Drugstore and dermatologist brands own the AI skincare shelf

Recommendation share is concentrated. SkinCeuticals, CeraVe, La Roche-Posay, Paula's Choice, Vanicreamlead, and most of the names that win are drugstore staples and dermatologist-credentialed brands rather than the luxury labels that dominate paid media. AI's default beauty shelf is a short, stable list, and it skews to accessibility and clinical credibility.

That shelf isn't idle browsing. Across the buyer conversations, 12 of 17 ended in a purchase decision: when a shopper brings the question to AI, the assistant usually moves them to buy something. The consideration set AI assembles is the one that converts, so being absent from it is a lost sale, not just a lost impression.

Across engines, roughly 45% of the pages AI cited for beauty were published more than twelve months ago, and the median cited source is 6 to 13 months old, depending on the engine. A formula reformulated, a claim updated, a hero product relaunched since then, none of it reaches the answer if AI is reading the version of you from last year.

And it's reading the wrong kind of page. About 89% of cited sources are uncategorized third-party pages, blogs, listicles, and roundups, rather than your own site, a marketplace listing, or major media. The brand reputation AI recommends on is assembled mostly from secondhand pages you didn't write and may never have seen.

Where AI actually looks: no dominant domain, just a long tail of pages

In our October 2026 source scan, no single domain dominates. The most cited is allure.com, at 4.8% of all 4,725 citations (229 of them). Below it the field breaks into 842 distinct domains, and reddit.com (3.6%), nbcnews.com (3%) and byrdie.com (2.2%) also reach 2% or more. By category, the largest share goes to independent blog / other third-party pages (48%), then editorial / press (25%).

One change since July 2026: between our July 2026 and October 2026 runs, the search layer stopped ranking YouTube videos, and ranked far fewer Reddit threads, for standard buying questions. Both still appear when a shopper asks for videos or opinions. The older models show the same drop, so this is a retrieval change, not a model or market shift.

The industry has started ranking these sources. A June 2026 analysis by beauty-tech company Novi, covered by CEW, counted 10.7 million citations inside ChatGPT and put Reddit first by a wide margin, with Wikipedia in the top five. In live grounded retrieval across four engines we see a different picture: Reddit sits at 3.6% and Wikipedia was cited 2 times. Part of that gap is the search layer we use, which returned far fewer Reddit results this month than in July. The larger lesson holds either way. A citation count only sees the answers that carry citations. In October 2026, ChatGPT answered 30% of these beauty questions without searching the live web, and those answers are invisible to that method. Source rankings tell you where AI looks. mapou measures what AI recommends, for what reasons, and whether the shopper buys, across four engines, on a monthly clock.

When AI explains why it recommends a brand, the reasons concentrate on being widely available and ingredient-led, with gentleness (barrier-safe, fragrance-free, safe for sensitive skin) close behind. These are functional, accessibility-driven reasons. The buzz angles brands spend on, luxury signaling, going viral, celebrity association, barely register: together they are about 9% of all reasons given. Clinical results is the exception: it ranks fifth of the 14 reasons we tagged.

For a drugstore brand like CeraVe, being recommended as widely available and barrier-safe is on-message. The risk is a prestige or clinical brand that spends to signal luxury or efficacy and is resold by AI as a basic, available option. In the layer that now mediates discovery, the reason attached to your name may not be the one you're paying for.

Grounding is whether the engine stops and searches the live web before answering, instead of replying from training memory that can be a year or more out of date. ChatGPT grounded on only 70% of these beauty questions; Gemini, Claude and Grok grounded 82 to 99%. The existence of citations is the proof an answer is influenceable: when an engine grounds, it builds the answer from pages you can change.

These rates come through our standardized retrieval harness, one shared search layer across all four engines; on its native consumer search the same model can ground more or less often, so the cross-engine gap and the monthly trend are the durable read, not any single rate.

This is where the freshness problem bites hardest. When ChatGPTanswers from memory, it answers from the same stale picture that carries the year-old sources and the off-strategy reasons, with no live page to correct it. The fix isn't to change the model; it's to give the grounded engines current pages to find, and to publish the structured signals that pull the memory-first engine onto the web more often.

This month the measured engines moved to a new model generation. To tell a real category shift from a model change, we re-ran July 2026's models on a frozen subset of the same beauty prompts the same day. Held constant, the category barely moved. What did move was the models themselves.

The clearest effect is grounding, how often an engine searches the live web before answering. On the same prompts, the upgrade changed it sharply, and in opposite directions by engine.

The upgrade also made answers more concentrated: 10 of the 23 tracked beauty brands were named less often by the new models than the old ones on the same prompts. Newer models give shorter, more decisive answers, so a few names hold, and the long tail thins. That is a model property, not a change in the category, and it is worth watching: the shelf AI shows a shopper is getting narrower.

You can't change a model's training. You don't need to. The three things AI gets most wrong about you, whether you're on the shelf, how current the story is, and why you're recommended, all trace back to pages. Every time an engine grounds, and on these questions the engines that ground did so up to 99% of the time, it builds the answer from pages you can change. Change the grounded pages, change the answer. The ungrounded share moves on the model's clock, so we measure both and report where each engine sits.

This is SEO's lesson one layer up: you can't move the algorithm, but you can change what it reads. The brand whose current claims are the easiest thing for AI to find is the one it recommends correctly.

1. 1

   Find where AI looks

   We pull the pages AI cites for your category and split them into what you own, what you influence, and what to dispute.

   mapou service: AI Visibility Monitor, Source & Citation Map

2. 2

   Fix the stale and off-strategy inputs

   Make your current claims, ingredients, and positioning machine-readable on the pages AI reads, and refresh the year-old third-party pages still framing you.

   mapou service: AI Content & Schema Studio

3. 3

   Re-measure, prove it moved

   Same prompts, same engines, next month. You see whether your recommendation share, source freshness, and narrative fit moved.

   mapou service: Tests & Lift

Each finding above maps to a report you can act on: a monthly read of whether AI puts your brand in the consideration set, by engine and segment; a source and freshness map of the pages AI reads for you; a narrative-fit report on the reasons AI attaches to your brand; and a grounding read by engine, verified by next month's re-run. Explore the live beauty intelligence.