas deserve a real focus group.
Not because the other thirty-seven were bad. Because the budget ran out at three, and the calendar ran out before the season did. The ideas you never tested are the expensive part — and you will never see the bill for them.
Name, price, pack size, claims, benefits, the occasion you are buying into. What you would put on a concept card.
Personas are sampled from India-representative pools — city tier, income conditioned on tier, language, category habit. Same seed, same cohort, every run.
Every persona responds in natural language, in character, with their own budget and scepticism. Nobody is asked to rate anything.
Each answer is compared against calibrated Likert anchors and resolved into a distribution across the 1–5 scale.
Scores roll up by construct and by segment — who leans in, who backs away, and where the concept loses them.
Ask a language model to rate something one to five and it will cluster on four and avoid the ends — the same scale-use artefacts you get from a bored human panel. So Shrota asks for prose, then measures the prose. The score is derived, never volunteered.
γ(r) = cos(response, anchor_r) p(r) ∝ γ(r) − γ(ℓ), ℓ = argmin γ score = Σ p(r) · r ∈ [1, 5]
Embed the answer, measure its cosine similarity to five calibrated anchor statements, rebase on the least-similar anchor, and take the expectation. After Maier et al., “LLMs Reproduce Human Purchase Intent via Semantic Similarity Elicitation of Likert Ratings”. The full derivation, including the worked example above, is documented in the open.
Every construct — purchase intent, value for money, claim believability, packaging appeal — scored across the cohort and cut by segment. Which tier leans in. Which income band walks. Where the price stops being defensible. The concerns are quoted back in the words the personas used, because “too expensive” and “not for someone like me” are different problems with different fixes.
Every synthetic research vendor will tell you their personas are uncannily human. We would rather tell you where the method breaks, so you know what you are buying.
Synthetic consumers approximate patterns, not people. Treat the output as directional signal for screening and prioritising — not as evidence for a bet-the-year launch.
The anchors define the construct. Shrota measures proximity to a set of calibrated statements. Change the anchors and you change what the number means.
It does not replace fieldwork. It decides what to send there. Screen forty concepts here, take the surviving three to real consumers.
The fastest way to judge Shrota is to run something you have real numbers for, and see whether we would have told you the same thing. Forty minutes, your category, your data.