Wet Claude be shoppin'- An AI retail experiment gone wild

What happened

Wet Claude be shoppin' — An AI retail experiment gone wild

Deana Burke set out to answer a mundane question: when an AI recommends a product while you shop, how does it actually decide what to recommend? To find out, she built a fake deodorant brand, watched how quickly major large language models picked it up, and started posting her findings to Instagram under the handle wetclaude. Within a few weeks, the account exploded from 400 followers to over 80,000 — not because Burke had cracked AI's product-ranking code, but because the question she was asking turned out to touch something much rawer than consumer tech.

This piece covers Burke's research approach to what she calls agentic commerce, the social dynamics that turned a niche experiment into a viral culture-war flashpoint, and what the episode reveals about the opacity of AI recommendation systems. It is aimed at product builders, brand marketers, and anyone trying to understand how LLMs are shaping the next layer of online shopping.

How it works

Deana Burke, a co-founder of the news aggregator Clone and a long-time collaborator with the media company Dirt, invented a fake deodorant brand and documented how long it took leading large language models to recommend it. The answer, she reported on Instagram, was not long. Dirt CEO Daisy Alioto, who had known Burke for almost five years through their joint work on Clone and the media podcast Tasteland, interviewed her for Dirt's publication on August 27, 2026. By that point Burke had already branched beyond the shopping thesis, posting about AI agents passing "mind viruses" — persistent obsessions, like a special interest in whales — between one another, about how ChatGPT decides what brands to surface, and about what the Hugging Face security incident reported by CNBC on August 26, 2026 means for data security.

The account's jump from 400 to over 80,000 followers happened in just a few weeks, a speed Burke described as a shock. The name wetclaude references an older internet meme that imagined an unconstrained, "uncucked" version of Anthropic's Claude — an AI not limited by safety guardrails. Burke had claimed the TikTok and Instagram handles when that meme first circulated and had been posting under the handle advilextrastrength before that pivot.

Burke's method is deliberately low-tech and accessible. She uses AI tools to write Python scripts and sort the data she collects, describing this as a kind of "renegade research" unlocked for non-technical people by the current generation of coding assistants. Her core approach is behavioral: rather than attempting to audit model weights or training data — which she acknowledges is impossible from the outside — she queries LLMs with shopping prompts and logs what comes back. She then looks for patterns across repeated queries and across different models, treating recommendation outputs as the only legible signal available.

Why it matters

For the mind-virus experiments she went further, testing whether a behavioral quirk or framing introduced into one AI agent conversation could propagate into or influence another agent's behavior. She has also been systematically scraping and logging every comment on her TikTok and Instagram posts, sorting them by theme to track AI sentiment in the public. The approach mirrors what search engine optimization researchers did in the early days of Google: measuring outputs because the inputs are opaque. Burke's framing is that LLM recommendation ranking is essentially a black box in the same way early SEO was, before the industry developed legible models of how PageRank operated.

AI-powered shopping recommendations are on a trajectory to become the primary discovery layer for consumer products, and right now the mechanics of that layer are almost completely unauditable by brands or consumers. Burke's observation — that being recommended by an AI is going to be worth an unknown but large amount of money, and that no one knows how it works — describes an emerging asymmetry of the same kind that shaped a decade of search marketing. Companies that spent the 2010s ignoring SEO lost distribution; companies that ignored social algorithms lost attention. The question Burke is probing is whether brands are about to make the same mistake with LLM recommendation ranking.

Who is affected

The secondary finding is about interpretive framing. Burke notes that her content is claimed equally by people who are pro-AI and people who are anti-AI, each reading the same research as confirmation of their position. Her mind-virus video drew comments spanning Israel-Palestine, trans issues, the woke mind virus, and MAGA, with AI functioning as a container for pre-existing political anxieties. This dynamic matters because it suggests that public discourse about AI capabilities and risks is being filtered heavily through culture-war lenses, making neutral empirical measurement politically legible in ways its producers do not intend.

Brands and e-commerce operators are the most directly exposed group. If LLMs are already recommending products — including fake ones created as experiments — then any brand without a deliberate strategy for AI-era discoverability is effectively leaving recommendation placement to chance. Direct-to-consumer brands with small SEO budgets that historically relied on social discovery are particularly vulnerable, since an AI shopping layer could route around influencer marketing and paid social entirely. The Hugging Face incident Burke referenced also flags a supply-chain risk for AI-dependent workflows: if training data or model infrastructure is compromised, the recommendations themselves could be manipulated upstream in ways that are invisible to downstream brands and consumers.

The 80,000-person audience that assembled around wetclaude in weeks also suggests a large population of non-technical people who are motivated to understand how AI systems make consequential decisions but have no current venue for that understanding. Developers and product teams building shopping agents face the reciprocal problem: their systems are making brand and product recommendations that carry commercial weight, and there is no established accountability framework for how those recommendations are generated or contested.

What to watch next

The most concrete thing a builder or brand should verify now is whether their products appear at all in LLM shopping responses, and under what prompting conditions. Burke's experiment shows that a fabricated brand can enter LLM recommendations without any formal submission or optimization process, which implies that LLM product knowledge is being absorbed from web crawls and training corpora in ways that current brand teams are not monitoring. Testing which models surface which products, and whether responses shift across prompt phrasings, is the minimum viable audit.

Longer term, the wetclaude project is attempting to build what Burke calls shared group experimentation — a community of people running their own probes and sharing results. Whether that community forms, and whether it generates findings legible enough to pressure AI labs or regulators toward disclosure, is the more structural question. Burke herself anticipates a cease-and-desist from Anthropic over the wetclaude name, which would itself be a data point: it would signal that Anthropic is watching the account closely enough to act, and that the research is landing somewhere it registers.