The window for treating AI search as a "wait and see" channel closed quietly in the last 60 days. Two announcements, two weeks apart, ended the conversation about whether agentic commerce in beauty was real.
Ulta Beauty went live in late April with a Google Gemini integration that pushes product recommendations directly into AI Mode in Google Search and into the Gemini app. The retailer confirmed on its 5 June Q1 FY26 earnings call that ecommerce grew in the mid teens, and named the AI deployment as a contributor. Sephora followed days later with a ChatGPT integration that connects a Beauty Insider account, lets customers ask questions in natural language and ships products back without leaving the chat.
Both companies were quick to frame this as a "discovery experience". The operational reality is sharper than that. A meaningful slice of customer journeys at the two largest US beauty retailers now begins with an AI agent reading product data, then deciding which brands to surface. The human sees three to five suggestions. They almost never see the rest.
If you sell into Ulta or Sephora, or you compete with brands that do, this is a category-level change in how discovery works. It is not a marketing trend to test. It is a structural change in the front of the funnel.
What the AI is actually doing
When a customer types or speaks something like "I have combination skin and a small budget, what cleanser should I try", the agent is doing four things in parallel. It is parsing the question for intent (cleanser, combination skin, price-sensitive). It is filtering the catalogue down to a candidate set. It is ranking that set by the data it can read. Then it is summarising the top three or four in natural language.
The brands that surface in the answer share a pattern. Their product data is structured. Their ingredients are named in the right place with the right syntax. Their reviews are accessible to the model, not locked behind a JavaScript widget. Their use cases are written somewhere the agent can find them. Their efficacy claims are framed in a way the model can verify against a source.
The brands that do not surface usually have the same product story buried in a brand-led PDP that reads beautifully to a human and incoherently to a machine. The information is there. The agent cannot extract it.
The four layers that decide whether you surface
The PDPs that read well to an AI agent share four data layers, in this order.
The first layer is structured ingredient data. Not a paragraph that says "powered by hyaluronic acid". A clean ingredient list with the actives named, the percentages where you have them, and the function of each tied to a benefit the customer might search for. If your PDP buries the ingredient story in a product description block with no semantic markup, the agent has to guess. It usually guesses against you.
The second layer is structured review data. Reviews are the single highest-leverage signal for an AI agent in beauty because they answer the customer's real question: did this work for someone like me. If your reviews are inside a third-party widget that does not expose its content to the page DOM in a crawlable way, you are invisible. If your reviews are crawlable but unstructured, the agent reads them but cannot weight them. The brands that are surfacing in AI answers right now have crawlable reviews with structured metadata about skin type, age range, concern and outcome.
The third layer is structured use-case data. AI agents are asked questions like "what should I use for redness after retinol" or "is there a cleanser that works with hard water". The brands that surface are the ones with explicit answers to those questions in their content, written in language that maps to how customers actually ask. The brands that do not surface have brand-led product pages that lead with the founder story or the science but never answer the customer's situation.
The fourth layer is structured claim evidence. AI agents are increasingly trained to distrust unsupported efficacy claims, especially after the rolling enforcement waves at the ASA and the rising US FDA scrutiny under MoCRA. The brands that surface are the ones whose claims are paired with the study, the methodology or the result, in a format the agent can verify. "Clinically proven" with no source attached is now a downgrade signal in many models.
Why this hits £500k-£5m brands hardest
Larger brands have product information teams, taxonomy tools and engineering capacity to restructure their data. They are doing it now. Some have been doing it for 18 months.
A brand at £500k-£5m usually does not have that team. The PDPs were built once, the ingredient story was written by the founder, the reviews live in a Yotpo or Okendo widget, the use-case content is on a blog that points at the homepage, and the claims have been the same since launch. None of it was wrong when human discovery was the dominant flow. All of it is now suboptimal.
The worst version of this is the brand that is doing well on TikTok, getting featured at retail, and quietly losing share in AI answers because the agent cannot map the buzz to the data. The reviews say the product is brilliant. The PDP says the product has hyaluronic acid. The agent has nothing to connect those two signals to a customer asking "what is the best cleanser for sensitive skin under £25".
The audit to run this quarter
Pick your three highest-traffic SKUs. Open each PDP in a private window. Ask three things.
First, can a basic text crawl extract the active ingredients, the percentages, the function and the suitability? If the answer is "only by guessing from the description block", the data layer is incomplete.
Second, can a crawler read the reviews in plain text outside of an iframe or a heavy widget? If the answer is no, the review signal is invisible to AI agents that do not run a full headless browser, which is most of them today.
Third, do you have explicit content somewhere on the site that answers the top five questions a customer might ask about that product, in the language they would use? If the answer is "kind of, in our blog", the use-case layer is too far from the PDP for the agent to connect it.
Then take the same questions to the live agent. Ask Gemini, Perplexity and ChatGPT what they recommend for the scenarios your customers actually face. If you do not appear, look at what brands do appear and read their data. You will see the pattern immediately.
The brands that win the next 12 months in beauty will not be the ones with the biggest TikTok presence or the loudest founder story. They will be the ones whose product data reads well to a machine. The AI is now part of the customer. Talk to it like one.