ChatGPT Adds Virtual Try-On for Clothes and Accessories: What Reference Photos, Saved Finds, Privacy Controls, and Buying Limits Actually Mean

Evidence checkpoints

Documented point: On 1 October 2026, ChatGPT introduced virtual try-on for clothes and accessories product listings, plus Favorites and folders in Library; OpenAI states mobile and web availability. The note says the Try on button will appear on clothes/accessories listings; it does not promise every account, merchant, product, country, or plan has it. [official source 1]

Documented point: The updated official guide documents selfie capture/upload, reuse/change/deletion of reference photos in Settings > Personalization > Reference photos, and product/merchant/result-selection rules. It expressly warns that try-on images may not represent the product or user appearance exactly and do not guarantee fit or size; prices, labels, reviews, and merchant rankings have stated limitations. [official source 1]

Documented point: OpenAI’s release-notes hub describes the 1 October ChatGPT shopping update as generally available and repeats the Try on, reference-photo, Favorites, folder and mobile/web details. Generally available status does not establish universal per-listing or per-region eligibility. [official source 1]

Documented point: OpenAI documents that ChatGPT Images supports creation/editing from uploaded images and is available on web, iOS, and Android across tiers. General ChatGPT Images availability is not proof that every account has the Shopping try-on entry point, nor a guarantee of output accuracy. [official source 1]

Documented point: The current consumer-data guide says submitted content can include images/files for model improvement depending on settings; it documents Data controls, Personalization, and Temporary Chat distinctions. It does not say deleting a reference photo erases all related chat/image records or that Temporary Chat automatically removes a saved reference photo. [official source 1]

Documented point: OpenAI documents that Improve the model for everyone can be turned off for new conversations, that it does not delete saved chats, and that Temporary Chats may be retained for up to 30 days for safety. Available controls depend on sign-in, plan, workspace, and other settings; training, history/retention and personalisation are separate controls. [official source 1]

News summary — 1 October 2026: OpenAI added a Try on entry point that can appear on clothes and accessories product listings in ChatGPT. When it is present, a shopper can capture or upload a selfie and use ChatGPT Images to generate an illustrative view of how an item might look on them. The same update added Favorites and folders in Library for saving products. OpenAI documents the shopping update for ChatGPT on mobile and web.

What changed

OpenAI’s ChatGPT release notes date this consumer shopping change to 1 October 2026. The announcement has two connected parts. First, a clothes or accessories product listing can show a Try on entry point. Selecting it lets the shopper take a selfie or upload one, after which ChatGPT Images creates a virtual try-on image. Second, a shopper can save a product to Favorites or organise products into folders in Library.

The generated image is an output of ChatGPT Images; Favorites and folders are organisational tools. Saving an item does not validate the image, reserve stock, fix a price or create an order. Conversely, generating an image does not require the shopper to treat the item as a saved buying candidate. A sensible decision rule is to separate exploration from purchasing: generate or save items only where the privacy implications are acceptable to the user, but do not progress towards payment until the merchant facts have been checked independently.

The official shopping guide provides more operational detail than the brief release announcement. It says a shopper can take or upload a selfie and can reuse a reference photo for later try-ons. It also documents a management path under Settings > Personalization > Reference photos, where the photo can be changed or deleted. This persistence can reduce repeated uploads, but it creates a separate housekeeping task: users should decide deliberately whether they want a reusable photo saved rather than assuming that every image-related control has the same effect.

A reference photo is the image used to guide later generated depictions of the person. It is not documented as a body scan, a set of measurements or a verified identity record. Nothing in the cited OpenAI sources says that the feature measures chest, waist, sleeve length or any other dimension. It should therefore not be used to infer a clothing size. The practical rule is simple: if the decision depends on measurements, use the merchant’s current size chart and measure against its stated method rather than interpreting the generated silhouette.

The workflow also sits inside ChatGPT’s broader shopping-results experience. Shopping results may not cover the whole market, identify the cheapest offer or guarantee that every supplied detail is current. A result can be non-advertising while still being incomplete, simplified or out of date.

For a historical comparison, Google Unveils AI Fashion Preview: Try on Clothes Virtually! describes Google Shopping’s 2023 generative-AI preview on diverse real-life models and its initially limited women’s-top rollout. It is category context only: its launch details, availability, and capabilities may have changed since publication.

What the generated view represents

The try-on output represents a model-generated visual possibility based on the supplied photo and product information available to the system. It can help a shopper consider a broad visual question, such as whether a short jacket shape works with an intended outfit. It cannot establish that the physical garment will fall in the same place, reproduce every seam, preserve exact texture or colour, or react to movement as shown. Those are material and manufacturing questions, not conclusions supported by an illustrative image.

OpenAI expressly warns that try-on images may not represent the product or the user’s appearance exactly and do not guarantee fit or size. That warning should govern how the image is read. If an output appears to show comfortable sleeve room, for example, the image is not evidence that the ordered size will have that room. If a colour looks warmer or darker than expected, neither the output nor a display is sufficient to settle the physical colour. Check the merchant’s photographs and written colour description, while recognising that those sources can also be imperfect.

Accessories require the same restraint. A generated view of a bag, scarf or other accessory can suggest visual proportion, but it does not authenticate the product, confirm exact dimensions or prove that straps, fastenings and pockets work as depicted. Where dimensions matter, compare the merchant’s stated measurements with an object of known size. Where authenticity matters, inspect seller identity and provenance using appropriate independent evidence. The OpenAI sources do not make ChatGPT an authenticity-checking service.

A practical, non-screenshot workflow

1. State the shopping intent. Begin with a bounded request that describes the category and constraints relevant to discovery, such as “Show me navy casual jackets with a stand collar.” Avoid putting secrets, financial details, identity documents or unrelated sensitive data into the prompt. A shopping description need not include a home address, payment details or private background information.

   Decision rule: include only information needed to find a product. If a fact is required only at checkout, provide it to the merchant through the merchant’s appropriate process, not in the ChatGPT shopping prompt.

2. Review the product result as a lead, not a complete market survey. OpenAI says product coverage is incomplete. Descriptions can simplify information from third parties, and the initially displayed price may not be the lowest available price. Read the result as a candidate that warrants investigation rather than as proof of best value.

   Decision rule: reject language such as “best”, “cheapest” or “complete selection” unless independently supported. A visible product card does not establish those claims.

3. Look for the Try on entry point. If it appears on the clothes or accessories listing, the item has an available route into the documented workflow. If it does not appear, do not infer an error or attempt to force the feature through assumptions about plan, region or merchant eligibility. Generic access to ChatGPT Images is not proof that the shopping listing qualifies.

   Decision rule: presence of the entry point permits an attempt; absence means continue without the listing-specific try-on or choose another candidate. It does not prove that the account lacks all image-generation capability.

4. Choose a selfie or an existing reference photo deliberately. The shopping guide documents capture, upload and later reuse. Select a clear image that serves the visual task while revealing no more than necessary. Check the background for documents, screens, badges, other people, addresses or possessions that should not be submitted.

   Decision rule: if the image contains sensitive or third-party information that cannot be removed, do not upload it. Crop or choose another photograph. Never include passwords, payment credentials, government identifiers or confidential workplace material.

5. Generate the visualisation. ChatGPT Images creates the output from the supplied material. OpenAI’s general Images documentation says ChatGPT can create images and edit uploaded images; that mechanism explains the visual output but does not turn it into a measurement process.

   Decision rule: describe the result as “an illustrative generated view”, not “the fit”. If the output changes body shape, garment details or proportions, treat that as an image limitation rather than product evidence.

6. Inspect for decision-relevant uncertainty. Compare visible features with the listing: overall length, collar type, fastening, pockets, pattern placement and accessories. Note discrepancies rather than asking the image to resolve them. The output can help form questions, but the product page must answer product-specific facts.

   Decision rule: if a feature affects function or purchase suitability, verify it in merchant specifications or with the merchant. Do not infer construction from the generated image.

7. Save the candidate if it remains useful. Add it to Favorites for a simple shortlist or place it in a Library folder when grouping products by purpose, recipient or comparison set. Folder membership is an organisational judgement, not a ranking or endorsement.

   Decision rule: save only enough candidates to compare meaningfully. Use folder names that record the decision context—for example, “Autumn jackets to verify”—rather than a conclusion such as “Correct size”.

8. Move to the merchant’s website carefully. Open the destination merchant and verify the exact product variant, seller and offer. Check the current price, currency, taxes, delivery cost, stock, materials, care instructions, dimensions or size chart, dispatch estimate and returns terms. Confirm that the product page corresponds to the colour and model considered in ChatGPT.

   Evidence boundary: the merchant handoff is where the generated visualisation and ChatGPT’s product summary stop being sufficient evidence for a purchase. Price, inventory, seller, fulfilment, return eligibility and transaction terms must be established at the destination and, where consequential, reviewed by a person before payment.

