From 38 to 58: how a skincare brand's AI Win-Rate™ moved in five months
A premium skincare brand asked one question: when a shopper asks an AI assistant to choose an anti-aging product, how often does it choose us? In May 2026 the answer was 38 times in 100. By October 2026 it was 58. This is what we measured, what we changed, and what the result does and does not show.
AI Win-Rate up 20 points, and the brand now recognized by every assistant
AI Win-Rate at each reading: decision conversations the brand won outright, per 100 tested. Each reading re-ran the same scenarios. Readings are published only when the margin of error is within ±10 points.
| Measure | May 2026 | October 2026 |
|---|---|---|
| AI Win-Rate | 38 | 58 |
| Brand recognition | Three of the four assistants could not say who the brand was | All four describe the brand, its founder, its technology and its clinical study accurately |
| Public evidence | Clinical study not public; founder expertise not set out anywhere an assistant could use it | Clinical study and a conference talk published with permanent DOIs, founder pages, 17 educational articles |
The same buying conversations, run to a decision, every reading
- Shopper scenarios. Written briefs such as a woman in her fifties looking for an anti-aging serum for mature skin, some opening without the brand named, as a real shopper would.
- Run to a choice. Each conversation works through follow-up questions a real shopper asks (does it work, has it been tested on older skin, is it worth the price) and ends by asking the assistant to pick one product.
- Four assistants. ChatGPT, Gemini, Claude and Perplexity, each run several times per scenario, because the same question gets different answers on different runs.
- Strict scoring. Only outright wins count. Shared picks, losses to a competitor, referrals to a retailer or doctor, and refusals to choose are reported separately.
- Same tests each time. Every reading re-runs the same scenario set, so a change in the number reflects a change in the assistants, not in the questions.
Spoken of well, then dropped when the shopper pushed
The baseline found six evidence gaps, led by category recognition, product-specific clinical evidence and the founder's medical authority. The diagnostics showed where the brand was losing. Assistants dropped it at the second turn, when the shopper asked about the science, and a medical-grade competitor or a mass-market retinol serum took the recommendation. In the first answer the brand's sentiment scored 71 out of 100, but its resilience, whether the mention held when challenged, scored only 25.
Evidence an assistant can find and trust, tied to each gap
| Gap | What was done |
|---|---|
| Clinical evidence | An eight-week clinical study published as a public record with a permanent DOI, a Clinical Evidence page, and an article on the study design |
| Founder authority | Four founder pages (biography, credentials, clinical evidence, press), a conference talk deposited with a DOI, and independent press coverage archived and linked |
| Answering the shopper's questions | An ingredient mechanism explainer and 17 educational articles aimed at the mid-conversation questions where the brand was being dropped |
| Claims assistants can repeat safely | Approved claims aligned with the live pages; unsupported and comparative claims removed until peer-reviewed |
| Recognition | A rewritten llms.txt stating what the brand does and does not claim, and structured data linking brand, founder and products |
Thirteen items were delivered between May and October 2026. Every change was approved by the client before it went live.
What this shows, and what it does not
The brand still loses about four in ten decision conversations. The main reasons: the founder's authority does not yet carry to individual products, most evidence is still self-published, some assistants hedge on health-related wording, and much of the new material has not yet reached the assistants' training data, which refreshes on a schedule no one outside the AI companies controls. Gains have been uneven: ChatGPT moved first, with Gemini and Perplexity slower.
The case measured the brand on its own site and on its retailer product pages. Every run is recorded. Full run logs and raw outputs are available to prospective clients under a mutual NDA.
Where does AI drop your brand?
An enterprise diagnostic shows your AI Win-Rate by audience and assistant, who wins instead and why, and the evidence gaps behind each loss.
Ask for a diagnostic 30-day Sprint, US$3,500