# Q07 ChatGPT — "Best fragrance-free sunscreen for oily skin that doesn't pill under makeup"

**Three simultaneous constraints — the hardest query in the set.** Captured 2026-09-02, logged
out, 5 fresh sessions, full-page.

## Runs

| Run | Top pick | Cards (price · merchant) | Citations | Ad |
|---|---|---|---|---|
| 1 | **LRP Anthelios UV Air** SPF 50 | LRP $39.99·Ulta; Supergoop! Unseen $38.00·Target; **EltaMD UV Clear $23.00·EltaMD** | **Glamour**; **Allure** | none |
| 2 | **Supergoop! Unseen** | Supergoop! $38.00·Target; Eucerin Oil Control $19.99·CVS; **DRMTLGY SmrtSun 45** $28.00·Ulta; Supergoop! Mattescreen $40.00·Nordstrom; LRP Mineral UL $42.97·Walmart | **Allure** ×3; NBC News | none |
| 3 | **Supergoop! Unseen** | Supergoop! $38.00·Target; DRMTLGY $28.00·Ulta; Eucerin $19.99·CVS; Supergoop! Mattescreen $40.00·Nordstrom; Supergoop! PLAY $16.00·Ulta | **Allure** ×2; **Glamour**; **New York Post** | none |
| 4 | **EltaMD UV Clear** SPF 46 | **EltaMD $45.00·Walmart-Pattern**; Supergoop! $38.00·Target; LRP Clear Skin $20.99·CVS; Eucerin $19.99·CVS | **EltaMD, Supergoop!, La Roche-Posay, Eucerin — all own sites**; Allure | none |
| 5 | **Supergoop! Unseen** | Supergoop! $38.00·Target; DRMTLGY $28.00·Ulta; LRP Mineral UL $42.97·Walmart; **CeraVe Ultra-Light SPF 30 $15.97·Walmart** | **Glamour** ×2; allure.com | none |

**Top-pick churn: Supergoop! Unseen 3/5, LRP 1/5, EltaMD 1/5.**
**Brands: Supergoop! (5/5), LRP (4/5), Eucerin (4/5), DRMTLGY (3/5), EltaMD (2/5), CeraVe (1/5).**

## 1. Refined hypothesis CONFIRMED

Magazine-saturated citations (**Allure in 5/5**, Glamour ×4, NY Post, NBC) with **zero medical
sources** — no AAD, no Cleveland Clinic — and **no price constraint**.

Prediction: divergence from the commerce shelf. **Result: top 40 = 5/6 = 83%. Top 10 = 3/6 = 50%.**

**DRMTLGY appears in 3 of 5 runs and is absent from Google's top 40 entirely** — a
direct-to-consumer derm brand heavily covered in beauty press, lightly ranked on commerce.
Exactly the q03 profile (SkinCeuticals, BeautyStat).

Constrained-with-a-price-ceiling (q06) → 100%. Unconstrained magazine-led (q03, q07) → 70%, 83%.
The refinement holds across four tests.

## 2. Deepest re-rank in the study

**Supergoop! Unseen, the top pick in 3 of 5 runs, is Google's #38** — the very bottom of the
top 40. Deeper than INKEY at #35 (q05) and EltaMD at #15 (q02).

On the hardest query, the AI's answer comes from the bottom of the commerce shelf.

## 3. NEW MECHANISM: the constraint type determines the source type

This is the sharpest finding of the query.

"Doesn't pill under makeup" is **experiential**. No clinical body publishes on it. There is no
AAD guidance for pilling. So the model falls back on whoever *does* make claims about it:
**beauty magazines and the manufacturers themselves.**

Run 4 is the extreme case — **four recommendations, four brand-own-site citations**: EltaMD,
Supergoop!, La Roche-Posay and Eucerin, each cited as evidence for its own product.

ChatGPT's wording stays honest — *"**The brand says** it layers well under makeup"*,
*"**Supergoop! describes it as** scent-free"* — but the visible badge does the persuading.

| Constraint type | Example | Sources the model reaches for |
|---|---|---|
| **Medical** | acne (q04) | AAD, exclusively |
| **Ingredient/safety** | sensitive skin (q01) | Cleveland Clinic, AAD, hospital content |
| **Experiential** | pilling under makeup (q07) | **Beauty magazines + brand marketing** |

**A brand cannot win an experiential query with clinical evidence, because none exists. It wins
by being in Allure and by making the claim on its own site.** That is a directly actionable
finding for a marketer and nobody has published it.

## 4. Biggest price spread in the study: 96%

**EltaMD UV Clear SPF 46:**
- **$23.00 · EltaMD** (run 1, brand direct)
- **$45.00 · Walmart - Pattern** (run 4)

**Nearly double, same product name.** "Pattern" is a third-party marketplace seller. The AI
surfaced a marketplace listing at ~2x the brand-direct price shown in another run, with nothing
signalling the difference.

This is the clearest consumer-harm example in the dataset, and it needs no interpretation.

Also: CeraVe Ultra-Light SPF 30 at **$15.97** matches Google's **$15.97** exactly — eighth
to-the-cent match.

## 5. Zero ads — and this breaks the regulatory theory

**0 of 5 runs.** Second query with none, after q04.

But **q02 was also sunscreen and ran ads in 3 of 5.** Same product category, opposite outcome.
So regulatory category alone does not explain it.

**Better hypothesis: head vs long tail.** Ads and Google's curated "Best…" modules both
concentrate on broad queries and vanish on specific ones. q02 ("best face sunscreen for daily
use") is a head query with ads and Google modules. q07 (three constraints) is long-tail with
neither. q04 may be both regulated *and* narrow.

**Prediction to test on q08-q10:** the two long-tail queries should carry no ads.

Running total: **15 of 35 runs (43%)**.

## 6. The model flagged its own constraint uncertainty

Run 3, unprompted: *"if you're specifically looking for zero fragrance AND zero essential oils,
check the current ingredient list before buying, since formulations can change."*

The model hedged on the fragrance-free constraint rather than asserting compliance. Worth
recording — it is the opposite of the confident-and-wrong failure mode, and it complicates any
simple "AI is overconfident" narrative.