18/11/2026
ChatGPT Shopping is the feature that makes products appear in ChatGPT's answers: the user describes what they are looking for, and the assistant offers a selection with images, prices and links to merchants. This page is for e-commerce and marketing leads who want to know how their products get retained there and what needs to be prepared.
It deals with one specific entry point into agentic commerce, for which our reference page sets out the framework: mandate, delegated payment and liability. Here there is a single subject: being selected and recommended in ChatGPT.
What ChatGPT Shopping is
ChatGPT Shopping refers to the set of purchasing features built into the assistant: product search in natural language, product cards compared within the conversation, links to merchants and, depending on the country, payment without leaving the interface. The scope is changing quickly and varies from one market to another: what follows describes how it works at the time of writing.
Product search in natural language
On a retail site or a comparison engine, the buyer filters: category, price, brand, size. In ChatGPT, they describe a situation: a use, a budget, constraints, sometimes a photo. The assistant translates that request into criteria, looks for products that meet them and presents a short selection, each product accompanied by an image, a price, a summary of reviews and an explanation of the choice.
The difference lies in that explanation. ChatGPT does not simply list: it says why a given product matches the request, and what sets it apart from the others. The buyer can then follow up ("cheaper", "delivered before Friday", "compatible with my equipment"), and the selection is rebuilt. Your product is therefore judged several times in the course of a single conversation, on criteria that become more precise.
Recommendations that are not adverts
OpenAI presents these results as recommendations selected by the assistant, not as advertising placements: you do not buy a place in the selection. Products are retained on the basis of their data, how well they match the request and the information available about them on the web.
This has a direct consequence for a merchant. The lever is not a bidding budget, but the quality and availability of information: a product whose price, stock or specifications cannot be read or verified does not enter the selection, whatever its real quality.
Why the subject already weighs on your sales
The volume of this channel remains below that of classic search, but it is growing fast and its traffic stands out for its quality. Three benchmarks put it in perspective, each on its own scope:
These figures cannot be combined with one another: they come from different sources and scopes. They do, however, support the same decision: treat the presence of your products in ChatGPT's answers as a channel in its own right, with its data, its rules and its measurement. Other benchmarks on behaviour linked to AI engines appear in our GEO statistics.
A visitor who arrives from a recommendation does not behave like one from a results page. They have already compared, read a summary of reviews and retained two or three options. Your product page no longer has to convince them of the product's value: it has to confirm what the assistant announced. A different price, an item out of stock or a missing specification breaks that confirmation, and the sale is lost at the last step.
How ChatGPT selects products
Selection follows a logic that agentic commerce makes familiar: the assistant does not guess, it reads data and compares it with a request.
The data it reads
ChatGPT relies on several sources to describe a product and decide whether to offer it:
- Product feeds supplied by merchants: titles, descriptions, prices, availability, variants, images, identifiers.
- Structured data on product pages, read on the site: price, stock, reviews, specifications.
- Page content: descriptions, guides, compatibility tables, delivery and returns conditions.
- Third-party sources: reviews, tests, comparisons and discussions, which feed the summary displayed next to the product.
The first three depend on you. The fourth is built over time: a product nobody talks about anywhere other than on your site offers little material for a confident recommendation.
The criteria that separate two products
For the same request, four criteria make the difference between a product that is offered and one that is set aside:
- Match with the request: the product's attributes explicitly answer the constraints expressed (dimension, use, compatibility, budget).
- Verifiability: final price, availability and lead time are readable and up to date.
- Richness of information: complete specifications, images, variants, returns conditions.
- Reputation: numerous, recent reviews that are consistent across sources.
One criterion does not appear in this list: your sales pitch. The assistant does not retain a promise, it retains an attribute it can compare with the request. "Ideal for small spaces" is worth nothing without a width in centimetres. Reread your product pages with this question: for each benefit claimed, which measurable piece of data proves it? Where the answer is missing, the attribute needs to be added, in the feed as well as on the page.
Preparing your catalogue: the five workstreams
Making a catalogue recommendable does not require a redesign of the site. The work divides into five workstreams, to be carried out in this order because each conditions the next.
