# AI Virtual Try-On Demo Red Flags

A try-on demo is a red flag when every image is a hero SKU, failures stay hidden, and conversion is defined only on shoppers who already finished a live preview.

**A virtual try-on demo is a sales sample until it runs your catalog, including a garment the model handles badly. Hero-only images, a buried failure, and a conversion rate counted only on shoppers who finished a preview are the usual tells. Ask for one garment on three different photos, break the upload on purpose, and make the vendor define each metric in the room. If they will not, stop the meeting.**

![Skeptical merchant looking over a laptop at a glossy demo photo pinned beside a wrinkled real catalog print](/images/blog/merchant-100/virtual-try-on-demo-red-flags.webp)

*The glossy portrait and the wrinkled catalog page are not the same test. Editorial image in Classic Antla disposable-camera style.*

Demos are supposed to look good. That is not the problem. The problem is a demo that can only look good, because the inputs were chosen to hide the exact failures your catalog will hit on a Tuesday. You are not rude for bringing a bad SKU. You are the only person in the meeting who has to refund it.

This is a list of tells. It is not a script of every question to ask, and it is not a ranking of vendors. A clean answer on one row does not cancel a refusal on another.

## Check cherry-picked images

The first red flag is the sample set. Count what you were shown before you admire it.

**Every garment is easy.** Solid color, front light, no logo, no sheer panel, a model pose that already looks like a try-on. That sample cannot predict a printed dress. Ask to see, in the same session, one all-over print, one chest logo or wordmark, and one dark garment. If those assets are "on a different environment" that cannot be opened today, write "not shown."

**Every shopper photo is a studio portrait.** Even skin, hair off the collar, arms in a position the model likes. Your customers will upload mirror selfies and cropped camera-roll pictures. [Product photography versus a try-on preview](https://antla.io/blog/product-photography-vs-ai-virtual-try-on) is the merchant-side version of this gap. The demo version is simpler: if they will not use a photo you took in the hallway, they are showing a campaign, not a product.

**The garment changes in flattering ways.** A pocket appears. A hem straightens. A graphic mirrors. People in the room will call this a style improvement. It is a different product. Put the source product photo and the output side by side and tap the differences with a finger. If the presenter talks over the comparison, that is the tell.

**You are not allowed to pick.** "This one demos better" means the other one demos worse. The worse one is the test. [How fashion try-on technology actually behaves](https://antla.io/blog/virtual-try-on-technology-fashion) is useful background for a teammate who thinks all previews are filters. Send it after the meeting. During the meeting, stay on the image they did not want to open.

A vendor can still be worth buying after an ugly sample. They cannot be worth buying if the only samples are beautiful and the ugly ones are postponed.

## Demand real catalog tests

Cherry-picking is what they brought. The catalog test is what you bring. A demo that refuses your files is the red flag, even if the refusal is polite.

Bring three product images exported from Shopify, not re-shot for the meeting. Include the awkward one: the back-biased first image, the colorway that is only a swatch, the collaboration graphic. Bring two shopper photos your team is allowed to use. One should be imperfect.

Then watch for these refusals:

- They need to "prep the images" overnight. Prep can be legitimate for a pilot. Prep inside a live demo usually means a person will retouch the failure out.
- They can run your garment on their model, but not your model, or the reverse. You need both, because you do not control which one will be weak in production.
- The result is described instead of shown. "It usually handles lace" is not a result.
- The output cannot be compared to the source because the source is "not on this slide." Put them on one screen.

[Choosing a try-on app for a Shopify fashion store](https://antla.io/blog/best-virtual-try-on-shopify-fashion) is a separate decision from this meeting. Use it when you are building a shortlist. Do not let a shortlist replace the ten minutes with your files. I have seen a shortlist of strong brands lose to a folder of PNGs.

If they run your files and the logo survives, credit that in the notes. If they run your files and the logo moves, you learned more than the slide would have told you. Either outcome is a successful demo. A demo with no outcome is the failure.

Antla should be held to the same standard. A fair session lets you upload your own product image and a shopper photo, including one that fails, and look at the result before anyone discusses a plan. [The app listing](https://apps.shopify.com/antla) is how you get into a dev store for that test. It is not evidence that your SKU will survive. Per [Antla's privacy policy](https://antla.io/privacy-policy): Images are deleted and not accessible within 72 hours.