9. Compare independently before buying. OpenAI says prices may lag, the first price shown may not be the lowest, and product coverage is not exhaustive. If cost matters, compare like-for-like offers, including size, colour, seller, delivery and tax. A lower headline price for another variant is not necessarily a lower price for the item being considered.

   Decision rule: purchase only when the current merchant page supports the required facts. For money or other consequential decisions, require human review rather than relying on a generated image, label, summary or recommendation-like presentation.

The evidence boundary at a glance

Stage	Useful for	Does not establish	Next check
ChatGPT product result	Finding a possible clothes or accessories item	Complete market coverage, lowest price or current stock	Compare the exact offer with merchant information
Try on entry point	Starting the supported visual workflow when present	Universal account, region, merchant or item eligibility	Use only when shown on the listing
Reference photo	Guiding a reusable generated depiction	Body measurements, identity verification or size	Review the photo and its saved-reference settings
Generated image	Exploring a possible overall appearance	Exact product appearance, construction, drape, fit or size	Check specifications, photographs and size chart
Favorites or Library folder	Maintaining a shortlist	Reservation, price lock, endorsement or order	Reopen and re-check each candidate
Merchant destination	Checking the offer and transaction terms	That all claims are necessarily correct	Apply normal seller, payment and returns due diligence

For broader commerce context, ChatGPT Shopping Revolution: How the Agentic Commerce Protocol Changes Online Shopping explains the site’s account of visual discovery, comparison, and the Agentic Commerce Protocol. It is not a guide to the 1 October try-on release, and protocol names, capabilities, and product scope should be independently checked.

Where the documentation places it

The official sources place this release in consumer ChatGPT rather than presenting it as a new developer service. The 1 October 2026 ChatGPT release note names mobile and web, while the OpenAI product release-notes hub lists the update as generally available. The shopping guide then explains how the product-listing entry point, selfie and reusable reference-photo process operate. The general Images guide supplies context for image creation and editing from uploads.

These documents overlap but answer different questions. The release notes establish what changed and when. The shopping guide describes shopping-result behaviour, merchant selection and output limits. The Images guide establishes broad image functionality. Consumer-data and Data Controls guides explain relevant data choices, but they should not be used to invent a single master switch that simultaneously governs saved photos, chats, image records, personalisation and model improvement.

Fact versus inference

Issue	What the official sources support	Unsupported inference to avoid	Editorial and user decision rule
Eligible listings	The Try on entry point can appear on product listings for clothes and accessories.	Every garment, accessory, merchant or product result includes it.	Check the individual listing. Do not promise the control before it is visible.
Platform scope	OpenAI documents the October shopping update for ChatGPT on mobile and web.	Every operating system, desktop application, account configuration or managed workspace has identical access.	Name only mobile and web unless a newer official source establishes another surface.
Release-status wording	OpenAI’s release-notes hub labels the update generally available.	Every country, plan, account and listing must expose it simultaneously.	Treat the generally available label as the release status, not proof of per-user or per-item eligibility.
Generic Images availability	OpenAI documents ChatGPT Images across tiers and on web, iOS and Android, including creation and editing from uploaded images.	Any account able to create an image necessarily has the shopping Try on button.	Separate general image capability from listing-specific shopping access.
Reference-photo reuse	The shopping guide documents reusable reference photos and the path Settings > Personalization > Reference photos for changing or deleting them.	Deleting one automatically removes every associated chat, generated image or other record.	Manage each data category through its documented control; do not assume cascading deletion.
Universal access	The sources announce a consumer feature with stated categories and surfaces.	Access is guaranteed for every plan, region, age, merchant, workspace or product.	If the entry point is absent, report only that it is absent in that context; do not diagnose an undocumented eligibility rule.

This separation prevents a common category error: taking a visible capability in ChatGPT and assuming it is automatically exposed to developers or governed identically in a managed organisation. A consumer interface announcement can rely on internal product orchestration without creating a public API. Unless OpenAI publishes separate developer documentation, there is no source-grounded basis for implementation instructions.

What availability does not mean

“Available on mobile and web” is a platform statement, not a universal-access guarantee. It does not say that every account sees the feature, that every country is covered or that every product listing qualifies. It also does not establish parity between browser and mobile layouts, a desktop application, managed workspaces or any developer surface. Where documentation is silent, do not invent product or account details.

Likewise, the generally available label in OpenAI’s release-notes hub describes release status, but the shopping workflow remains conditional at the listing level. The practical sign of eligibility is the entry point appearing on a relevant listing. A user who can generate ordinary images but cannot see Try on on a given jacket has evidence only about that immediate context. They do not have enough evidence to conclude that their plan is excluded, their region is blocked or the merchant has opted out.

Availability also does not imply accuracy. ChatGPT Images being offered across tiers does not make a virtual garment depiction an exact representation. The shopping guide’s warning about appearance, fit and size applies at the point where confidence can otherwise become costly. A polished image can be visually plausible while still varying from the actual wearer, garment or combination of both.

Nor does a shopping result imply commercial completeness. OpenAI says product coverage is incomplete. Product descriptions may simplify third-party information, model-generated labels and review summaries are not guarantees, prices can lag, and the first displayed price might not be the lowest. OpenAI documents factors that affect merchant ordering, but the order does not certify the best seller. These limitations mean that a product can be useful to discover without being sufficiently verified to buy.

Advertising should remain a separate question. OpenAI’s shopping guide characterises product results under its own documented selection rules; advertising is a separate product surface. OpenAI’s advertising guide describes clearly labelled ads and test conditions involving particular plans and eligible regions. The presence or absence of advertising therefore should not be used to infer how a particular organic product result was selected, and ad-free status does not prove that organic shopping coverage is complete.

Finally, access to the workflow does not settle the privacy decision. A selfie or reference photo is user-submitted content. OpenAI’s consumer-data guidance says submitted content, including images and files, can be used for model improvement depending on settings. Its Data Controls guidance says Improve the model for everyone can be switched off for new conversations, but switching it off does not delete saved chats. Temporary Chat has separate behaviour, and OpenAI says Temporary Chats may be retained for up to 30 days for safety. These are distinct controls, not interchangeable promises.

Deleting a saved reference photo should therefore not be described as deleting associated chats or generated images, because the supplied sources do not establish that outcome. Disabling model improvement should not be described as deleting content. Temporary Chat should not be described as automatically removing a separately saved reference photo. Before uploading, users should review the controls available to their account and keep secrets, sensitive records and unnecessary third-party data out of prompts and images.

The governing buying principle is narrower but firm: visual plausibility is not purchasing evidence. A try-on image may help decide which jacket deserves closer attention; only independent checks of the exact product, variant, seller, dimensions, price, stock, delivery and returns can support the consequential decision to buy.

A generated visualisation, not a fitting-room measurement

ChatGPT’s virtual try-on output is an image generated from a product listing or item image and a selfie or other reference photo. It is not a body scan, a physical fitting, a measurement session or a size recommendation. The distinction matters because an image can help someone explore an overall look while leaving the practical questions that determine whether an item is wearable, genuine, available and returnable unanswered.

Explanatory example: suppose the generated view shows a cropped jacket apparently ending at the wearer’s waist, with sleeves reaching the wrists. That is an example of a visual composition, not evidence that the purchased jacket will have those proportions on that person. Sleeve length depends on the actual garment measurements, selected size, shoulder construction, the wearer’s measurements and the merchant’s sizing conventions. The image can illustrate “dark cropped jacket with these trousers”; it cannot establish “size 12 will fit with 61-centimetre sleeves”.

Decision rule: use a generated try-on to decide whether a design is worth investigating, not whether a particular size is safe to order. If a buying decision depends on a measurable property, obtain that property from the current merchant documentation and compare it with measurements taken outside the generated image.

Appearance is an interpretation, not a faithful preview

Appearance uncertainty begins with the two inputs. A product listing may provide only selected angles, edited catalogue photographs or simplified descriptive information. A reference photo may contain perspective distortion, loose clothing, occlusion, shadows or a camera angle that does not reveal body proportions accurately. ChatGPT Images then creates a new picture from those inputs; it does not place a physically measured garment onto a measured three-dimensional model of the user.

This means a plausible-looking output may still differ in silhouette, garment length, pattern position, fastening details or how much of the wearer is visible. Photographic polish is not evidence of precision. Indeed, a smooth, coherent result can make uncertainty less obvious because it may visually resolve details that the supplied evidence did not establish.

Practical procedure:

1. Identify the broad visual question before generating: for example, “Does this jacket’s shape work with wide-leg trousers?”
2. After generation, write down only the high-level observations the image supports, such as “the colour combination appears balanced in this rendering”.
3. List every purchase-critical detail separately: size, length, fibre content, fastening, stock, price and returns.
4. Verify those details on the current product page or in merchant documentation rather than reading them from the picture.

