Asos AI Stylist Sends Shoppers to Competitors When Inventory Falls Short

Asos launched Stylist in ChatGPT this month, a shopping assistant that surfaces fashion picks and video content for UK and US customers. The app runs on Bambuser’s video commerce platform and turns Asos’s product library into machine-readable data that ChatGPT can retrieve and return as shoppable videos. Shoppers ask for outfit ideas, browse by occasion or trend, and click through to buy on Asos. The pitch is frictionless discovery inside an environment where 17 million people already spend time. The execution reveals a structural problem no one designing AI commerce tools wants to admit: the AI will always try to be genuinely helpful, and genuine helpfulness does not respect single-retailer distribution.

Steve Webster, an e-commerce executive whose career includes stints at Barbour and Liwa Trading Enterprises, tested Stylist and documented the failure mode in a LinkedIn post. He asked the app to build a smart casual wardrobe for a middle-aged man. Stylist returned a Mango blazer, Jack & Jones shirts, Thomas Crick Evers trainers. Competent enough. Then it reached fragrance. The AI recommended Tom Ford Oud Wood, a luxury scent that fits the brief perfectly. It also linked directly to tomfordbeauty.co.uk, a competitor Asos does not stock. Within three minutes, the AI stylist had sent a paying customer to another retailer.

What Bambuser’s Intelligence Layer Actually Does

Bambuser’s new Intelligence Layer converts Asos’s video library and product catalogue into structured data that large language models can process and retrieve in real time. The system ingests product metadata, video timestamps, styling context, and availability flags, then packages everything so ChatGPT can return shoppable video clips alongside product cards. The technical architecture is sound. A human stylist working exclusively for Asos would navigate the Tom Ford gap by suggesting a different fragrance the retailer actually carries. ChatGPT does not operate within those constraints.

The AI optimizes for the stated need, not for the conversion rate. When a customer asks for a specific product category and the retailer’s catalogue does not contain the best answer, the model fills the gap with whatever it knows. ChatGPT knows Tom Ford Oud Wood is the right fragrance for a mature man building a smart casual wardrobe. It also knows Asos does not stock it. So it linked out, because the right answer for the customer is not always the commercially convenient answer for the retailer.

The Attribution Problem No One Is Measuring

Webster raised a question that will prove uncomfortable for every retailer building on top of third-party intelligence layers: for every session that produces a shoppable Asos purchase, how many produce a visit to Tom Ford, Hermès, or wherever else the AI concluded was the better answer? The channel looks like distribution. In some sessions it probably is. In others, Asos is funding a sophisticated referral engine for everyone else.

No public data exists on how often AI shopping assistants route customers to competitors. The metric is not tracked in standard analytics dashboards, and the platforms hosting these tools have no commercial incentive to surface it. A session that ends with a click to tomfordbeauty.co.uk still counts as engagement. Whether that engagement converts into revenue for the retailer who paid to build the experience is a different question.

The structural tension is this: AI models are trained to be helpful across the entire internet, not helpful within the boundaries of a single retailer’s inventory. When you build a commerce experience on top of someone else’s intelligence layer, you inherit that layer’s associations, its breadth of knowledge, and its definition of a good answer. A good answer for the customer is not always a good answer for the retailer.

Why Human Stylists Do Not Have This Problem

A human stylist working for Asos would never recommend a product the retailer does not carry. The constraint is built into the job. If a customer asks for Tom Ford Oud Wood, the stylist pivots to a fragrance Asos stocks that shares similar notes or positioning. The recommendation is still helpful, but it operates within commercial boundaries the stylist understands implicitly.

ChatGPT does not understand those boundaries because it was not trained to respect them. The model’s objective is to provide accurate, useful information. When the most accurate answer involves a product outside the retailer’s catalogue, the model provides it. The fact that this behavior undermines the retailer’s business model is not a bug the AI recognizes. It is a feature of how the system was designed.

