Virtual try on for small shops: what it really cuts

Here is a thing most apparel sellers do not know about their own listings. Google's Merchant Center documentation on apparel virtual try-on states that all brands with a shopping feed and high quality imagery are automatically opted into it. Nobody signs up. If your photographs are good enough, your garments are being rendered onto other bodies already, and the only decision you were offered was an opt out you were never told about.

That is not a scandal, and the mechanics are more reasonable than the framing suggests. It does mean the question for a small shop is not whether to adopt virtual try-on. It is whether your images are getting you a good version of it, and whether the thing everybody promises it delivers, which is fewer returns, has any evidence behind it.

Are you already in it without having decided?

Probably, if you sell clothing through a feed. The threshold Google describes is a shopping feed plus high quality imagery, and the opt out route runs through Google Customer Support from your Merchant Center account. Domain owners without a Merchant Center account can complete an opt out form to have their images removed.

Worth pausing on what that means practically. Your product photograph is being used as an input to a generative model that produces a new image, which a shopper then sees attached to your listing. You did not make that image. You cannot review it before it appears. Your brand is attached to it. For most sellers that trade is fine, because a rendered garment on a body is more informative than a flat lay. For a seller whose whole proposition is the exact drape of a specific fabric, it is worth thinking about.

Two systems, not one, and they behave differently

The confusion in every article written about this comes from collapsing two products into one name. They are documented separately and the differences matter commercially.

Read the second row twice if you buy Shopping ads. The documentation on how the Try-on tool works says the shopper facing button appears only on non sponsored product results, while the apparel try-on page says qualifying products are eligible across paid Shopping ads. Those are different features doing different jobs, and a seller who tests only the sponsored placement will conclude, wrongly, that none of this applies to them.

Does virtual try on actually reduce returns?

Nobody credible has told you that it does, and the silence is the interesting part. Neither Google help page makes a returns claim. The announcement of the try on update describes a custom image generation model for fashion that understands the human body and the way materials fold, stretch and drape, and frames the benefit as seeing how a garment looks on you. It talks about a Shopping Graph of more than 50 billion product listings. It does not talk about returns.

The vendors selling try-on widgets do make that claim, usually with a percentage attached and no methodology. Be careful with those numbers, because the mechanism they imply is doing a lot of work. A try-on render answers a styling question: does this shape suit me, does this colour work with my hair. It does not answer a sizing question, and sizing is what drives apparel returns.

There is a plausible second order effect worth naming honestly, because it cuts the other way. If a render makes a garment look better on a shopper than it will in life, you sell more units and take more of them back. A tool that increases conversion without increasing fit accuracy increases returns in absolute terms. Whether it increases the return rate depends on which effect is larger, and no published data settles that for a small shop.

What do your product images have to look like?

The specification is short enough to act on this afternoon. At least 512 by 512 pixels, ideally 1024 or higher. One garment per image. A front facing model or mannequin in a simple pose with arms down at the side, or the garment laid flat with folds and wrinkles minimised. Minimal obstruction of the item.

Every one of those constraints exists because the model has to segment the garment from everything else in the frame. Arms down means the sleeve outline is unambiguous. One garment means it does not have to guess which layer you are selling. Flat and unwrinkled means the fabric edge is findable. A styled editorial shot with crossed arms and a jacket over the shoulders is a lovely photograph and a hostile input.

The practical consequence for a small shop is that you now have two image jobs rather than one: the photograph that sells, and the photograph that machines can read. They are not the same picture, and the second one is cheap to produce once you know it is a separate deliverable. The same split shows up in the rules covered in our piece on what marketplaces require from a product photo, and it is the recurring theme of the last two years: your catalogue is now read by software before it is seen by a person.

Is fit prediction a different product from try on?

Entirely, and conflating them is the most expensive mistake in this area. Try-on is a rendering problem: given a garment and a body, produce a plausible image. Fit prediction is an inference problem: given this person's purchase and return history, and the measurements of this garment, will they keep it.

The second one is what actually moves returns, and it needs data you may already have. Order history with sizes. Return reasons, if you collect them properly. The garment's real measurements rather than its size label, which is the piece almost nobody has in structured form. A size 12 from two of your suppliers are different garments, and until that is in your catalogue as numbers rather than as a letter, no model can help you.

That is the unglamorous work, and it is worth more than any widget. A size guide built from your own measured samples, per supplier, changes returns in a way a render does not. Our notes on which parts of returns management are safe to automate go through where the automation genuinely pays, and measurement capture is the input that everything else depends on.

How do you check whether the render is any good?

You cannot preview it, which is the frustrating part, but you can inspect the output the same way a shopper does. Open the Shopping tab in a private window, find your own product, and look at what comes back. Then look at three things in order.

