# Q10 ChatGPT — "Gentle retinol alternative for sensitive skin that won't cause purging"

Captured 2026-09-02, logged out, 5 fresh sessions, full-page. Final query.

## Runs

| Run | Cards | Brands | Citations | Ad |
|---|---|---|---|---|
| 1 | Medik8 Bakuchiol Peptides $58·Sephora; Naturium Azelaic 10% $19.88·Walmart | Medik8, Naturium | **PubMed**; **AAFP**; Cleveland Clinic | none |
| 2 | Medik8 $58·Sephora; Cetaphil Healthy Renew $26.40·iHerb; Beekman 1802 $43.26·Celaura; Herbivore Moon Fruit $68·Herbivore; Good Molecules Azelaic $12·Ulta | Medik8, Cetaphil, Beekman, Herbivore, Good Molecules | AAD ×2 | none |
| 3 | **none** | **none** | **PubMed ×3, exclusively** | none |
| 4 | **none** | **none** | **PubMed ×2; DOI (direct journal link)** | none |
| 5 | Medik8 $58·Sephora; Herbivore $68; Beekman $43.26; ANUA Azelaic $22·Ulta | Medik8, Herbivore, Beekman, ANUA | **PubMed**; Allure ×2; Cleveland Clinic | none |

## 1. The mechanism completes: a fourth source tier — primary literature

**This is the only query in the study citing PubMed and a DOI.** Not magazines, not brand sites,
not even professional bodies — actual randomised trials. Run 3 cited PubMed three times and
nothing else.

The reason is the constraint. *"Won't cause purging"* is an **efficacy claim that requires
evidence**, and bakuchiol-vs-retinol has real RCT literature. So the model reaches for it, and
quotes it: *"In a 12-week randomized trial, bakuchiol performed similarly to 0.5% retinol, while
retinol caused more scaling and stinging."*

### The completed mechanism — constraint type determines source type

| Constraint type | Query | Sources the model reaches for |
|---|---|---|
| **Experiential** | pilling under makeup (q07) | Beauty magazines + brand marketing |
| **Medical** | acne (q04) | AAD, exclusively |
| **Safety-critical** | pregnancy (q08) | ACOG, MotherToBaby — **and no products at all** |
| **Efficacy-claim** | won't purge (q10) | **PubMed, DOI — primary literature** |

Four tiers, four queries, cleanly separated. **This is the study's most defensible original
finding.** A brand cannot win an experiential query with clinical evidence, and cannot win an
efficacy query with marketing copy — the model looks in a different place depending on what
kind of question it is.

## 2. Lowest overlap in the study — 29%

| ChatGPT brand | Google position |
|---|---|
| Medik8 | **#1** ✅ |
| Herbivore | #4, #11 ✅ |
| Beekman 1802 | absent ❌ |
| Cetaphil | absent ❌ |
| Naturium | absent ❌ |
| Good Molecules | absent ❌ |
| ANUA | absent ❌ |

**Top 40: 2/7 = 29%. Top 10: 2/7 = 29%.**

### But the reason is a fourth divergence cause: category expansion

Four of the five absent brands (Naturium, Good Molecules, ANUA, and arguably Cetaphil) are
**azelaic acid** products. Google's shelf is strictly *retinol alternatives*. **ChatGPT answered
the underlying need — gentle anti-ageing for sensitive skin — rather than the literal query.**

Restricted to actual retinol-alternative products, overlap is 2/4 = 50%.

And **ChatGPT's top pick, Medik8 Bakuchiol Peptides, is Google's #1** — the second query where
the top picks align exactly (q01 was the other).

### Divergence has four distinct causes

| Cause | Query | Signature |
|---|---|---|
| Editorial promotion | q03, q07 | Magazines push brands the shelf doesn't rank |
| Semantic intent match | q05 | Product *named* for the query |
| Query-shape mismatch | q09 | Shelf sells kits, AI builds from parts |
| **Category expansion** | **q10** | **AI answers the need, not the literal query** |

## 3. Partial commerce shutdown — 2 of 5 runs

Runs 3 and 4 returned **no cards and no brands**, like q08. Both were the PubMed-heavy runs.

Pattern worth noting: **the runs that reached for primary literature sold nothing.** When the
model is in evidence mode, it stops being a shopfront. 7 of 50 runs across the study returned
zero commerce, and all 7 were on safety- or evidence-constrained queries.

## 4. Ads: 0/5 — the long-tail hypothesis is confirmed

Fifth consecutive zero-ad query. Final tally: **15 of 50 runs (30%)**.

All 15 ads landed on the three broadest queries (q01, q02, q03, q06). Every specific query —
q04, q07, q08, q09, q10 — drew zero. **Ad inventory tracks query breadth, not category.**

## 5. An unusual merchant appears

Run 2 and run 5 both surfaced Beekman 1802 at **"Celaura Beauty | Premium Beauty, Skincare &
Self-Care Essentials USA"** — a merchant name that reads like an SEO title tag rather than a
retailer. Also **iHerb** for Cetaphil. On the long tail, the merchant layer gets stranger.