Trade-off: a generated visual can reduce the effort of imagining a combination, but its realism can encourage unwarranted confidence. The more consequential a detail is to the purchase, the less weight the rendered image should receive.

If you need general image-tool orientation, ChatGPT Images 2.0 Complete Tutorial: How to Create Stunning AI Images with DALL-E 3 in 2026 covers ChatGPT Images 2.0 generation, native editing, prompt construction, and API setup. Those creative-workflow instructions do not establish Shopping try-on eligibility, image accuracy, fit, sizing, or current availability.

What the visual can show and what it cannot establish

The generated visual can illustrate	It cannot establish	Evidence to use instead
An example of a colour or style combination	Exact colour under daylight, indoor lighting or a particular display	Current product photographs, colour description and merchant information, with the recognition that screens also vary
A possible overall silhouette	Actual fit, ease, tightness or mobility	Size guide, garment measurements, construction details and a physical fitting where possible
An approximate visual relationship between an item and a person	Body measurements or the correct labelled size	Measurements taken independently and compared with the merchant’s current guide
A stylised impression of fabric	Fibre content, weight, softness, stretch, opacity or breathability	Materials, care and product-construction information
An example of how an accessory might complement an outfit	Its exact dimensions, weight, capacity or attachment strength	Merchant specifications and measurements
A candidate item worth opening	Authenticity, seller authorisation or provenance	Seller identity, merchant terms and product documentation
A product discovered at one point in the shopping flow	Current stock in the required variant	Live merchant product page and checkout availability
A displayed or summarised price	The current or lowest total price	Merchant page and checkout total, including delivery and taxes where applicable
A visually appealing use case	Suitability for work rules, safety needs, allergies, disability requirements or another consequential context	Applicable requirements, professional advice where needed and human review

Decision rule: if the right-hand question would still matter after the image disappeared, it requires evidence outside the image.

Fit and size require measurements outside the rendering

“Fit” describes how a garment relates to a body in three dimensions, including room for movement, pressure, length and the intended amount of ease. “Size” is the merchant’s label or sizing category. Neither can be inferred reliably from how neatly an image generator depicts a garment.

A reference photo does not turn the workflow into measurement software. Even a full-length photograph can be affected by lens perspective, pose and distance. A generated picture can also adjust the garment or the apparent body to produce a coherent composition. It therefore cannot determine chest, waist, hip, inseam, rise, shoulder width or sleeve measurements.

Practical procedure for clothing:

1. Open the current merchant product page for the exact item and variant.
2. Find the size guide associated with that product or brand rather than assuming a familiar label is equivalent.
3. Check whether figures are body measurements or finished-garment measurements. They answer different questions.
4. Measure a similar garment that fits as desired, laid flat where appropriate, and compare like with like.
5. Check the product description for intended cut, such as fitted, relaxed, cropped or oversized.
6. If essential figures are absent or ambiguous, ask the merchant rather than using the generated image to fill the gap.

Example: a shopper comparing a medium jumper with one already owned should not compare their body chest measurement directly with an unexplained “chest width” figure. The latter may be a flat garment measurement. The buyer should first identify the merchant’s measurement convention, then compare the same type of measurement.

Decision rule: do not select a size from the try-on image. Select only after reconciling the merchant’s size guide, the item’s intended cut and relevant independent measurements. Where a poor fit could create a health, mobility, workplace or safety problem, require human review by an appropriate person and do not rely on ChatGPT’s output.

Construction and drape are physical properties

Two garments with a similar outline can behave very differently. Construction includes seam placement, lining, interlining, closures, pockets, panels, reinforcement and the way pieces are cut. Drape is how material hangs and moves under gravity. It depends on fibre, weave or knit, thickness, finish, stretch, cut and garment weight.

A generated image may suggest folds, volume or structure, but those visual cues are part of the synthesis. They do not prove that a coat has a functioning vent, that a dress is lined, that a pocket is usable or that a hem has the stiffness shown. A rendered lapel can look sharply structured even when the product uses softer material; conversely, a structured item may be depicted with flowing folds.

Practical procedure: inspect the merchant’s materials and care section, construction description and all available product photographs. Look specifically for lining, stretch, closure type, pocket function, seam position and care requirements. If those attributes determine whether the item meets the need, treat silence as missing information rather than a positive answer.

Worked example: an attractive visual of a pleated skirt may suggest that the pleats retain their shape while walking. Before purchase, the buyer still needs to check the fibre composition, whether the pleats are permanent or require pressing, whether the item is lined, and the care instructions. The image can motivate that investigation; it cannot answer it.

Trade-off: visualisation is useful for exploring proportion quickly, while construction research is slower but addresses durability, maintenance and physical behaviour. For an inexpensive styling experiment, a buyer may tolerate more uncertainty; for an item that is difficult to return or costly to maintain, missing construction details should carry greater weight.

Colour and texture remain uncertain

Colour can vary across listing photographs, generated output, display settings and lighting conditions. Texture presents a related problem: pixels can depict a surface as glossy, brushed, ribbed or smooth without establishing how it feels or performs. The official warning about imperfect representation applies to both the product and the user’s appearance.

Practical procedure:

• Read the named colour on the exact variant rather than selecting solely from the generated view.
• Compare the merchant’s available product photographs, including close-ups if supplied.
• Check fibre content and fabric description for evidence about texture and sheen.
• Confirm that the chosen colour swatch and the item in the basket are the same variant.
• If exact colour matching is essential, use a physical sample or an in-person check where available rather than treating any screen image as definitive.

Example: a generated bag may appear matte burgundy beside a navy coat. That can answer the exploratory question “Would a dark red accent work?” It cannot establish whether the merchant’s “wine” colour is brown-toned, purple-toned or glossy in daylight.

Decision rule: treat colour-family judgements as exploratory. If exact shade, surface finish or tactile quality is a purchase condition, require merchant evidence or physical inspection.

For optional visual-generation literacy, ChatGPT Images 2.0 Advanced Prompting: 25 Patterns That Get Production-Quality Outputs presents structured prompt specifications, negative constraints, and reference-image conditioning for production images. Treat it as a creative prompting resource, not evidence that Shopping try-on produces reliable fit, size, product results, or current feature access.

Advanced prompting guidance is optional background only. A more detailed instruction may change the composition, pose or requested emphasis, but it cannot convert the result into proof of fit, size, stock, price or feature availability.

Authenticity and seller identity are not visible properties

A picture resembling a branded item does not authenticate the item offered for sale. Authenticity concerns provenance: who manufactured or supplied the goods, who is selling them and what evidence accompanies the transaction. A generated try-on does not inspect a physical object, validate serial information or establish that a seller is authorised.

The distinction is especially important where product results may include more than one merchant. The item’s apparent design and the seller’s identity are separate facts. A convincing logo, clasp or pattern in a generated image cannot verify either.

Practical procedure:

1. Open the merchant destination rather than relying on the product card or generated image.
2. Identify the legal or trading seller named for the transaction.
3. Check whether the product is sold directly by the retailer or by a third-party marketplace seller.
4. Read the merchant’s product description and terms for statements about condition, provenance or authorised retailing where relevant.
5. Do not submit serial numbers, identity documents, payment details or other secrets to a prompt for an authenticity judgement.

Decision rule: if authenticity materially affects value or safety, a generated image contributes no verification. Require documentary evidence and human assessment appropriate to the purchase. For high-value transactions, obtain independent professional advice rather than asking ChatGPT to make the final financial decision.

Accessory scenario: an appealing picture leaves four questions open

Consider a shoulder bag shown on a qualifying accessories listing. The shopper supplies a reference photograph, and the generated image depicts the bag sitting neatly at hip height with a proportion that complements the outfit. This is a useful style illustration, but it leaves at least four independent uncertainties.

1. Scale: the image does not prove the bag’s dimensions, strap drop, capacity or position on the person. Perspective and synthesis can alter apparent size. The merchant’s measurements are required.
2. Construction: the picture cannot establish lining, pocket layout, zip quality, strap adjustability, hardware strength or material thickness. Product specifications and photographs are required.
3. Authenticity: visible branding in the rendering does not prove the offered item’s provenance or the seller’s status. Seller identity and transaction documentation are required.
4. Return eligibility: nothing about the generated appearance establishes whether the selected seller accepts returns, whether a sale item is excluded, what condition rules apply or when the return period begins. The current merchant policy for that transaction is required.

Reproducible buying procedure: record the exact variant, open the merchant page, compare the listed width, height and depth with an object of known dimensions, check the strap range, inspect the materials and closure details, identify the seller, then read the return terms before adding the item to the basket. At checkout, reconfirm seller, variant, availability and total cost.

Decision rule: if one of those four facts is missing and would change the purchase, pause. An attractive generated composition is not evidence that resolves the missing fact.