Silhouette first. Does the garment keep its actual shape when it is placed on a body, or has the hem crept up, the neckline changed, the sleeve length shifted. Shape is the thing a returns email will be about, and it is the thing a segmentation error breaks first.

Colour second, and be strict here. Rendering a garment onto a new scene changes the light on it, which changes the colour a shopper perceives. If you sell a navy that is easily mistaken for black, or a cream that photographs as white, the rendered version will make that worse rather than better. That is a returns risk you can partly control by improving the source photograph, since a flatter, better lit input gives the model less to reinterpret.

Detail third. Prints, logos, buttons and any text on the garment are the parts most likely to be reconstructed rather than preserved, for the same reason small text fails in every other generated image. A graphic tee is the hardest case in your catalogue and the one worth checking by hand.

If any of the three comes back wrong, the lever you have is the input image rather than a complaint. A cleaner front facing shot at 1024 pixels or better is the whole of your control surface, which is a good argument for producing those shots deliberately rather than hoping your editorial photography happens to qualify.

What does a Shopping Graph of 50 billion listings mean for you?

That this is a volume surface, not a boutique one. Google puts the Shopping Graph at more than 50 billion product listings, from global retailers to local shops. Try-on is being applied across that scale automatically, which has two consequences that pull in opposite directions for a small seller.

The good one is that eligibility does not depend on your size or your budget. The gate is a feed and adequate imagery, both of which a one person shop can meet, and the enrolment is automatic. There is no minimum spend and no partnership to negotiate, which is unusual enough to be worth noticing.

The awkward one is that a rendering system applied at that scale is tuned for the average garment, and your product is interesting to you precisely where it is not average. An unusual cut, an unusual fabric or an unusual construction is exactly what a general model handles least well. The smaller and more distinctive your range, the more likely the rendered version flattens what makes it worth buying.

What if you do not sell clothing?

Most of this does not reach you, and the honest advice is to stop reading the try-on coverage rather than to look for an equivalent. The consumer tool covers shoes, tops, bottoms and dresses, and explicitly excludes lingerie, bathing suits and accessories. If you sell homewares, tools, food or anything else, there is no analogous automatic enrolment to worry about.

What does carry across is the underlying lesson, and it is the more valuable half. Your product images are now inputs to systems you do not operate, and the qualities that make an image useful to a machine differ from the qualities that make it attractive to a person. That is true for a saucepan on a marketplace feed as much as for a dress on the Shopping tab, and the shop that produces both kinds of image deliberately is doing something its competitors are doing by accident.

What about the shopper's photograph?

It is the part customers ask about and the part sellers cannot control, so it is worth knowing the stated position. Google says it does not collect or store biometric data during the experience, and that people's photos are not used for training purposes. That is a clear statement and it is the one to quote if a customer emails you about it, since they will assume the retailer is involved.

You are not the data controller for that upload, which is a genuine relief, but you are the brand the customer associates with the experience. If you sell in Europe, being able to answer the question accurately in your own words is part of the job, and it sits alongside the broader consent questions we set out in the piece on where consent limits ecommerce personalisation.

Does any of this change your returns obligations?

No, and it is worth remembering what the baseline is before you spend money reducing it. For distance sales in the European Union, the consumer has the right to return the item within 14 days, described in the Your Europe guidance for businesses as the cooling off or withdrawal period, and no reason or justification has to be given. That right does not shrink because your listing had a better picture.

Which reframes the whole exercise. You are not trying to eliminate returns, you are trying to stop the ones that were caused by a misunderstanding you could have prevented. That is a much smaller target and a much more achievable one, and it responds to information rather than to imagery.

What we would do this quarter

Check whether you are in it, first. Look at one of your own listings on the Shopping tab, logged out, and see whether a try on control appears. Five minutes, and it settles a question most sellers are guessing about.

Then produce the plain machine readable shot for your twenty best selling garments: front facing, arms down, one item, 1024 pixels or better. It is a morning's work and it upgrades the version of your product that a generative system produces, which you otherwise have no control over at all.

Finally, measure ten garments and publish the numbers. Chest, length, waist, in centimetres, per item rather than per size. That single change does more for returns than anything in this article, costs nothing, and works on every channel you sell through rather than only on the one that rendered the picture. If your product pages cannot carry that data cleanly, that is a storefront limitation rather than a merchandising one, and our ecommerce website builder generates pages whose fields you can add to, which matters the day you decide a size table belongs in structured data rather than in an image.

The rendered image is the visible half of this and the smaller half. What shoppers keep is decided by numbers you have not written down yet, and the piece on how an AI shopper reads a product page makes the same argument from the discovery side.