OpenAI added virtual try-on to ChatGPT this week, letting users upload a selfie and generate images that show how clothing and accessories look before purchase. The feature puts the assistant closer to Google's product visualisation tools and, more importantly, moves a visual decision into a conversational surface. For retailers the consequence is structural: the product page is no longer the only place where a shopper evaluates fit, and the data that feeds an assistant now determines whether a product can be shown at all.
Try-on arrives inside a chat window
The mechanic is simple. A shopper uploads a photo, selects an item, and receives a rendered image of themselves wearing it. Simple mechanics are what make adoption possible, because they remove the two barriers that killed earlier try-on tools: app installs and body scanning. Uploading a photo is a small ask compared with standing in a scanning booth, and it works on the device people already use.
The strategic point is where the interaction happens. Try-on inside an assistant means the shopper does not visit a retailer site to evaluate fit, and the recommendation that follows is generated by the platform rather than by the merchant. Retailers have spent a decade optimising the on-site experience; this shifts part of that evaluation upstream, to a surface they do not control.
How the image pipeline changes assortment
Why some catalogues cannot be shown
Rendering requires clean product imagery, consistent angles and accurate size data. Catalogues built from lifestyle photography, where garments are worn by models in motion, are much harder to use than catalogues built from flat-lay or ghost-mannequin shots. Retailers with disciplined asset production will find their products easier to render, and that becomes an invisible competitive advantage.
Size and fit data are the second gate. A rendered image is only useful if the shopper trusts the sizing behind it, and trust depends on published measurements and return data. Brands that publish detailed measurements and honour returns generously will see better conversion from visual tools, because the tool amplifies whatever confidence the underlying data supports.
The privacy question nobody answers well
Uploading a selfie is a meaningful disclosure, and the terms that govern it are usually written for lawyers rather than shoppers. Retailers integrating try-on need a plain answer to three questions: where the image is stored, how long it is retained, and whether it trains anything. Brands that answer those questions on the product page will convert better than brands that bury them in a policy document.
Regulatory pressure will arrive unevenly, which complicates global rollouts. The practical approach is to treat facial imagery as a special category with its own consent flow, rather than as another user-generated asset. That costs development time but avoids the scenario where a feature has to be withdrawn from a major market after launch.
Why discovery moved into assistants
Shoppers increasingly start with a question rather than a category. Adobe's data shows AI-referred traffic engaging at a higher rate and generating more revenue per visit than other traffic, 原文 which tells retailers that the assistant is not a novelty channel but a pre-purchase surface where consideration happens. Try-on accelerates that shift by moving evaluation, not just research, into the conversation.
Attributes become the language
The implication for content is direct. If an assistant must describe a garment to recommend it, then structured attributes such as fabric, fit, care and occasion become the language of discovery. Marketing copy does not survive that filter, while a well-maintained attribute table does. This is the same lesson that has been visible in search for years, now applied to a channel that reads product data literally.
What product feeds must contain now
Four fields do most of the work. Accurate material and composition, because assistants answer questions about fabric first. Fit and measurement, because that is what try-on cannot invent. Availability by size and location, because a recommendation that cannot be bought is worse than no recommendation. And a canonical product identifier, because duplicate listings split the evidence an assistant can gather.
Retailers should also audit how their products appear when asked about directly. A simple test is to ask an assistant to describe a specific SKU and compare the answer with the catalogue. Gaps in that answer are the to-do list, and they are usually cheaper to fix than a new campaign. 原文 Store associates face the same shift, moving from answering questions to correcting the answers customers already received.
How to judge whether to invest
Judge try-on by return rate and conversion on the categories where fit drives the decision, not by engagement with the feature. If returns fall on dresses and footwear while conversion holds, the tool is working. If engagement rises and returns do not move, the feature is entertainment, and the budget belongs elsewhere.
The longer view is that visual evaluation is becoming infrastructure. Google built it over four years, OpenAI has now added it to a conversational surface, and the retailers that benefit will be the ones whose product data is clean enough to be rendered, described and compared. US online holiday spend is projected to top $275 billion this year source The channel mix will keep changing, but the data requirement does not.