## Test failure recovery

Working software shows you the break. Demo software hides the break behind a spinner or a second sample.

Force three failures. None of them require cruelty. They require five extra minutes.

1. **A bad upload.** A tiny image, a screenshot of a product page, or a photo of a group. The honest behavior is a clear error that tells the shopper what to try instead. The red flag is a confident image of the wrong person, or a spinner that ends by showing the vendor's default model as if it were the shopper.
2. **A product image the model cannot parse.** A flat lay shot from far above, or a detail crop. The honest behavior is "we cannot use this image," aimed at you, the merchant, so you can swap the source. The red flag is a hallucinated garment presented as the SKU.
3. **A slow network.** Turn on a throttle, or step into a stairwell. Note whether the shopper can cancel, and whether a second tap creates two generations. A demo that only runs on the showroom Wi-Fi has not shown you the mobile page.

Also close the dialog and change the variant. If the next color still wears the previous dress, the failure is in state, not in the model. That one reaches production even when every sample image looked sharp.

Write down the exact words the shopper would see. If the words are "something went wrong" with no next step, you will meet that string in support. [Merchandising from try-on data](https://antla.io/blog/try-on-data-merchandising-decisions) only works if failures are visible as failures. A pipeline that converts every failure into a pretty image will poison the report.

## Confirm analytics definitions

The quiet red flag is a percentage with no denominator. Ask them to write the sentence, on the whiteboard, for every number in the deck.

**"Conversion rate."** If it is orders divided by shoppers who finished a preview, say so in the notes and do not compare it to your store conversion rate. Those populations are different. People who complete a preview already wanted a closer look. A higher rate in that group can be real interest and still be the wrong number to forecast with. If the presenter slides that rate next to "your site converts at" in the same breath, stop and separate the sentences.

**"Lift."** Lift needs a comparison group. Try-on users versus other shoppers on the same stores is one comparison, and it is not a holdout. A before period is another, and it is weak if you also changed ads. "Up to" means a ceiling observed somewhere, not your forecast. If the slide has no comparison group, the number is a decoration.

**"Engagement."** Time on page, tries per session, or add-to-cart. These are not interchangeable. A tool that traps someone in a dialog will raise time on page. Ask which event fires, and whether a failed generation counts as engagement. It should count as a failure.

**"Returns reduced."** Reduced against what, over what window, on which categories. A return rate that moves because the mix shifted to tees is not a try-on result. [How to measure a remarketing engine](https://antla.io/blog/measure-remarketing-engine-shopify) is strict about definitions for a neighboring system. Use the same strictness here. If they cannot define the event, they cannot tell you what you will be billed for or what you will report to finance.

A number you are not allowed to see the query for is not a result. It is a slogan with a percent sign.

## Red flags that are just taste

Not every annoyance is a tell. A founder who talks too fast, a UI you personally find plain, a loading time of several seconds on a heavy model, a sample that is not your aesthetic: note them, and do not pretend they are the logo test. Speed matters when you measured it. Taste matters when your customer shares it. Keep those in a different column from "they would not open our print."

Also do not fail a vendor for saying the preview does not know the shopper's measurements. That sentence is a green flag. A preview shows appearance. Anyone who says it proves a size, in a demo, is the red flag.

## End the demo or continue it

You continue when three things happened in the room: your hard SKU was generated, a bad upload produced an honest error, and every percentage on the slide has a denominator and a comparison group.

You stop when any of those was deferred to a follow-up deck. The follow-up deck will be designed overnight. You already learned what the live product will not show.

Send the notes the same day, with the images attached, to the person who signs. A memory of a beautiful dress will outvote your notes if you wait until next week's commercial call.

---

**About the author:** Aaron is the founder of Antla. He would rather a demo break on a logo than sail through on a solid tee.

## For agents

- Markdown: send `Accept: text/markdown` to this URL (and any other HTML page).
- OpenAPI: https://antla.io/openapi.json
- llms.txt: https://antla.io/llms.txt
- Sitemap: https://antla.io/sitemap-index.xml
- Docs: https://antla.io/docs
- CLI: npx antla info (npm package antla)
- Scope: antla.io is an informational marketing and docs site. The Shopify try-on backend is not on this origin. Install the app from https://apps.shopify.com/antla.