Inventory is a live merchant fact

A product appearing in ChatGPT does not establish that the desired size, colour or configuration remains in stock. Product coverage is not complete, and availability can change between discovery and checkout. The generated image is even further removed from inventory: it depicts a visual possibility rather than reserving a unit.

Practical procedure: verify availability on the current product page, select the exact variant, and check again in the basket or checkout flow. If timing matters, distinguish “listed”, “available for dispatch”, “pre-order” and any other merchant-supplied status rather than treating them as equivalent.

Example: a try-on may show a green shirt, while the merchant page has only blue in the required size. The generated green version does not create inventory and should not be used as evidence that the green variant can be purchased.

Decision rule: only the merchant’s current transaction flow can establish whether the selected variant is available to order. Even then, stock may change until the order is accepted under the merchant’s process.

Price displays can lag and may not be the lowest

OpenAI’s shopping guide states that displayed prices can be delayed and that the first price shown may not be the lowest available price. It also states that not all products are shown. These limits prevent three common inferences: a displayed figure is not necessarily current, the initially presented merchant is not necessarily cheapest, and the result set is not a complete market comparison.

Practical procedure:

1. Open the current product page for the exact variant.
2. Confirm the displayed currency and base price.
3. Check whether any discount has conditions, dates, membership requirements or variant exclusions stated by the merchant.
4. Add delivery charges and taxes where applicable.
5. Compare sellers only on the same product, condition, size or configuration and delivery destination.
6. Record the time of comparison if price is central to the decision, recognising that it may still change.

Trade-off: ChatGPT can narrow a broad search, but a wider merchant comparison takes additional time. Because product coverage is incomplete, the sensible depth of comparison depends on the purchase value and the buyer’s tolerance for missing alternatives. No shopping result establishes a lowest-market-price guarantee.

Decision rule: base a money decision on the merchant’s current total payable amount, not the generated image, a simplified result description or an earlier displayed price. Significant purchases require human review; do not delegate the final spending decision to ChatGPT.

Shipping, taxes and total cost are transaction-specific

A product’s headline price is not necessarily the amount the buyer will pay. Delivery destination, speed, taxes and merchant terms can change the total. The shopping visual contains no evidence about those charges, and a product-result price should not be assumed to include them.

Practical procedure: enter only the information required by the merchant through its legitimate checkout process, then inspect the item subtotal, shipping, taxes and any other disclosed charge before payment. Keep payment credentials, authentication codes and financial secrets out of ChatGPT prompts. If the total differs from the discovery-stage figure, use the checkout total as the immediate decision input and investigate any unexplained charge.

Example: two merchants may list the same accessory at different base prices. The lower base price need not produce the lower delivered total if shipping or taxes differ. Conversely, the initially displayed merchant need not be the cheapest because OpenAI warns that the first shown price may not be the lowest.

Decision rule: compare total delivered cost for the same variant and destination. If the checkout does not clearly explain the total, do not infer the explanation from ChatGPT.

Returns must be checked for the exact seller and item

Return eligibility is a contractual merchant matter, not a property that can be seen in a try-on. A general policy may also contain item-specific exclusions, condition requirements, deadlines or different rules for marketplace sellers. The relevant evidence is the policy applying to the exact transaction at the time of purchase.

Practical procedure:

• Identify the seller responsible for the sale.
• Find the return window and determine when it starts.
• Check exclusions for the category, sale status or product condition.
• Check who pays return postage and whether the original delivery charge is addressed.
• Record the policy supplied by the merchant with the order information.
• Ask the merchant to clarify ambiguity before purchase.

Decision rule: if trying the item depends on being able to return it, do not buy until return eligibility and likely return cost are clear from merchant evidence. Do not ask ChatGPT to make a legal determination about contractual rights; consequential legal questions require qualified human review.

Suitability depends on the intended use

An item can look appropriate in a generated scene while failing the requirements of its intended setting. Suitability may involve workplace rules, weather, visibility, protective performance, allergies, sensory needs, disability access, cultural context or care constraints. These are not reliably established by visual resemblance.

Practical procedure: turn the intended use into explicit requirements before considering the picture. For example: “must be machine washable”, “must not contain wool”, “must permit unrestricted shoulder movement” or “must meet an employer’s documented uniform rule”. Match each requirement to merchant documentation or an authoritative human source.

Decision rule: where clothing or an accessory affects health, safety, employment, government compliance or another consequential matter, require human review by the responsible person or qualified professional. A generated try-on may assist visual discussion but must not make the decision.

Which merchant evidence answers each buying question?

Buying question	Merchant evidence needed	What to verify
Is this the exact item and variant?	Current product page	Product name, model, colour, condition, configuration and selected variant
Which size should I consider?	Size guide	Measurement convention, units, body versus garment measurements and product-specific notes
What is it made from and how is it maintained?	Materials and care information	Fibre composition, lining, stretch, finish, washing or specialist-care requirements
Who will sell it to me?	Seller identity	Direct retailer versus marketplace seller, trading identity and relevant transaction terms
Can I order the required version now?	Availability	Exact size, colour, destination eligibility and dispatch or pre-order status
What will delivery add?	Shipping and taxes	Destination-specific delivery charge, estimated method and applicable taxes shown by the merchant
Can it be returned?	Return window	Deadline, start date, exclusions, required condition and return-postage responsibility
What will I actually pay?	Total cost	Item subtotal, discounts, shipping, taxes and other merchant-disclosed charges at checkout

Decision rule: a result is ready for purchase consideration only when each material buying question has a current evidence source. The image itself should not occupy any evidence cell in this table.

Shopping descriptions, labels and review summaries have limits

The Shopping with ChatGPT Search guide documents several limitations around product results. Product descriptions may simplify information obtained from third parties. Labels and review summaries are model-generated and are not guarantees or verified statements. Prices may be delayed, and an initially displayed price may not be the lowest. Product coverage is incomplete.

These disclosures affect how each element should be read:

• Simplified descriptions: use them for orientation, then consult the source product page for precise specifications. Simplification may omit a qualification that matters.
• Generated labels: treat a phrase applied to a product as a navigation aid, not as a certified attribute. Verify the underlying characteristic.
• Review summaries: do not treat a generated synthesis as a verified consensus or as proof that comments apply to the exact variant or seller. Consult the available underlying evidence and keep subjective reports separate from specifications.
• Delayed prices: recheck the merchant’s current page and checkout.
• Non-lowest initial prices: do not interpret ordering or prominence as a cheapest-price claim.
• Incomplete coverage: absence from the results does not establish that a product or merchant does not exist.

Example: a generated label such as “good for travel”, if shown, would not prove weight, durability, airline compliance or capacity. Those are separate claims requiring specifications and, where relevant, current carrier rules. The label may identify a line of enquiry but cannot finish it.

An unrun, reproducible documented-review protocol

Status: proposed and unrun. The following is an example method readers can perform themselves. It is not a benchmark, a product test, a scorecard or evidence of any result. It is designed to reveal which facts come from the generated image and which require independent documentation.

1. Choose one qualifying listing. Record the date, product name, visible merchant and exact variant. Do not assume every clothes or accessories listing will expose Try on.
2. Define three visual questions. Examples are overall colour pairing, approximate silhouette and whether an accessory style complements an outfit. Avoid measurement or authenticity questions because the output cannot establish them.
3. Define eight verification questions. Use the fields in the merchant-evidence table: product page, size guide, materials and care, seller identity, availability, shipping and taxes, return window and total cost.
4. Use a deliberate reference image. Exclude other people, documents, screens, location clues, financial information, identification numbers and sensitive material. Keep secrets and untrusted data out of prompts.
5. Generate once without repeated correction. Save the wording used and note that the output is an illustrative image-generation result. Do not interpret a more polished picture as a more accurate one.
6. Write observations before opening the merchant page. Restrict them to visible, non-verified impressions: for example, “the rendered combination uses a short jacket silhouette”. Do not write “the sleeves fit”.
7. Open the live merchant page. Record the exact variant, material, measurements, seller, stock status and current base price. Mark absent information as “not documented”, not as “no”.
8. Inspect transaction terms. Record delivery charges, taxes shown, return window, exclusions and total cost for the relevant destination without placing an order solely for this review.
9. Compare claims by category. Put visual impressions in one column and merchant facts in another. Do not assign accuracy percentages or scores; the method has not established a measurement standard.
10. Log unresolved issues. Examples include missing garment dimensions, ambiguous seller identity or unclear return exclusions. Each unresolved issue should lead to a pause or a question for the merchant.
11. Review privacy separately. If a reusable reference photo was saved, inspect the documented Settings > Personalization > Reference photos control. Do not assume that changing or deleting it also erases related chats or generated images; those are separate records and controls.
12. Stop before recommendation. The output of the protocol is an evidence sheet, not a verdict. The buyer reviews the evidence and decides whether further merchant or professional clarification is needed.

