Pre-production design intelligence
PRÉVU validates your unreleased jewelry & accessories concepts with synthetic consumer panels and a decade of category expertise. No guessing. No blind production runs. No designs left to wonder “would this have been the hit?”
Backed by methods proven at L’Oréal / ModiFace scale · Zero-PII synthetic panels · NDA-secure intake
The old way is broken
L’Oréal acquired ModiFace for $1.2B and runs 217 virtual product tests a year — proving pre-production validation is a real, valuable need. We bring that capability to the brands that can’t build ModiFace in-house.
How it works
3–7 unreleased designs. Three photos per piece plus the essentials: material, price point, target customer, and the story behind it.
Our model extracts each design’s visual attributes — silhouette, material language, scale, palette, and style family — the same way published conjoint research identifies a product’s “visual DNA.”
250+ synthetic consumer personas evaluate the whole cohort across six dimensions. Multi-persona forced disagreement cancels AI flattery.
A comparative ranking — not a guess — with per-dimension scores, audience-fit slices, and concrete redesign direction.
Every real try-on and sale we observe feeds back to recalibrate the personas — the report gets sharper with each cohort. A data flywheel, not a one-off.
Methodology
This is the single most important design decision in PRÉVU — and the reason our results are trustworthy where a naive “AI, rate my design” prompt is not.
Show an AI one product and ask “is this good?” and it will almost always say yes. Measured studies (Cheng et al., Science, 2026) found LLMs agree with users 39–79% more than humans do. A lone design gets flattered into a false positive.
Force the panel to choose between designs and the flattery cancels out. “Which of these five would you actually wear?” produces a real ordinal ranking — the same logic behind choice-based conjoint, the gold standard of design research.
People (and models) over-favor the first and last items in a list — primacy and recency effects. We randomize presentation order on every run so position can never decide the winner.
We benchmark each cohort against your own bestseller (kept outside the ranked set). It gives the panel a stable reference point, so scores mean something absolute over time — not just “relative to this batch.”
“Every try-on, every click, every skipped product is a valuable signal.”
— The dual-purpose VTO model, validated by Arbelle (NYC beauty-tech) and Wildberries (AI jewelry try-on)Ranking dimensions
Each design is scored on six independent dimensions — the ones that actually predict a jewelry hit. No single “overall score” hides the story.
Will it catch the eye in a feed? Proxy: dwell time / screenshot-save rate.
Will they show it off? Proxy: share / save rate.
Will they actually buy? Proxy: add-to-cart / conversion.
Does it match your target customer’s archetype? Per-segment slices.
Does it feel worth the price? Proxy: price-to-desirability gap.
Does it stand out vs. the category? Benchmark vs. your bestseller.
How many designs per cohort?
Choice-based conjoint research consistently recommends 2–5 concepts per screen; Hick’s Law puts the comfortable decision window at 4–7; the famous jam study showed choice overload kicks in past about six options. So we set the cohort at 3–7, defaulting to 5.
More importantly: persona count beats item count. With 250+ synthetic personas each “voting” on the cohort, five designs yield a stable, statistically meaningful ordinal ranking. Adding a sixth or seventh design adds noise, not clarity.
Why trust the results
We’ve built and sold jewelry and accessories across Amazon EU and direct-to-consumer for 15+ years — including a live pearl line. We know product, pricing, and channel. AI is the amplifier, not the foundation.
Visual-DNA extraction follows published conjoint research (Purdue / Yale). Our synthetic calibration cites Minds (85–100% agreement with traditional panels) and Articos (80–92% correlation, 86% recall, peer-reviewed across 46 studies).
Early results are clearly marked as directional comparative rankings, not absolute purchase predictions. As real try-on data accumulates, the report upgrades from “directional” to “predictive.” We’d rather under-promise.
Before we sell a single report, we run our own pearl collection through PRÉVU and publish the model-vs-actual match as a sample. You see the method working on real sales before you pay.
Who it’s for
Confidentiality & compliance
Every submission is covered by NDA and logically isolated. Unreleased designs are never used for training or shared with third parties without explicit authorization.
Our panels are synthetic. We extract statistical patterns from aggregated, anonymized data, then generate entirely new personas — never re-label real individuals. This is true anonymization, not pseudonymization, and sits outside personal-data regulation.
Any consumer try-on data we collect carries consent scoped to aggregated research. We run a Data Protection Impact Assessment before commercial launch. GDPR and PIPL aligned.
Pricing
One design cohort, end-to-end.
Ongoing validation for your new-collection pipeline.
Self-serve validation at scale.
Pricing is tailored to cohort volume and cadence. Tell us your pipeline and we’ll quote — no spreadsheet of mystery fees.
FAQ
Yes — if you show it one design in isolation. That’s exactly why PRÉVU always evaluates a cohort and asks the panel to choose between pieces. Relative judgment cancels the flattery bias, and we randomize order to kill position effects.
A minimum of three, a default of five, a hard maximum of seven. Three gives a real 1st/2nd/3rd; five is the sweet spot before choice overload; seven is the ceiling where signal starts to drown in noise.
Yes. Submissions are NDA-covered and logically isolated. Unreleased designs are never used for model training or shared without your explicit authorization.
No. Our panels are synthetic — built from aggregated, anonymized statistical patterns, then generated as new personas. That’s true anonymization (not pseudonymization), so it sits outside personal-data law. We run a DPIA before launch.
In later tiers, yes — you can upload anonymized panel data to calibrate the synthetic layer to your real audience. That’s how we graduate from “directional” to “predictive.”
A standard cohort returns in days, not the weeks a focus group demands. We’ve benchmarked the method as ~11 weeks faster to market than traditional validation.
Early reports are labeled directional. Your first cohort is often free, and we publish our own pearl-line self-validation so you can see the method’s track record before relying on it.
Get started
Tell us about your next collection. We’ll run one cohort on us and show you exactly what PRÉVU sees.
Prefer email? hello@prevu.studio