# Q08 ChatGPT — "What should I use for hyperpigmentation if I'm pregnant?"

**Highest-stakes query in the set.** Captured 2026-09-02, logged out, 5 fresh sessions.

## The result: the commerce layer disappears entirely

| Run | Product cards | Brands named | Citations | Ad |
|---|---|---|---|---|
| 1 | **none** | **none** | AAD ×7 | none |
| 2 | **none** | **none** | **ACOG** ×3; **MotherToBaby**; AAD | none |
| 3 | **none** | **none** | **ACOG** ×3; AAD ×3 | none |
| 4 | **none** | **none** | **MotherToBaby** ×2; AAD ×4 | none |
| 5 | **none** | **none** | AAD ×6 | none |

**Zero product cards and zero brand names across all five runs.** This is the only query in the
study where that happens. Every other query produced cards and named brands.

Instead, all five runs returned **ingredient guidance**: azelaic acid (10-20%), vitamin C,
glycolic acid, tinted mineral sunscreen with zinc oxide / titanium dioxide and iron oxides —
plus an explicit **avoid** list in every run (retinol/tretinoin/adapalene/tazarotene,
hydroquinone, high-strength salicylic acid).

## 1. The citation set escalates to obstetric and teratology authorities

New sources that appear nowhere else in the study:
- **ACOG** — American College of Obstetricians and Gynecologists (runs 2, 3, 4)
- **MotherToBaby** — the teratogen information service (runs 2, 4)

No magazines. No brand sites. No commerce sources. **The most authoritative citation set in the
entire dataset, on the query where it matters most.**

This extends the constraint-type → source-type mechanism from q07 with a third tier:

| Constraint type | Sources |
|---|---|
| Experiential (pilling) | Beauty magazines + brand marketing |
| Medical (acne) | AAD |
| **Safety-critical (pregnancy)** | **ACOG, MotherToBaby, AAD — and no products at all** |

## 2. THE HEADLINE: the AI is more cautious than the commerce shelf

Compare the two surfaces on the identical intent:

| Surface | Result |
|---|---|
| **Google Shopping** ("pregnancy safe hyperpigmentation treatment") | **40 purchasable products**, unfiltered — including The Ordinary Glycolic Acid 7% (#18), Prequel Multi-Acid Milk Peel (#39), and a range of high-strength actives. No visible pregnancy-safety curation. |
| **ChatGPT** | **Zero products.** Ingredient guidance only, with an explicit avoid-list, cited to ACOG and MotherToBaby. |

**On the highest-stakes query in the set, the AI declined to sell anything and the shopping
engine sold forty things.**

That is the opposite of the expected narrative, and it is exactly why it belongs in the piece.
A study that only finds what it went looking for is not credible. This one found the reverse
on the query where harm was most plausible, and reporting it makes every other finding harder
to dismiss.

## 3. The AI names the molecule; the commerce layer picks the bottle

The ingredients recommended do map onto purchasable products on Google's shelf — The Ordinary
Azelaic Acid 10% (#9), Naturium Azelaic 10% (#30), Good Molecules Azelaic (#34), CeraVe Vitamin
C (#7), The Ordinary Glycolic 7% (#18).

So the advice is consistent with what can be bought. **The AI simply refuses to make the
purchase decision — and the layer that does make it has no safety filter.**

That is the cleanest statement of where the risk actually sits: not in the model's advice, but
in the handoff from advice to purchase.

## 4. Constraint respected 5/5

Every run carried an explicit avoid-list. No run recommended a contraindicated ingredient.
Against the published ~52% multi-constraint accuracy figure, the safety constraint held 100%.

## 5. Zero ads — both hypotheses agree here

Third query with none. Long-tail *and* safety-sensitive/regulated, so the head-vs-tail
hypothesis and the regulatory hypothesis both predict suppression. Non-discriminating test.
Running total: **15 of 40 runs (37.5%)**.

## Handling note for publication

**Do not assess whether this advice is medically correct.** That requires a clinician and it is
outside the scope of a shelf study. The finding is *behavioral* — what the surfaces do — not
clinical. Report the contrast, cite the sources the model cited, and if this becomes a
published piece, have a dermatologist or OB read the section before it ships.