Trade-off: a single generation keeps the exercise manageable and limits the temptation to select the most flattering rendering. Multiple generations could expose visual variability, but they still would not establish which image, if any, matches the physical product. Neither approach substitutes for merchant verification.

Reference photos: start with the documented control

The documented management path is Settings > Personalization > Reference photos. OpenAI’s shopping guide, updated around the 1 October 2026 release, says a person can reuse a reference photo for later try-ons and can change or delete it from that location. This is a control for the saved reference photo. It should not be treated as a master deletion switch for every chat, uploaded image, generated image or saved product associated with the shopping workflow.

A careful procedure is therefore record-specific. Before uploading a selfie, open the reference-photo setting and establish what controls are visible on the account and device being used. After a try-on, return to the same setting to decide whether the saved photo should remain available for reuse. If it is no longer needed, use the documented reference-photo deletion control. Then review the relevant conversation and other visible image or chat records separately rather than assuming that the first action removed them too.

Decision rule: keep a reusable reference photo only when the convenience of later visualisations outweighs the preference not to retain that photo as a personalisation record. If the user would be uncomfortable finding the image saved there after the immediate task, the safer choice is not to create that persistent reference in the first place. Deleting it later is a useful control, but deletion after upload is not equivalent to never having submitted it.

Reuse, replacement and deletion apply to the reference-photo record

Reuse is the feature’s practical reason for saving a reference photo: a person need not provide a new selfie for every eligible clothes or accessories listing. That convenience also introduces persistence. A photograph selected for one jacket visualisation may remain available for a later accessory or clothing try-on until it is changed or deleted through the documented setting.

Changing a reference photo should be understood as choosing a different image for subsequent use, not as proof that every previous use has been rewritten. The supplied official sources do not state that replacing the saved reference photo retroactively alters generated images, removes an earlier upload from a conversation or deletes the conversation itself. Likewise, deleting the reference photo is documented as removing that reusable reference; the sources do not establish that it cascades into all related records.

For example, suppose a shopper takes a plain, front-facing photo for a coat visualisation, generates an image and later changes the reference to a newer photograph. The supportable conclusion is that the reference-photo setting has changed. It would be unsupported to conclude, without checking, that the first photograph has disappeared from any conversation in which it was uploaded or that the first generated visualisation has been removed. Those are distinct records with potentially distinct controls.

A practical clean-up sequence is:

1. Open Settings > Personalization > Reference photos and inspect the saved reference.
2. Change it if future try-ons should use another photograph, or delete it if reuse is no longer wanted.
3. Return to the conversation in which an image was submitted and determine whether the chat remains saved.
4. Review any generated image that remains visible in the relevant chat or image area.
5. Review saved products in Favorites or Library folders separately; those are product-saving features, not photo-deletion controls.
6. Check Data controls independently if the concern is whether new consumer conversations may be used to improve models.

This sequence deliberately avoids promising an all-in-one purge. Where the sources do not document a deletion equivalence, the correct status is unknown, not “automatically deleted”.

A records-and-controls matrix

The following matrix separates records that can appear in or around the workflow. “Separate check required” means the supplied OpenAI documentation does not justify treating another control as a substitute. “Unknown” means the cited sources do not establish the proposed deletion consequence.

Records, purposes and controls that must not be conflated

Record or setting	What it means here	Documented control or behaviour	What must not be assumed	Practical decision
Saved reference photo	A photograph retained for reuse in later virtual try-ons.	OpenAI documents changing or deleting it at Settings > Personalization > Reference photos.	Whether deleting it also erases uploaded chat images, generated outputs or associated chats is unknown.	Retain it only if repeated use is wanted; otherwise remove it through its own setting and inspect other records separately.
Uploaded chat image	An image submitted as content in a conversation, whether as a selfie, an item photograph or another input.	OpenAI’s Images guide documents creating or editing images from uploads. Consumer-data documentation says submitted content can include images and files.	Whether deleting the saved reference photo also deletes the uploaded copy in a chat is unknown.	Before uploading, assume the image is conversation content as well as visual input; afterwards, review that conversation itself.
Generated image	The illustrative ChatGPT Images output produced from the product context and photograph.	It is an image-generation result. The shopping guide warns that it may not represent the product or person exactly and does not guarantee fit or size.	Whether deleting the source reference automatically removes the generated image is unknown.	Inspect and manage the output as its own record. Do not retain or share it merely because the reference photograph was later deleted.
Saved chat	The conversation containing prompts, shopping context, uploads and outputs.	OpenAI’s Data controls guide states that turning off Improve the model for everyone does not delete saved chats.	Do not assume that disabling model improvement removes existing conversation history; OpenAI describes separate controls.	If the aim is to remove a saved conversation, deal with the conversation or history record rather than relying on the training preference.
Personalization or Memory	Features and settings that can influence how ChatGPT uses information across interactions. Memory is conceptually different from the saved reference-photo control.	OpenAI’s consumer-data guide describes Personalization and Data controls as separate settings areas; available options can depend on account and workspace conditions.	Whether deleting a reference photo clears every saved Memory or personalisation item, or changing Memory replaces that photo, is unknown.	Review visible Personalization or Memory controls independently when cross-conversation use is a concern.
Model-improvement choice	The consumer preference governing whether content from new conversations may be used to improve OpenAI’s models.	Improve the model for everyone can be turned off for new conversations. OpenAI says the change does not delete saved chats.	Do not assume that turning it off deletes old content, removes reference photos or prevents all retention; those outcomes are unsupported.	Disable it before starting a new conversation if the user does not want new consumer conversations used for model improvement, then make separate retention decisions.
Temporary Chat	A separate conversation mode with different history, Memory and model-improvement behaviour.	OpenAI says Temporary Chats may be retained for up to 30 days for safety.	Do not assume that “temporary” means instantaneous deletion, zero retention or automatic deletion of a separately saved reference photo; those outcomes are unsupported.	Use it only if its documented behaviour suits the task; do not use it as a substitute for checking the reference-photo setting.
Saved product in Favorites or a Library folder	A bookmark-like product record used to organise possible purchases.	The 1 October 2026 release notes document Favorites and folders in Library as shopping additions.	Whether removing a saved product also deletes photos, chats or generated images is unknown.	Treat product organisation as a separate convenience layer and remove stale finds independently.

The distinction between an uploaded chat image and a saved reference photo is especially important. The same photograph may participate in more than one role: it can be submitted as conversation content and also selected for future reuse. The existence of one control does not prove that only one copy or record exists. The official sources supplied for this article do not provide a universal record-level map connecting every upload, reference, output and chat.

Decision rule: identify the object first, then use the control named for that object. Use Reference photos for the reusable reference; conversation controls for a saved chat; Data controls for the model-improvement preference; Personalization or Memory controls for those concepts; and Favorites or Library for saved products. If a proposed action crosses those categories, regard the result as unverified until the interface or official documentation states otherwise.

Consumer content and model improvement are a separate question

OpenAI’s consumer-data guide says content submitted to consumer services can include images and files and may be used to improve models depending on the user’s settings. For this workflow, that means a selfie or other uploaded photograph is not outside the model-improvement discussion merely because it was supplied for shopping. The relevant choice must be reviewed in Data controls.

The control described by OpenAI is Improve the model for everyone. According to the Data controls guide, a consumer can turn it off for new conversations. The timing matters: it governs new conversations after the choice is made. It is not documented as a retrospective deletion tool, and OpenAI explicitly says turning it off does not delete saved chats.

A privacy-conscious procedure before starting is:

1. Sign in to the intended account rather than assuming another account’s settings carry over.
2. Open the available Data controls.
3. Review the status of Improve the model for everyone.
4. If model improvement using new consumer conversations is not acceptable, turn the setting off before creating the try-on conversation.
5. Return to Personalization and inspect Reference photos separately.
6. Decide whether a reusable reference is acceptable or whether the task should not proceed.

For example, a user may disable model improvement and still see an existing conversation in chat history. That is consistent with OpenAI’s documented separation: training preference and saved-chat retention are not the same control. Conversely, deleting a chat should not be described as proof that the model-improvement setting has changed. The user must inspect both controls when both concerns matter.

This separation prevents two common but unsupported conclusions. The first is, “I turned off model improvement, so my old chats and images were deleted.” OpenAI says saved chats are not deleted by that switch. The second is, “I deleted my reference photo, so no related content can be retained anywhere.” The cited sources do not establish that consequence.

Temporary Chat has different persistence, but “temporary” does not mean “never retained”

Temporary Chat is a distinct mode rather than a synonym for deleting a normal conversation later. OpenAI documents separate behaviour for chat history, Memory and model improvement. It also says Temporary Chats may be retained for up to 30 days for safety. A user should therefore not interpret the label as a promise of immediate erasure or zero server-side retention.

