ChatGPT will show you wearing the clothes. It will not promise they fit.

Three taps, one generated image

ChatGPT's release notes for 1 October 2026 added a line that sounds small and is not: "A 'Try on' button will appear on product listings for clothes and accessories in ChatGPT."

The flow is three steps. Select Try on from a product listing. Take or upload a selfie. View the generated image. You can also skip the button entirely, upload a photo of a garment into any conversation, and ask ChatGPT to show how it could look on you.

Note the verb in OpenAI's own documentation. The try-on is "created with ChatGPT Images." It is not a fit simulation, a measurement model, or a render of the actual garment on your actual body. It is an image generator making a picture of a plausible you wearing a plausible version of that thing.

That distinction is the whole article, and OpenAI puts it in writing.

The disclaimer is the most useful sentence on the page

Here is the line, from the Shopping with ChatGPT Search help page: "Try-on images may not represent the product or your appearance exactly and do not guarantee fit or size. Check the merchant's measurements, product details, and return policy before buying."

Read that against the feature's promise, which is to "see how the item could look on you." Both things are true at once. The picture will show you something. What it will not do is tell you whether the medium fits, whether the sleeves are the length shown, or whether the colour on your screen is the colour in the box.

This is a generative image standing in for a physical measurement, and generative images are confident by construction. They do not have an error bar. A garment that is cut wrong for your shoulders will still render beautifully on your shoulders, because the model is drawing what the request implies rather than what the tape measure says.

So the honest framing is narrow and still genuinely useful: try-on answers "would I wear this" and does not answer "does this fit me." Those are different questions, and only one of them drives returns.

Your selfie is kept until you delete it

The part that deserves a slower read is storage, because the default is retention.

"Your reference photos are saved for future try-ons, so you can reuse them without uploading them each time," the help page says. Convenient, and also a standing copy of your face and body held against your account until you act. Removal is manual: Settings, then Personalization, then Reference photos, then select the photo and either change it or use the delete icon.

Nothing here is hidden, and OpenAI documents the control clearly. But the asymmetry is worth naming. Uploading takes one tap inside a shopping flow you were already in. Deleting takes four deliberate steps inside a settings menu you have no reason to visit. Features that collect a body photo by default and release it only on request will accumulate body photos.

If you try this once out of curiosity, the useful habit is to go and clear the reference photo afterwards.

The 19.3% problem this is aimed at

Why build it at all? Because online apparel has a returns problem large enough to justify almost any intervention.

The National Retail Federation and Happy Returns estimated that 19.3% of all online sales would be returned in 2025, part of a total they put at $849.9 billion across retail, with the overall return rate at 15.8% of annual sales. Those are estimates from a named industry body, not hard counts, and they cover retail broadly rather than clothing alone. They still establish the scale. Roughly one in five things bought online comes back.

A tool that reduces "I could not picture it" purchases by even a few points is worth real money to merchants, which is why the plumbing underneath this matters more than the party trick on top. For merchants on Shopify, the help page notes, product data is already integrated into ChatGPT through Shopify Catalog, with no additional work required from individual merchants. OpenAI also runs a direct product feed programme for merchants who want their own data flowing in. We looked at the other end of that pipe when Shopify shipped its own chat-driven store builder.

What else is in the fine print

The same help page quietly documents several things that have nothing to do with try-on and are more likely to cost you money.

Prices first. "Prices shown in ChatGPT's initial response typically reflect the price from the first merchant listed, which may not be the lowest available price." The cheapest option, if OpenAI knows of one, surfaces on the product detail page rather than in the carousel you see first.

Then the labels. Badges like "Budget-friendly" or "Most popular" are generated by the model, and OpenAI states plainly that they are "not guarantees or verified statements, and may not reflect all available market data." A "Budget-friendly" tag might only mean reviewers mentioned good value. Review summaries are model-generated from public websites and are "not verified by OpenAI."

One clarification is worth crediting, especially given what OpenAI announced four days later. "Product results are selected independently by ChatGPT and are not ads, nor influenced by any OpenAI partnerships." That is a firm commitment, and it is now load-bearing, because the company is also testing a paid visual ad format inside ChatGPT. Shopping results and advertising are separate systems today. Watch whether that separation reads as clearly a year from now.

Our take

Virtual try-on is a fun, genuinely useful shopping aid and a bad substitute for a size chart, and OpenAI says so itself in the only place most people will not look. Use it to decide whether a jacket is your kind of jacket. Do not use it to decide whether it is your size.

The two things we would actually change about the default: clear your reference photo when you are done, and click into the product detail page before you buy, because the price in front of you is the first merchant's price and not necessarily the best one. Both are one-line fixes to a feature that is otherwise better documented than it is marketed.