Open access to search crawlers
If ChatGPT's search crawler cannot read your pages, your products cannot be retained from your site. Check that OAI-SearchBot is not blocked in your robots.txt file, nor by a firewall or a content delivery network configured against "AI bots" in general. The GPTBot crawler, for its part, is used for model training: blocking it is a separate choice and does not remove your products from search.
Also check rendering: price, stock and specifications must appear in the HTML that is served, not only after scripts have run. The syntax of access rules is detailed in our guide to the robots.txt file.
Make the product feed reliable
The product feed is the most direct source. It must be complete, accurate and up to date. Sort its fields by lifespan, because the risk is not the same:
- Stable data: identifiers (including the global trade item number), brand, dimensions, materials, compatibilities.
- Changing data: price, promotions, stock, delivery times.
- Appraisal data: ratings, number of reviews, labels.
Changing data is what makes a recommendation fail. Set an update frequency and keep to it: an outdated price in a feed ends up in the answer, then contradicted on your product page. Appoint an owner for each family of fields, since attributes generally belong to the product information system and availability to the stock system.
Structure the product pages
The product page remains the page the assistant reads and to which it sends the buyer. Three elements make it usable:
- Product structured data consistent with the visible text: price, currency, availability, rating, identifiers.
- Specifications in a table, with values and units, rather than a descriptive paragraph.
- Answers to selection questions: for whom, for what use, with what limits, compatible with what.
The markup is described, type by type, in our guide to structured data; building a page that ranks and reads well is the subject of our article on the SEO product page.
Put your policies in black and white
Delivery, returns, warranty and payment methods are part of the comparison. An assistant that cannot find your conditions cannot set them against a competitor's, and retains the offer whose terms it can verify. Publish conditions with figures (lead time in days, charges, withdrawal period) on an accessible page and, where the format allows it, in the feed.
Make sure they are constant: the same information on the product page, the help page and the terms and conditions. A discrepancy between two pages is read as uncertainty.
Pay particular attention to exceptions, because they are what buyers ask about in follow-ups: bulky items delivered on a different schedule, products excluded from returns, regions you do not ship to, charges that depend on basket value. State each exception next to the general rule, in a sentence that can be read on its own. A condition that only appears in a footnote or a pop-up window is unlikely to be read, and the assistant will then describe your offer less precisely than a competitor's.
Look after what third parties say about the product
The review summary displayed next to a product comes from outside sources as much as from your site. Encourage verified reviews, reply to them, and make your products available for testing by the media and comparison sites in your sector. Users place a high level of trust in these summaries: 66% rely on AI outputs without checking their accuracy (Squid Impact, 2025). False information about your product, picked up in a summary, is therefore unlikely to be corrected by the reader: it is up to you to spot it.
The mistakes that keep a product out of the selection
When a product never appears in recommendations, the cause rarely lies in the quality of the product. Five flaws recur, and all of them are corrected in the data:
- Missing attributes: without a dimension, a compatibility or a material filled in, the product cannot be compared with a constraint expressed by the buyer.
- A title built for the site's search engine: a string of keywords does not say what the product is; a descriptive title, with the brand, the type and the distinguishing attribute, reads better.
- Poorly linked variants: sizes and colours presented as separate products, with no link between them, scatter reviews and blur the price.
- A price displayed before charges: if delivery charges only appear on the final screen, the price given in the answer will be contradicted on arrival.
- Missing or old reviews: the assistant lacks material to summarise what customers think of the product.
Deal first with the categories that carry your margin, not the whole catalogue: a batch of complete, accurate product pages teaches more than an entire catalogue that is half filled in.
Payment in the conversation: what changes when the purchase is completed in ChatGPT
Depending on the country and the merchant, the buyer can complete their order within the conversation, without going through the site's checkout. This payment relies on an exchange protocol between the assistant, the merchant and the payment provider, which is being rolled out gradually at the time of writing. Check availability in your markets before committing to a project.
Where this option exists, three things change for you. The product page is no longer necessarily seen: the feed carries all the information. The order arrives through an integration, not through your basket: your system must confirm price and stock at the moment of purchase. Finally, you remain the seller: shipping, after-sales service, returns and the customer relationship stay with you.
Before accepting these orders, lock down three points:
- Real-time verification of price and stock before confirmation.
