ChatGPT Turned Itself Into a Fitting Room — and Your Products Are Already in It

On October 1, 2026, OpenAI pushed a “Try On” button into ChatGPT Shopping — globally, on mobile and web, with no waitlist. Shoppers can upload a selfie, tap the button on any eligible clothing or accessory listing, and watch ChatGPT render that item on their own body. The feature runs on ChatGPT Images 2.5, a model OpenAI released September 8 that the company says produces more natural lighting and richer fabric textures.

That’s the headline. The part that should keep store owners up at night is buried one paragraph deeper in the announcement.

Shoppers don’t have to discover your product inside ChatGPT to try it on. They can upload a screenshot of an outfit from any website — including yours — and ask ChatGPT to render it on their photo. Your products are in that fitting room whether you signed up for it or not.

The storefront just slid further down the funnel

Here’s what actually changed. For years, virtual try-on lived on the retailer’s site: you went to the brand, you used their tool. ChatGPT flipped that. Now the trying-on happens in the same chat window where the shopper finds the product. Discovery, consideration, visualization — all inside one conversation.

We’ve seen this pattern before. When comparison shopping moved to Google, product pages had to survive being judged in a grid next to five competitors. Now your product has to survive being judged on a stranger’s body, rendered by an AI, before the shopper ever visits your site.

The same release added a second feature that’s easier to overlook: Favorites. Shoppers can save products to a library inside ChatGPT, organize them into folders, and keep their try-on images alongside them. Think about what that replaces. It’s a wishlist that lives outside your store. The consideration phase — comparing three dresses, sleeping on it, coming back Thursday — can now happen entirely inside ChatGPT. Your site gets the visit only when the decision is already 90% made.

That has a real implication for how you measure things. If a growing share of your traffic arrives from AI interfaces, your “time on site” and “pages per session” will look worse. The browsing happened somewhere else. Don’t misread that as your site getting worse.

What it doesn’t do (read the fine print)

OpenAI’s own materials warn that the try-on images cannot confirm fit or size. That’s honest, and it matters. The technology renders drape and appearance convincingly, but it doesn’t know the garment’s actual measurements against the shopper’s body. Google launched its own virtual try-on back in 2025 with the same limitation.

And here’s the caveat we have to be straight about: nobody knows yet whether this reduces returns. Better visualization should cut the “looked different than I expected” returns. But it might also create a new kind of confident buyer who orders three sizes because the try-on looked great on all of them. OpenAI has published no conversion or return data. Anyone telling you this definitively fixes apparel returns is guessing.

There’s a second reason for humility. This isn’t OpenAI’s first attempt at commerce. Its earlier Instant Checkout push fell short of expectations — shoppers didn’t want to complete purchases inside the chat. So OpenAI pivoted: forget checkout, own the decision. That’s a smarter play, and it’s the one store owners should take seriously.

Don’t build your own virtual try-on

Let me say the thing some agencies won’t: do not spend $10,000–$20,000 building a virtual try-on widget for your own storefront right now. We’ve watched stores do this twice now — once for AR in 2021, once for “AI stylists” in 2024. The mistake is the same every time. You pay a fortune to add a feature to the page the shopper visits last, while the discovery layer gives it away for free at the page they visit first.

The money goes somewhere more boring and more valuable: your product data.

ChatGPT’s try-on and product cards are built on the information it can find about your products — titles, attributes, images, descriptions. Garbage in, garbage out. If your size data is vague, the AI’s render of your dress on a shopper is built on vague inputs. If your hero images are inconsistent, the try-on output is inconsistent. Your product feed is no longer just fuel for Google Shopping. It’s the raw material an AI uses to dress a human in your clothes.

This is the same work we keep coming back to: product descriptions that actually convert aren’t a copywriting exercise anymore. They’re infrastructure.

Five things worth doing this month

Audit your attributes like an AI is reading them — because one is. Material, fit, size accuracy, color naming. “Oversized boxy fit” beats “comfortable.” A shopper asking ChatGPT “will this be tight on me?” gets its answer from your data.

Clean up your hero images. The try-on model builds on your product photos. Blurry, inconsistently lit catalog shots produce weird renders — and the shopper blames your product, not the AI. Our rule of thumb from this morning’s piece still holds: AI visuals are fine, but real photos close the sale.

Add real fit information, not just a size chart. A size chart nobody reads is decoration. Fit notes (“runs small in the shoulders,” “model is 5’9” wearing M”), customer photos, even a short video of the garment moving — these are the inputs that make try-on outputs believable.

Check how your store looks to AI agents. ChatGPT isn’t the only AI touching your catalog. Shopping agents are already crawling product pages, comparing prices, and building carts — we wrote the readiness checklist here, and the try-on launch makes every item on it more urgent. Structured data, clean URLs, accurate stock status: none of it is glamorous, all of it is what the machines read.

Watch your referral traffic sources. Most analytics setups lump AI referrals into “direct” or “referral” without breaking them out. Segment them. You want to know — month by month — what share of your converting traffic first met your product inside a chat window. That number is only going up.

The honest bottom line

Your storefront is becoming the warehouse of the funnel: the place where the transaction happens, not where the decision happens. That’s not a catastrophe — warehouses make money. But it changes where you invest. Pretty PDP animations matter less than clean data. The fitting room on your site matters less than the one in the chat app, which you don’t control.

The winners here won’t be the stores with the fanciest on-site tech. They’ll be the ones whose products look correct, feel accurately described, and arrive as promised when an AI dresses a stranger in them. Boring fundamentals, new venue.

We’re not sure yet whether try-on shrinks return rates or just moves the uncertainty earlier in the journey. What we are sure of: the stores treating their product data as a competitive asset are about to pull ahead of the ones still treating it as a chore. October 1 was the starting gun.

FAQ

Does a try-on confirm fit or size?

No. OpenAI’s own materials are blunt about this: the images render drape and appearance convincingly, but they don’t know the garment’s measurements against the shopper’s body. Real fit notes, model measurements, and the merchant’s size chart still do that job.

Do shoppers have to discover my product inside ChatGPT for this to matter?

No, and that’s the uncomfortable part. A shopper can screenshot an outfit from any website — including yours — and ask ChatGPT to render it on their photo. Your products are in that fitting room whether you signed up or not.

Should I build my own virtual try-on widget?

The post’s answer is no: don’t spend $10,000–$20,000 adding a feature to the page the shopper visits last, while the discovery layer gives it away free at the page they visit first. Put the money into product data instead — attributes, hero images, fit notes.

Will virtual try-on cut my return rate?

Nobody knows yet, and anyone claiming otherwise is guessing. Better visualization should cut “looked different than expected” returns, but it could just as easily create confident buyers who order three sizes because the render looked great on all of them. OpenAI has published no conversion or return data.