Nor should Temporary Chat be assumed to override the reference-photo system. The official material supplied here does not say that opening a Temporary Chat automatically removes an existing saved reference photo, prevents one from being selected, or deletes one after use. If the concern is reference-photo persistence, the user must still inspect Settings > Personalization > Reference photos.

A practical choice is to use Temporary Chat only after answering two separate questions:

• Conversation question: Is Temporary Chat’s documented treatment of history, Memory, model improvement and possible safety retention for up to 30 days suitable?
• Reference question: Is it acceptable for the selected photograph to exist as a reusable reference under the visible Personalization controls?

If the answer to either question is no, do not upload the photograph. A mode that addresses one form of persistence does not neutralise another undocumented one.

A privacy decision tree before supplying a photograph

1. Is a personal photograph necessary? The Try on route documented by OpenAI uses a captured or uploaded selfie, but using the feature is optional. If viewing the ordinary product listing and merchant photographs is sufficient, stop before uploading. The trade-off is less personalised visualisation in exchange for sharing less personal content.

2. Would disclosure of the image create significant harm or distress? Do not upload intimate images, identity documents, financial documents, medical material, workplace-confidential content or photographs exposing another person’s private information. Keep passwords, access tokens and other secrets out of prompts and images. If the photograph contains sensitive material that cannot be cropped out safely, do not use it.

3. Can a lower-sensitivity photograph do the job? Prefer a simple image containing only what is needed for the illustrative task. Check the background for letters, screens, badges, school or employer identifiers, house numbers, vehicle registrations, reflections, children and bystanders. Crop or take a new photograph rather than relying on the model to ignore those details.

4. Whose image is it? Use an image the account holder is entitled to provide. Do not upload another person’s face merely to see how clothing might look on them without their informed agreement. For children or other vulnerable people, do not treat the visualisation as a substitute for an adult’s privacy judgement or merchant sizing checks.

5. Have the current controls been reviewed? Inspect Data controls, Personalization and Reference photos on the account actually being used. Do not infer their state from a different device, another person’s screenshot or a previous session. Controls can depend on sign-in, plan, workspace and other settings.

6. Is model-improvement use acceptable? If not, turn off Improve the model for everyone before starting the new consumer conversation. Remember that this does not delete saved chats and is not a reference-photo removal tool.

7. Is reusable-photo persistence acceptable? If yes, proceed with awareness that the image can be reused and can later be changed or deleted in Reference photos. If no, do not assume Temporary Chat solves the issue; avoid the upload unless the visible workflow provides terms and controls that meet the user’s requirement.

8. Would Temporary Chat be preferable? Consider it for its documented history, Memory and model-improvement distinctions, while accepting that OpenAI says it may retain Temporary Chats for up to 30 days for safety. Do not describe it as anonymous or retention-free.

9. Can the output itself be handled safely? A generated image may still reveal appearance-related information. Decide before generation whether it may be saved, downloaded or shared. Do not post it publicly by default, and do not use it to make consequential judgements about another person.

10. Will the controls be revisited afterwards? Schedule an immediate review: inspect Reference photos, the conversation, generated images and saved products. Remove each unwanted record through its own available control rather than assuming that one deletion propagates everywhere.

Consider a shopper who wants to visualise a scarf. The least-data route may be to use ordinary product photography and skip the selfie. If a personal visualisation is still useful, the shopper could take a new image against a blank wall, with no other people or identifying material in view; review Data controls first; decide deliberately whether reuse is wanted; and inspect the reference-photo setting and chat afterwards. This is an example of data minimisation, not a guarantee that the service creates no other records.

Review the controls again after the task

Privacy decisions should not end at upload. The workflow can leave several useful but distinct artefacts: a reusable photograph, a saved conversation, a generated visualisation and a product in Favorites or a Library folder. Each may have a different reason for remaining. A reference photo may be convenient for another garment, while a generated image may no longer be useful; a product bookmark may be worth keeping even when the personal image is not.

Use this post-task procedure:

1. At Settings > Personalization > Reference photos, confirm which photograph is saved.
2. Delete or change that photograph if continued reuse is unwanted.
3. Open the relevant chat and review the prompt text, uploaded content and generated image that remain visible.
4. Review Favorites and Library folders, removing products that are no longer under consideration.
5. Recheck Improve the model for everyone rather than assuming a previous choice persisted across accounts or workspace contexts.
6. If Temporary Chat was used, remember the documented possibility of safety retention for up to 30 days.

Decision rule: preserve only records with a continuing purpose. Convenience alone may justify keeping a product bookmark, but it does not automatically justify retaining a personal photograph. Conversely, deleting the photograph should not require discarding useful merchant information if the product can be retained without it.

Managed workspaces require a separate policy check

OpenAI distinguishes consumer ChatGPT services from managed offerings. Organisation-controlled workspaces can have different default model-training treatment and administrative controls. That distinction does not establish that virtual try-on is available in Business, Enterprise or Education workspaces, nor does it establish that the consumer reference-photo path appears there in the same form.

If using an employer, school or other managed account, first ask whether personal shopping and personal-photo uploads are permitted under the organisation’s rules. Then inspect the controls that are actually available and consult the workspace administrator where necessary. Do not move a personal photograph into a managed workspace merely because its data treatment may differ from a consumer account; organisational access, acceptable-use requirements and retention arrangements may create different considerations.

The reverse inference is also invalid: consumer availability on mobile and web does not prove workspace availability. The 1 October 2026 release notes provide a mobile-and-web statement for the ChatGPT shopping update, while the supplied sources do not promise try-on for every managed workspace, plan, region, account or listing.

None of the OpenAI sources cited in this article announce a public API virtual-try-on endpoint or a Codex capability. General ChatGPT Images support for creating and editing from uploads is not evidence that developers can reproduce the shopping workflow through an API, or that workplace deployment controls are identical to consumer settings.

Organic product results and advertisements are separate systems

OpenAI’s shopping guide says product results follow a documented selection process, while separately labelled advertisements are described in another OpenAI guide. This describes the shopping product-result process. It does not mean the market is comprehensively covered, that the first item is objectively best, or that every merchant participates. The same guide documents limitations in product coverage, price freshness, descriptions, labels, review summaries and merchant ordering.

Advertisements are treated separately in OpenAI’s Ads in ChatGPT guide. The guide says ads are clearly labelled and do not shape answers. As documented for the current test, ads may appear to Free and Go users in eligible testing regions, while Plus, Pro, Business, Enterprise and Education plans do not have ads. These are advertising eligibility conditions, not rules for organic product ranking.

How to distinguish shopping results from ads

Question	Organic shopping product result	Advertisement
How does OpenAI characterise it?	The shopping guide documents product-result selection; a separate guide covers labelled advertising.	A separate advertising placement.
How should the reader identify it?	Do not relabel an ordinary shopping result as sponsored without evidence.	Look for the clear advertising label documented by OpenAI.
Does plan status establish different organic ranking?	No such distinction is established by the assigned sources.	Plan and eligible-region conditions determine exposure to the documented ad test.
Does absence of an ad prove complete market coverage?	No. Product coverage can still be incomplete.	No. Whether an ad appears says nothing about completeness of organic results.
What should happen before purchase?	Verify the product, seller, current price, stock, size information, delivery and returns at the merchant.	Apply the same merchant verification; a label identifies a paid placement, not product suitability.

A useful reading procedure is to inspect the result before interpreting it. First, determine whether it is presented as an ordinary product result or carries an advertising label. Second, do not infer that an organic result is exhaustive or that an advertisement is an endorsement. Third, open the merchant only after recording which item, seller and variant are under consideration. Finally, verify the transaction details at the destination.

For example, a Free or Go user in an eligible test region might encounter a separately labelled ad alongside the broader ChatGPT experience. That does not convert every clothing result into an ad. A Plus user’s documented lack of ads does not establish a superior organic ranking, broader product coverage or access to more try-on listings. The cited sources support an ad-eligibility distinction, not a claim of plan-based organic shopping quality.

Decision rule: use labels to distinguish paid placement from product results, but apply independent merchant checks to both. “Not an ad” answers a funding and selection question; it does not answer whether the item is authentic, in stock, correctly described, suitable, returnable or offered at the lowest total price.

Three responsible ways to use the try-on workflow

The try-on workflow can help with visual comparison, but it does not turn a listing into a verified product, a photograph into body measurements, or an image into evidence of fit. The following scenarios are procedures readers can adapt; they are examples, not tests of the feature or guarantees that any particular account, listing or output will qualify.

Scenario one: considering a jacket

1. Qualify the listing before supplying a photograph. Begin with the product result itself. Confirm that it is a clothes listing and look for the documented Try on entry point. Do not assume that every jacket result is eligible merely because jackets fall within the general clothing category. The release notes say the button will appear on clothes and accessories listings, but they do not establish universal coverage across every item, merchant, account, plan or region.

