198 years old. Over 1,100 perfumes made. Zero paid influencer campaigns — until now. What broke the streak? A viral $660 fragrance, followed immediately by the dupe conversation, followed by the realization that staying quiet is itself a statement, and not the one they wanted.

This is the oldest story in luxury social: you can ignore the algorithm until your product becomes the subject of it. The price point didn't insulate Guerlain from TikTok; it made the dupe comparison more legible. Mystery and restraint are the precise qualities that creator content cannot preserve and cannot afford to ignore simultaneously.

We wrote last week about Pinterest pitching luxury brands on appearing when taste is still forming. Guerlain spent 198 years forming taste through other channels. Paying an influencer is the admission that those channels no longer reach first. The dam has cracked.

"Wenn KI einkauft: Warum Betrugsprävention im Check-out neu gedacht werden muss" — when AI shops, why fraud prevention needs rethinking. Pascal at etailment frames the problem cleanly: velocity rules, device fingerprinting, and behavioral biometrics were calibrated for humans. An AI agent that purchases at 3am, skips the browsing phase, and ships to an address it hasn't used before looks exactly like a fraudulent transaction. Except it isn't.

This is the same gap Neritus Vale identifies from the other direction in today's piece on persistent agent memory arriving before retailers built the customer database to hold it. The agent arrives. The fraud stack blocks it. The customer complains. The fraud team says working as intended. At scale, this incident is coming, and whoever triggers it first will force the conversation about agent identity standards that nobody has needed to have yet.

The argument: AI agents shop for you, brands lose the discovery layer, the impulse moment, the browsing journey — therefore AI threatens online shopping as we know it. The conclusion is probably right. The implication that this is bad is less obvious. AI doesn't destroy online shopping; it destroys the friction that online shopping built its average order values on. Fourteen clicks to checkout. Three contradictory size charts. Returns framed as a feature. If an agent sweeps that away, the things that die deserved to.

What's harder to argue is that discovery — real discovery, not "customers also bought" — survives the agent layer without deliberate redesign. That's where Neritus Vale's piece today on the training data problem becomes load-bearing. The agent can only discover what it was taught to recognize. If the training set doesn't include the customer who actually exists, discovery fails before the fraud stack gets a chance to.

Founder Alana Pallister literally sold her house to fund 300,000 units. She sold a million. Now the bet is Coachella. The tracksuit math worked because it was one product with no competition at its specific intersection of stretch, silhouette, and price. Coachella is a crowded field where every brand with a festival strategy is competing for the same iPhone shot. The logic isn't irrational — virality compounds when the brand has already demonstrated it once — but the terms change. The first million units were a surprise. The second viral moment has to be designed.

Connection worth noting: I.AM.GIA and Topshop (more on that below) are using a cultural event as their primary distribution mechanism from opposite brand trajectories. I.AM.GIA is scaling up from a garage; Topshop is scaling down from a collapsed empire. Both end up at the same place — the event as the brand. The more interesting comparison is with today's Swan Beauty piece: one brand built product first and stumbled into a moment; the other found the moment and is now building product credibility. One of these sequences is easier to repeat.

Bain puts the Chinese market on track for RMB 1.5 trillion in e-commerce. Worth holding against Alibaba's ¥380 billion AI infrastructure pledge — which we read as defensive capex, not growth investment. Securities Times (证券时报) ran the sharper Chinese-language version this week: Taobao's "dual engine" has stalled while Pinduoduo, JD, and Douyin accelerate past it. The market grows regardless. The question is whose infrastructure earns the margin on 1.5 trillion yuan of transactions. Alibaba's answer is to own the cloud layer underneath everyone else. That is the bet, and it is not a small one.

Snap is rolling out augmented reality try-on tools for fashion retailers — in-app fitting rooms, product overlays, the familiar pitch. The timing lands one month after the Tstars-Tryon paper demonstrated that virtual try-on model quality is effectively solved and the bottleneck has shifted to catalog metadata: structured size, fabric drape, and fit attributes that most retailers still don't maintain at the required granularity. Snap's AR surface is only as good as what brands feed it. The model problem is solved. The data operations problem is not, and Snap cannot fix it for you.

Southeast Asia's e-commerce market is projected at 16% growth this year, with video commerce and AI cited as structural drivers. TikTok Shop's 84% beauty surge is the loudest data point, but the 16% aggregate hides serious variance: Indonesia and Thailand, which have dense creator infrastructure and established live commerce formats, are pulling the average up. Markets without that infrastructure are growing at roughly half the rate. Video-as-discovery is not evenly distributed, and the gap is widening faster than regional policy can respond.

While Sir John Crabstone's piece today covers the AI mirror startup that accidentally went viral on a bachelorette jet, the Pheloung sisters ran a deliberate version of the same cultural play. The jet had a name before anyone boarded. The content strategy was set. In 2026, the engineered stumble and the actual stumble produce the same coverage. The difference shows up in the second campaign, when one brand has a repeatable system and the other has a great story.

The Met Gala is controversially attended, controversially dressed, controversially sponsored — and beauty brands showed up in force anyway. Gifting suites, editorial adjacency, carpet adjacency. The calculation is simple: controversy is part of the moment, and attaching to the moment is cheaper than originating one. This is the same logic we traced in the Devil Wears Prada 2 partnerships. Beauty doesn't need the Met Gala to be tasteful. It needs it to be watched.

Shark Beauty and Lookfantastic are joining an AI-driven Topshop catwalk event that incorporates live social commerce. Topshop — collapsed in 2021, became an ASOS digital license, then a quieter IP property — is now a format for live commerce experiments. There are worse fates. If the catwalk sells Shark Beauty products and Lookfantastic bundles while models wear Topshop designs, Topshop has found a business model the physical estate never quite landed. The brand becomes a stage set. Philip Green did not have this in mind. Stages are useful.

H&M and Lotto launch a football-heritage collection timed for summer. The mid-market athletic crossover is increasingly crowded at every tier: Lululemon owns premium performance, Forta is betting today that performance beauty is the unsolved gap, and H&M is reminding the floor that sporty is a visual language, not a performance category. Lotto's Italian 1980s archive gives H&M credibility it couldn't build from scratch. Fair trade.

McKinsey publishes a live commerce explainer. Thorough. Accurate. Would have been timely in 2023. The Gulf had built live commerce infrastructure before McKinsey wrote the MENA playbook. China's platforms were running it at scale heading toward 1.5 trillion yuan annually when this piece was being commissioned. McKinsey is not wrong. McKinsey is just on a different time zone from the industry it describes.

AI e-commerce startup Onton raises $7.5 million. The headline doesn't explain what Onton does beyond "AI e-commerce." The deck almost certainly says "personalized discovery" or "AI-native shopping assistant." We noted last week that retailers have launched AI shopping apps faster than benchmarks exist to evaluate them. The same gap applies to the startups selling the underlying tools. Raise fast, define later.