Webster’s observation is not a criticism of the ambition. Placing a brand inside an agentic platform where millions of customers already spend time is a reasonable strategic bet. The Bambuser integration is technically well conceived. The problem is structural, not executional. The AI will always try to be genuinely helpful, and genuine helpfulness and single-retailer distribution are not the same objective.

What This Means for AI-Mediated Commerce

The Asos case is not an outlier. Every retailer building shopping experiences on top of third-party AI platforms will face the same tension. The more helpful the AI becomes, the more likely it is to recommend products the retailer does not carry. The less helpful it becomes, the less reason customers have to use it.

One solution is to constrain the AI’s knowledge base to only the retailer’s inventory. This eliminates the competitor-referral problem but introduces a new one: the AI can no longer answer questions about products the retailer does not stock, which makes it less useful than a standard search bar. Another solution is to accept that some sessions will route customers elsewhere and treat the AI as a top-of-funnel awareness tool rather than a direct conversion channel. This requires a different attribution model and a willingness to fund discovery that does not always convert.

A third option is to build the intelligence layer in-house, training a model specifically on the retailer’s catalogue and styling philosophy. This is expensive, time-consuming, and requires machine learning expertise most retailers do not have. It also does not solve the underlying problem: customers will still ask for products the retailer does not carry, and the AI will still need to decide whether to admit the gap or pretend it does not exist.

The Uncomfortable Question Retailers Are Not Asking

The fundamental question is whether AI-mediated commerce serves the retailer or the customer. If the objective is to maximize customer satisfaction, the AI should recommend the best product regardless of who sells it. If the objective is to maximize retailer revenue, the AI should recommend only products the retailer carries, even when better options exist elsewhere. These objectives are not compatible.

Most retailers building AI shopping assistants have not decided which objective they are optimizing for. The assumption is that the two objectives align, that helping customers find what they want will naturally drive revenue to the retailer. The Asos case proves that assumption is false. The AI helped the customer find the right fragrance. It just did not help Asos make the sale.

Webster’s post has not prompted a public response from Asos or Bambuser. The silence is telling. The problem he identified is not unique to Asos, and it is not a problem any retailer has figured out how to solve. The AI will always try to be genuinely helpful. Genuine helpfulness and single-retailer distribution are not the same objective. Until retailers decide which one they are optimizing for, every AI shopping assistant will face the same structural tension.

Frequently Asked Questions

What is Bambuser’s Intelligence Layer?

Bambuser’s Intelligence Layer is a capability that converts a retailer’s product catalogue and video library into structured, machine-readable data that large language models can process and retrieve in real time. It allows AI platforms like ChatGPT to return shoppable video content and product recommendations based on customer queries.

Why did Asos’s AI stylist recommend a competitor’s product?

ChatGPT recommended Tom Ford Oud Wood because it is the most accurate answer to the customer’s request for a fragrance suitable for a smart casual wardrobe. The AI does not restrict its recommendations to products Asos carries; it optimizes for the best answer based on its training data, which spans the entire internet.

Can retailers prevent AI assistants from linking to competitors?

Retailers can constrain the AI’s knowledge base to only their own inventory, but this makes the assistant less useful when customers ask about products the retailer does not stock. Another option is to build a proprietary AI trained exclusively on the retailer’s catalogue, though this requires significant machine learning expertise and investment.

How common is this problem among AI shopping assistants?

No public data tracks how often AI shopping assistants route customers to competitors. The metric is not part of standard analytics dashboards, and platforms hosting these tools have no commercial incentive to surface it. The Asos case suggests the problem is structural and likely affects every retailer using third-party AI platforms.

What is the best way for retailers to use AI shopping assistants?

Retailers must decide whether they are optimizing for customer satisfaction or conversion rate. If the goal is awareness and discovery, accepting that some sessions will route customers elsewhere may be acceptable. If the goal is direct revenue, constraining the AI to the retailer’s inventory is necessary, even if it reduces the assistant’s usefulness.