If the button is absent, stop treating that result as a supported try-on listing. You may still read the displayed product information or visit the merchant, but a generic request to create an outfit image is not equivalent to using the shopping listing’s documented try-on path. Decision rule: proceed with the listing-based workflow only when the listing itself presents the entry point; otherwise use the merchant’s ordinary product information and do not infer eligibility.

2. Decide whether a photo is justified. A reference photo is useful only if seeing an approximate jacket-and-person composition will materially help the comparison. Before capture or upload, inspect the frame for unrelated people, identity documents, payment cards, house numbers, workplace badges, computer screens, private correspondence and location clues. Crop those elements out rather than expecting the generated image or a privacy control to neutralise them.

A plain, recent image in which the relevant area is visible may make the intended visualisation easier to interpret, but this is a suggested method rather than an accuracy guarantee. Do not upload intimate imagery, confidential material or a photograph you lack permission to use. Do not put passwords, financial details, government identifiers, medical records, employment records or other secrets in the prompt. If another person is the subject, obtain their informed agreement before using their photograph; if that is not practical, decline the upload.

The shopping guide says reference photos can be reused and managed through Settings > Personalization > Reference photos. Reuse is convenient, but it also means the photo should be chosen as a saved reference rather than treated as a disposable input. Trade-off: a reusable image reduces repeated uploads, while an image you are unwilling to retain should not be made a reusable reference merely for convenience.

3. Interpret the visualisation narrowly. Suppose the generated example appears to place the jacket over the user’s clothing. The defensible interpretation is: “This provides an illustrative composition that may help me consider overall styling or silhouette.” It is not: “This proves the shoulders fit, the sleeves are the correct length, the fabric hangs this way, or this is my size.” OpenAI’s shopping guidance expressly warns that try-on images may not represent either the product or the user exactly and do not guarantee fit or size.

Look for issues that make the picture unsuitable even as a visual aid: altered fastenings, missing pockets, changed lapels, inconsistent hems, implausible layering, distorted hands or accessories, or a body shape that does not correspond closely enough to the reference. These are inspection categories, not a claim that every output will contain such errors. If a construction detail matters to the purchase, the merchant’s photographs and specification should control; if the generated rendering conflicts with them, disregard the generated detail.

4. Save only if the candidate remains worth checking. The October 2026 release notes add Favorites and folders in Library for saved products. A jacket may be placed in Favorites for a short shortlist, or in an example folder such as “Jackets to verify” if the interface offers the documented Library-folder workflow. Saving is an organisational action, not an endorsement, stock reservation or price lock.

Before saving, decide whether the benefit of retrieval outweighs the persistence of the shopping record. If the candidate is already rejected on material, style or budget grounds, do not retain it merely because the visualisation was appealing. Decision rule: save only an unresolved candidate that you intend to compare or verify; otherwise remove it from the working shortlist.

5. Name the unresolved uncertainty. For a jacket, the outstanding questions commonly concern chest and shoulder measurements, sleeve length, garment dimensions, lining, fabric composition, fastening construction, care requirements, colour, stock, seller identity, delivery and returns. The image answers none of those conclusively. Write the uncertainties down before opening the merchant page so that visual appeal does not displace the buying criteria.

6. Verify independently with the merchant. At the merchant’s current page, match the exact product and variant. Check the size chart and whether it gives body measurements or finished-garment measurements; these are not interchangeable. Read the material and care description, inspect merchant imagery, confirm the selected colour and size are in stock, identify the seller, and check the current item price, delivery charge, taxes where shown, dispatch estimate and return conditions. A marketplace may have seller-specific terms, so verify the actual seller attached to the chosen variant.

OpenAI’s shopping guide warns that displayed prices can lag and that the first price shown may not be the lowest. It also says product coverage is incomplete. Therefore, a price shown in ChatGPT is a lead to verify, not a payable quotation or proof of market-wide value. Purchase rule: pay only on the basis of the merchant’s current checkout information and policies, not the generated image or an earlier ChatGPT price display.

Scenario two: comparing accessories

1. Check that each candidate qualifies separately. Imagine comparing two accessories returned by a shopping query. One result may present Try on while another does not. Eligibility is attached to the available listing workflow, not transferred by category similarity. Do not assume that because one accessory can be visualised, every comparable product can be processed through the same listing button.

Record the asymmetry in the comparison. Candidate A may have a generated visualisation, while Candidate B may have only merchant photographs and listing details. That does not make A better, more authentic or more strongly recommended. It means the available evidence differs. Decision rule: compare products on verified buying criteria, and treat visualisation availability as an optional aid rather than a ranking signal.

2. Choose the least revealing adequate photograph. For an accessory worn near the face or upper body, the user may decide that a closer crop is sufficient. A narrower frame can reduce unrelated background information, although it does not eliminate the fact that a recognisable image may be personal content. For an item where a useful view would require an image the user does not want retained or processed, the responsible choice is to decline try-on and use product dimensions or merchant imagery instead.

If a reusable reference already exists, review whether it remains appropriate before selecting it. A convenient saved photo may be old, poorly framed or more revealing than necessary. The documented path for changing or deleting reference photos is Settings > Personalization > Reference photos. Deleting that reference-photo record should not be described as deleting associated conversations, generated images or every other stored record: the official sources treat these controls separately and do not promise such cascading deletion.

3. Inspect differences without converting them into measurements. An example comparison might ask whether one shape appears more visually prominent than another. The generated images may support that limited aesthetic judgement if both are coherent enough to inspect. They cannot establish exact scale, weight, comfort, clasp reliability, skin response, optical performance, metal composition or durability. Nor can an image authenticate a branded item or identify whether a marketplace seller is authorised.

Use a two-column note rather than relying on visual memory. Under “illustrative impression”, record matters such as relative prominence or compatibility with an outfit. Under “requires evidence”, record dimensions, materials, weight, included components, warranty, seller and returns. This separates interpretation from factual verification. It also prevents a more polished output from being mistaken for a more suitable product.

4. Save comparable records deliberately. If both items remain candidates, save them to the same relevant Library folder where the documented feature is available, or use Favorites for a smaller general shortlist. Use neutral organisation: for example, “Accessories to compare”, not “Best choice”. Folder names are the user’s labels and do not make the contents verified recommendations.

If only one item can be saved or visualised, preserve the missing-data note. Do not fill the gap with a fabricated description or assume that ChatGPT’s product set covers the entire market. OpenAI’s shopping guide says not all products are shown, and its explanation of merchant and product selection does not amount to a complete-market guarantee.

5. Carry unresolved questions to the source of sale. Verify the precise dimensions and materials, the exact colour or finish, included parts, stock status, seller identity, delivery terms and return restrictions on the merchant page. For an accessory whose hygiene seal, piercing status, personalisation or final-sale classification could affect returns, read the item-specific policy rather than relying on a general returns banner. These are examples of merchant checks, not assertions about a particular product.

Product labels and review summaries shown by ChatGPT are also subject to the shopping guide’s limitations: they are model-generated and are not guarantees or verified statements. Use reviews to identify questions worth investigating, not to establish material composition, authenticity or universal comfort. A description may simplify third-party information, so compare critical wording with the merchant’s current specification.

6. Apply a fair comparison rule. Proceed to merchant verification when at least one candidate satisfies the intended use on known facts and the unknowns are answerable before purchase. Pause when the visualisations conflict with product photographs or important dimensions are absent. Decline when the only reason to prefer an item is that its generated image looks more convincing. For expensive or otherwise consequential purchases, have a person review the evidence, total cost and terms before payment.

Scenario three: buying a gift

1. Separate product eligibility from permission to use a person’s image. A clothing or accessory listing may qualify for Try on, but that does not grant permission to upload the intended recipient’s photograph. A gift does not create an exception to privacy. If the recipient has not agreed to the use of their image, do not upload it or make it a reusable reference photo.

An alternative is to inspect the product without person-specific visualisation, use merchant photographs, or ask the recipient for measurements and preferences without revealing the gift. Using the buyer’s own image may offer a rough styling reference, but it cannot establish how the item will look or fit another person. Decision rule: no informed permission, no recipient-photo upload.

2. Qualify the listing without overstating access. Confirm that the particular gift listing displays Try on. If it does not, do not infer that changing devices, repeatedly regenerating or moving to an unrelated ChatGPT surface will necessarily enable it. General ChatGPT Images availability is not proof that the Shopping try-on entry point exists for that listing or account.

3. Limit what a consented visualisation means. Where the recipient has agreed and a suitable photo is used, the image can be considered as an illustrative presentation of style. It remains unable to guarantee size, fit, colour accuracy, texture, drape, comfort or the recipient’s preference. A pleasant-looking picture is especially weak evidence for a gift because the buyer may not know the recipient’s tolerance for particular fabrics, fastenings, proportions or care requirements.

