What the October 4 test showed
A hands-on test of ChatGPT’s live Try On feature was published on October 4, 2026. The tester uploaded her own photo and tried five products in sequence: a moon pendant, an everyday outfit, a vintage wedding dress, costume wings and a cat collar. This is one user’s observation in practical scenarios, not a new feature launch or a laboratory model comparison. [1 · Business Insider · Hands-on test · October 4, 2026]
The result looked convincing in simple cases. The pendant retained recognisable proportions and appeared in the expected position, while the everyday outfit was reproduced with almost no conspicuous distortion. The collar also looked plausible on the cat, although such an image cannot by itself confirm size or comfort. [1 · Business Insider · Hands-on test · October 4, 2026]
Complex products produced material discrepancies. The system replaced the vintage dress’s smooth fabric and solid sleeves with lace, layers and sheer sleeves; it also changed the costume wings’ shape and visually lengthened the body. The tester therefore recommended relying on product measurements rather than a single generated image. [1 · Business Insider · Hands-on test · October 4, 2026]
A visual cue is not a fit measurement
OpenAI’s official description draws an important boundary: the user uploads a reference photo and ChatGPT creates a new image with the selected clothing or accessory. The company separately warns that the result may not represent the product or appearance exactly and does not guarantee fit or size. It advises checking measurements, product details and return policies before purchase. [2 · OpenAI Help Center · Shopping with ChatGPT Search]
ChatGPT Images 2.5 creates and edits images from instructions. That ability is useful for visual discovery and inspiration, but it is not the same as physically simulating fabric, pattern, tension or movement. The dress and wing discrepancies fit the risk of generative reconstruction: the system can plausibly add a detail that is absent from the sold product. [1 · Business Insider · Hands-on test · October 4, 2026] [3 · OpenAI · ChatGPT Images 2.5 · September 8, 2026]
OpenAI’s system card evaluates ChatGPT Images 2.5 for harmful-content safety categories, but publishes no tests of retail-product identity, garment fit or size accuracy. The fresh journalistic test is therefore useful as a signal about error types, but it cannot establish how often they occur. [1 · Business Insider · Hands-on test · October 4, 2026] [4 · OpenAI · ChatGPT Images 2.5 system card · September 8, 2026]
How retailers should test the value
For a retailer, try-on may shorten the path from inspiration to the product page by helping a shopper decide whether the overall look appeals to them. The technology should nonetheless be evaluated separately by category and product complexity. Jewellery, a basic shirt and a layered dress require different fidelity, and an aggregate usage rate can conceal the riskiest cases. [1 · Business Insider · Hands-on test · October 4, 2026] [2 · OpenAI Help Center · Shopping with ChatGPT Search]
A minimum metric set includes product-page visits, add-to-cart actions, completed purchases, cancellations, expectation-mismatch returns and support contacts. It also needs a manual sample checking colour, material, silhouette and distinctive details. More clicks without return controls could indicate a more persuasive error rather than a better decision. [1 · Business Insider · Hands-on test · October 4, 2026] [2 · OpenAI Help Center · Shopping with ChatGPT Search]
Reference photos can be saved for future try-ons and changed or deleted in settings. Retail interfaces should therefore explain clearly where an image is stored, how it is reused and how consent can be withdrawn; convenient repeat try-on should not turn a sensitive visual profile into an invisible condition of shopping. [2 · OpenAI Help Center · Shopping with ChatGPT Search]
Sources
- Business Insider · Hands-on test · October 4, 2026 — First-hand checks of jewellery, everyday clothing, a vintage dress, costume wings and a pet collar; the source for observations about successful and altered visualisations.
- OpenAI Help Center · Shopping with ChatGPT Search — Official description of Try On, saved reference photos and the warning that a visualisation does not guarantee exact product appearance, fit or size.
- OpenAI · ChatGPT Images 2.5 · September 8, 2026 — Official description of the image model, instruction-based editing and claimed improvements; it is not an independent assessment of retail fidelity.
- OpenAI · ChatGPT Images 2.5 system card · September 8, 2026 — Official safety evaluations; the published set does not test product reproduction, garment fit or size accuracy.
Expert commentary
The fresh test shows neither failure nor proven maturity, but the boundary of the current product. The system did well where the task resembled placing a recognisable object or ordinary outfit, then began changing the merchandise when texture and shape became unusual. For e-commerce, this distinction matters: a convincing image can help exploration while remaining too imprecise for a fit or sizing decision. [1 · Business Insider · Hands-on test · October 4, 2026] [2 · OpenAI Help Center · Shopping with ChatGPT Search]
The mechanism works through the shopper’s imagination. When it becomes easier to picture an item on oneself, less effort is needed to continue to the product page. Yet a generative model optimises a plausible image, not correspondence with every catalogue detail. Unless retailers separate inspiration from specification checks, expectations may rise faster than decision quality. [1 · Business Insider · Hands-on test · October 4, 2026] [3 · OpenAI · ChatGPT Images 2.5 · September 8, 2026]
The competitive advantage will come from controlled product fidelity, not the most dramatic picture. Platforms should keep the original product photo beside the result, identify potentially altered areas and provide direct access to size, material and return information. This preserves trust by showing where merchant data ends and probabilistic visualisation begins. [1 · Business Insider · Hands-on test · October 4, 2026] [2 · OpenAI Help Center · Shopping with ChatGPT Search]
Customer relationships also depend on how the reference photo is handled. OpenAI allows it to be saved for later try-ons and deleted in settings, but users still need clear explanations of storage and reuse. A body or face image feels more sensitive than ordinary search history; opaque consent can outweigh convenience and reduce adoption. [2 · OpenAI Help Center · Shopping with ChatGPT Search]
The evidence is limited to one journalistic experience. Success on simple products may have depended on the particular photos, angles and listing quality, while the dress error may reflect its unusual texture. The reverse is possible as well: other complex items could render better. Without a predefined sample and blind review, these observations cannot be generalised to a catalogue or returns impact. [1 · Business Insider · Hands-on test · October 4, 2026] [2 · OpenAI Help Center · Shopping with ChatGPT Search] [4 · OpenAI · ChatGPT Images 2.5 system card · September 8, 2026]
Over the next 60–90 days, useful measures are fidelity by category, manual-correction rate, time to product-page visit, conversion and expectation-mismatch returns. A positive scenario would combine more purchases with no deterioration in those returns or privacy complaints. Until such data exists, Try On should be positioned as a way to explore a look, not as a digital fitting room with measured accuracy. [1 · Business Insider · Hands-on test · October 4, 2026] [2 · OpenAI Help Center · Shopping with ChatGPT Search] [4 · OpenAI · ChatGPT Images 2.5 system card · September 8, 2026]