- Substitution and refusal rules: what happens if the product is no longer available?
- Traceability: what information was presented to the buyer, at what moment, with what consent.
These safeguards belong to the general framework of delegated purchasing, covered in the reference page cited in the introduction. As long as payment in the conversation is not open in your markets, the useful work stays the same: clean data and written rules, which will serve whatever protocol is adopted.
Measuring what ChatGPT Shopping brings in
Measurement combines three readings. None is complete alone, and reading them together avoids drawing conclusions too quickly.
The presence of your products in answers
Build a set of purchase requests per category, worded as a customer would express them, with a use and a constraint. Submit them at regular intervals and note, for each one: are your products offered, at what rank in the selection, with what price and what summary, and which competitors appear. Repeat each request, because the selection varies from one session to another. Record errors too: wrong price, inaccurate specification, a discontinued product still being offered. This error log is worth as much as the presence reading: it points to the feed field or the page to correct, and it can be dealt with within the week.
Traffic and sales
In your analytics tool, isolate the sessions that came from chatgpt.com, then follow them through to the order: landing pages, conversion rate, average basket, return rate. Compare this segment with your other channels, without expecting the same volume. If orders are placed within the conversation, they do not appear as sessions: reconcile the order data from your integration. Finally, read the return rate for this segment: a rate above that of your other channels signals that the recommendation promises something other than what the product delivers, and so points to a badly filled-in attribute.
The warning indicators
Three signals deserve weekly monitoring: a gap between the price announced and the price on the site, a rise in orders cancelled because of stock-outs, and a drop in presence in a category where you used to be offered. Each one points to a specific workstream: feed, stock or product page content. Set a threshold and an owner for each, so that the alert triggers a correction and not just an observation. Always compare a category with itself, from one period to the next: a variation is a correlation over time, which seasonality or a stock-out at a competitor can also explain.
FAQ on ChatGPT Shopping
What is ChatGPT Shopping?
ChatGPT Shopping brings together ChatGPT's purchasing features: the user describes their need, and the assistant offers a selection of products with images, prices, a review summary and links to merchants. Depending on the country, the order can be completed within the conversation. The results are recommendations, not adverts.
How do you get your products to appear in ChatGPT?
Allow ChatGPT's search crawler to read your pages, supply a complete and up-to-date product feed, structure your product pages with measurable specifications and structured data, and publish your delivery and returns conditions. Selection then depends on the match between your attributes and the request.
Do you have to pay to be recommended in ChatGPT Shopping?
No. At the time of writing, the products offered are selected by the assistant on the basis of their data and their relevance; they are not purchased placements. The lever is the quality of information: complete attributes, accurate price and stock, verifiable reviews and conditions.
Is ChatGPT Shopping available in my country?
Availability varies by feature and by country, and it changes over time. Product search is offered more widely than payment in the conversation, which is being rolled out gradually. Check the status of the service in your markets when you start your work, and prepare your data without waiting.
What is the difference between ChatGPT Shopping and Google Shopping?
Google Shopping relies largely on adverts and on a ranking of results that the buyer filters. ChatGPT Shopping starts from a request in natural language and offers a short, explained selection, which the buyer refines through follow-ups. Both read product feeds, but selection does not follow the same rules.
Which product data does ChatGPT use?
It uses the feeds supplied by merchants, the structured data and content of product pages, and third-party sources such as reviews and tests. The decisive fields are price, availability, specifications, variants, images and delivery and returns conditions.
Continue reading
Depending on your situation:
- You want to be cited beyond product pages: if your guides and comparisons also need to appear in the assistant's answers, the rules are those of ChatGPT's web search → ChatGPT SEO.
- You are thinking about the agents to deploy on your own shop: if the question is which agents to install on the merchant side, and which to start with → AI agents for e-commerce.
- Organic traffic is still your main channel: if your categories and product pages first need to progress in classic search engines → e-commerce SEO.
- You need to enrich hundreds of product pages: if your descriptions lack specifications and answers to selection questions, the Incremys content production module generates the copy from your product data, in the brand's tone, with validation before publication.
.png)
.jpeg)

.jpeg)
%2520-%2520blue.jpeg)
.avif)