Do not use the output to infer sensitive characteristics, health information, employability, social status or other consequential judgements about the recipient. Virtual try-on is a shopping visualisation, not a basis for medical, employment, insurance, credit, government or other high-impact decisions. Any use touching those areas requires qualified human review and an appropriate evidence process outside this feature.

4. Save only what can be managed discreetly and responsibly. Favorites or a Library folder can keep gift candidates together, but saving may leave an accessible record in the account. Consider whether the account is shared or visible on a shared device. The official sources document saved products and reference-photo controls, but they do not guarantee that a saved shortlist will remain hidden from another person who can access the same account or device.

If surprise matters more than convenience, do not save the candidate in ChatGPT. If the recipient later withdraws consent to use their reference photo, review Settings > Personalization > Reference photos and separately consider relevant chats and generated images. Do not assume that changing one control automatically changes the others.

5. Identify the gift-specific uncertainties. In addition to size and material, check the exchange process, gift receipts where the merchant offers them, return deadline, personalised-item restrictions, delivery timing and whether the recipient can return the product without access to the buyer’s payment account. ChatGPT cannot guarantee any of those terms from a generated visualisation.

6. Verify at the merchant and preserve human judgement. Match the exact variant, inspect the current size guide, and confirm availability, seller, delivery and returns. Review the final amount at checkout. If the purchase would create financial strain, involve a large sum, or rely on credit, pause for human review rather than treating a shopping image as justification. ChatGPT should not make the financial decision.

Access problems and responsible troubleshooting

When the Try on button is missing

A missing button has several possible explanations, but the official sources cited here do not provide a diagnostic code for distinguishing among them. The listing may not qualify, the account or region may not currently expose the workflow, or the documented product experience may have changed. OpenAI’s 1 October 2026 release notes state that Try on is appearing for clothes and accessories listings on mobile and web; they do not guarantee that every eligible-looking result will show it.

1. Confirm that you are looking at a ChatGPT shopping product listing for clothes or accessories, rather than ordinary text, a generic uploaded image or a merchant page.
2. Check the listing itself for the entry point. Do not rely on another result having displayed it.
3. Consult the current ChatGPT release notes and shopping guide, because availability statements and workflows can change after the October 2026 announcement.
4. If the button remains absent, continue without virtual try-on. Visit the merchant for specifications or compare another qualifying listing rather than attempting to force access.

Do not interpret absence as proof of an account fault, and do not interpret presence on mobile as proof that an identical control must appear in every other surface. The cited sources document mobile and web; they do not announce a desktop-application entitlement, Codex capability or public API endpoint for virtual try-on.

When the product does not qualify

A nonqualifying product may still be discussed in a chat or represented in an ordinary image-generation request, but that is a different task from invoking Try on on a shopping listing. Keep that distinction visible. The shopping workflow carries product-result context and its own disclosures; a free-form generated composition should not be relabelled as an official listing try-on.

The practical response is to use merchant photographs, dimensions and specifications, or choose another listing that presents the documented control. Do not modify the product category in the prompt merely to make an ineligible item appear eligible. Decision rule: if the entry point is absent, treat virtual try-on for that listing as unavailable and do not claim otherwise.

When account or regional availability is uncertain

The OpenAI release-notes hub labels the October 2026 shopping update as generally available and repeats the mobile-and-web statement. “Generally available” does not establish universal per-listing, per-account, per-plan or per-region access. The correct troubleshooting step is to compare the current official wording with what the signed-in product actually presents, not to infer an entitlement from the general label.

Account conditions can also matter to privacy controls. OpenAI’s consumer-data documentation distinguishes Data Controls, Personalization, Temporary Chat and managed-workspace arrangements. Available controls can depend on sign-in status, plan, workspace and other settings. If the account belongs to an employer, school or other organisation, consult that organisation’s policy and administrator before uploading a personal image. Do not assume consumer documentation overrides workspace rules.

Advertisements require another distinction. OpenAI’s ads guide says ads are separately labelled and, as described for its test, may appear to Free and Go users in eligible regions, while the named paid and managed plans do not have ads. That test does not define try-on eligibility. Organic shopping product results are documented separately from advertisements, so changing an advertising condition should not be presented as a way to enable or alter organic try-on results.

When image generation is delayed or fails

Image generation may not complete immediately; the assigned OpenAI guide does not specify a completion time. A delay is therefore not, by itself, evidence that the product is unavailable or that the reference photo is unsuitable. Avoid repeatedly uploading the same photograph while a request may still be processing, because duplicate submissions can create extra records and confusion without resolving the underlying issue.

1. Wait for the current request rather than immediately starting several copies.
2. If it fails, check whether the product listing and Try on entry point are still present.
3. Review the uploaded image for obvious usability or privacy problems and replace it only if you deliberately choose to do so.
4. Retry once when appropriate, or abandon the visualisation and proceed with merchant evidence.
5. Consult the current official Images, Shopping and release-note documentation if behaviour differs from the October 2026 description.

Do not claim that a retry will succeed, estimate an undocumented completion time, or interpret a failed generation as a product-quality signal. It says nothing reliable about the garment, accessory or merchant. If the task is time-sensitive, the safer option is to bypass generation and verify the item directly.

When the output is visibly inconsistent

Failure is not limited to a technical error message. An image that completes but materially changes the item or person may be unusable for comparison. Inspect the generated view against the listing and reference photo. If a defining feature is missing, duplicated, repositioned or altered, do not mentally correct it and continue as though the output were accurate.

Try another generation only if the privacy and persistence decision remains acceptable. Otherwise, discard the visual inference and use source photographs. Decision rule: when the image conflicts with merchant evidence, merchant evidence governs product facts; when merchant evidence is itself unclear, pause rather than choosing whichever representation is more attractive.

When product information may be stale

ChatGPT’s shopping guide says prices may be delayed and an initially displayed price may not be the lowest. Product descriptions may simplify third-party information, while labels and review summaries are model-generated rather than guaranteed or independently verified statements. Stock, delivery, seller and returns can also change at the transaction destination.

Refresh the decision at the merchant, not merely the ChatGPT conversation. Match the product name or identifying details, exact variant, seller and quantity; then check the current price and final checkout total. If a key attribute differs between ChatGPT and the merchant, treat the ChatGPT detail as stale or unresolved. Do not ask the generated try-on image to arbitrate a textual discrepancy.

Incomplete coverage is also relevant. The absence of another product, seller or lower price from the result set does not establish that none exists. Search elsewhere when market comparison matters. For a high-cost purchase, make the comparison process and final terms subject to human review.

When official guidance may have changed

This article’s volatile product facts are attributed to OpenAI’s documentation as of the 1 October 2026 release and its associated shopping update. Product interfaces, availability and controls can change. Before relying on a navigation path or availability statement, consult the current official Help Center and release notes listed at the end of this article.

A documented path should be followed as written rather than supplemented with invented support menus. For reference photos, the supplied shopping guide names Settings > Personalization > Reference photos. For model-improvement choices, OpenAI documents Data Controls. These are separate areas serving separate purposes. If the current product no longer matches the documentation, pause and use the current first-party guidance rather than guessing at a replacement control.

Proceed, pause or stop

Action	Conditions	Next step
Proceed	The listing displays Try on; the photo is yours or used with permission; no secrets or unnecessary private details are present; reference-photo persistence is acceptable; and the image is needed only for limited visual comparison.	Generate the illustrative view, inspect it for inconsistencies, save the product only if it remains a genuine candidate, then verify every buying fact with the merchant.
Pause	Eligibility or regional access is unclear; the account is managed; the photo contains avoidable private information; the output conflicts with the listing; product data may be stale; or price, size, seller and return terms remain unresolved.	Consult current official guidance, reduce or replace the input where appropriate, involve the account administrator or another human reviewer, and obtain current merchant evidence.
Do not proceed	The listing lacks the entry point; image permission is absent; the task requires intimate or confidential content; the user expects measurement, guaranteed fit, authentication or a consequential judgement; or acceptable merchant evidence cannot be obtained.	Do not upload or purchase on the basis of the visualisation. Use non-personal product information, choose another verification route, or abandon the transaction.

ChatGPT’s virtual try-on can be useful as a limited visualisation tool when the listing qualifies, the user accepts the persistence implications of reference photos and saved records, and the generated image’s uncertainty is kept explicit. It is not a fitting, measurement, authenticity check or purchasing guarantee. The responsible endpoint is always an independently checked merchant record for the exact item, variant, seller, price and terms, followed by human judgement before payment.

Useful Links

• ChatGPT release notes
• Shopping with ChatGPT Search
• OpenAI product release notes
• Images in ChatGPT
• How OpenAI handles data in consumer services
• Data Controls in ChatGPT
• Ads in ChatGPT