Distribution, Not Recommendation

L'Oreal just bought structural presence inside ChatGPT's answers. The try-on demo is the decoy. The real story is what happens to advice when one brand owns the answer

Sometime this summer, someone will ask ChatGPT to help them find a good red lipstick, and Maybelline will already be living inside the answer. Not as a link. As the experience itself: a virtual try-on rendered right in the conversation, powered by L'Oreal's ModiFace and announced with OpenAI at VivaTech in June.

The demo is gorgeous and the coverage wrote itself. But the try-on is the decoy. The deal has three parts, and the AR is the least consequential one.

The three-part deal

Part one is the visible one. Maybelline's Makeup Virtual Try-On runs directly inside ChatGPT, a conversational surface with more than 900 million weekly users. Two decades of ModiFace R&D meeting the largest assistant on the planet. That's the headline, and it earned it.

Part two is quieter. L'Oreal will feed verified, structured product data straight into ChatGPT to strengthen organic discovery for brands like Lancome and Kerastase in the US. Not ads. Answers. When the model talks about a Lancome serum, it will draw on the brand's own clean feed instead of whatever scraped retailer copy it would otherwise surface.

Part three: SkinCeuticals, CeraVe and Garnier join OpenAI's advertising pilot, testing sponsored placements at what both companies call moments of consumer intent.

Add them up and the pattern is unmistakable. This isn't a feature launch. It's the world's biggest beauty company buying structural presence inside the answer, before the question is even asked.

What an answer with one ending can't do

Now run the interaction. A user asks for a red lipstick. The Maybelline experience opens. The face tracking is state of the art, the rendering is dynamically lit, the product is genuinely good. And the answer still has a ceiling it can never break: it can only end in Maybelline.

That's the distinction the industry keeps blurring, so let's make it precise. A recommendation requires the structural possibility of answering "not this brand." If every path through the experience terminates in the same catalog, what you've built is not an advisor. It's a channel. Distribution, not recommendation.

There's nothing wrong with distribution. L'Oreal is playing its hand perfectly, and every brand with the scale to cut its own deal will copy this move within a year. But nobody should confuse the two, because consumers won't. Not for long.

Why the model breaks at scale

L'Oreal's own executives supplied the counterargument, probably without meaning to. Chief digital and marketing officer Asmita Dubey talks about an "11-minute paradox": consumers researching beauty through conversational AI spend far longer per session than they do in traditional search. People aren't using assistants to shortcut the decision. They're using them to deliberate.

Deliberation is comparison. The whole reason to ask an assistant instead of a brand site is that the assistant, in theory, sits above the brands. A per-brand companion inverts that promise. Nobody will run one companion per brand any more than they'd install one search engine per website.

So the brand-owned integrations will multiply, they'll be beautiful, and every one of them will deepen the same structural gap: the assistant era still has no advisor.

The layer nobody has claimed

The gap has a shape. Between brand catalogs and the assistant, something has to sit that can compare across brands the way a good retail advisor would. It needs three properties.

It has to be callable: exposed as a tool an agent can invoke mid-conversation, not an app someone must open. It has to be legible: structured so a model can reason with the data, not just retrieve it. And it has to pass the Indifference Test: zero stake in which product wins. No house brand, no affiliate cut, no verified feed that quietly promotes its own portfolio.

A brand fails the test by definition. A marketplace passes only inside its own inventory. The layer has to be neutral by architecture, not by promise.

What this means if you're not L'Oreal

If you're a brand without a foundational-partner deal at OpenAI, your products now compete against a verified feed. Your move is legibility: structure your data so any model can reconstruct what your product does and who it's for. That work is unglamorous and it compounds.

If you're a retailer, the calculus is sharper. Your shelf is your asset, and the assistant era is currently pricing it at zero. Building your own companion fails the same scale test the brands are about to fail. The move is to put a neutral decision layer over your own catalog, so that when an agent, or a customer standing at your point of sale, asks what to buy, the answer draws on your assortment with advisor logic instead of brand logic.

That layer is what we build at SKINBOT: a neutral B2B decision layer that runs skin analysis and recommends exclusively from the partner retailer's catalog. It integrates as an API, QR entry point or iframe, and it holds no stake in which product wins. In production, the pattern holds: when the advice is neutral and the entry is self-service, roughly 80 percent of customers who start finish and take a recommendation.

L'Oreal just proved that the answer is the new shelf. The next thing this industry has to build is the part of the answer that doesn't belong to anyone.