ChatGPT Shopping
ChatGPT Shopping is the set of product results shown inside a ChatGPT answer. The user describes what they need and receives a selection of specific products with prices and links. Unlike Shopping in a search engine, you do not enter that selection with a bid, you enter it with the quality of your product data.
In short
| What it is | Product results inside a conversation, with price, description and a link to the shop |
| How a product gets there | Through structured data on the product page, feeds and other sources about the product |
| What it is not | Not an ad format you can buy with budget |
| Who it matters for | Shops whose catalog carries descriptive properties, not just a code and a price |
| What helps most | Complete data, consistent price and availability, reviews, and clear answers to what the product is for |
How ChatGPT Shopping works
Where the products come from
ChatGPT's product layer draws on three sources. The first is a product feed the shop sends OpenAI in the required format: identifier, title, description, price, availability, brand, GTIN, images, category, parameters such as material and dimensions, rating with review count, shipping and return terms. The fields largely overlap with the Merchant Center feed, so it can be generated from the same product data. The second source is the web: the OAI-SearchBot crawler fetches product pages and reads the text and the schema.org structured data, type Product with a nested Offer (price, currency, availability) and AggregateRating. The third source is third-party sites: reviews, comparison sites, marketplaces and expert tests, from which the model takes ratings and opinions. The data merges into one product record, which may list several sellers at different prices.
From the query to the product card
- The model recognizes shopping intent and breaks the query into conditions: product type, budget, dimensions, features, purpose. If the user has memory turned on, it adds known preferences.
- From the product index it picks candidates matching the type and category.
- It filters the candidates by the conditions: price under the limit, attribute at the required value, product in stock. The filter only works with values actually present in the data.
- It ranks the remaining products by match with the query, rating, price and availability, and for the top few writes why they fit, including pros and cons drawn from reviews.
- The answer shows cards with an image, price, rating and a link to the seller. The seller is chosen by availability and price, not by a bid per click.
Data refresh: what applies between two loads
The feed is loaded repeatedly; the crawler comes on its own schedule. Between two loads the last known price and availability apply, so for frequently changing products the export frequency matters, and the page and the feed have to say the same thing. A sold-out product is better marked unavailable than removed from the feed: the record stays and returns to the selection at the next load after restocking, instead of being created again.
Worked example: a user wants a children's bike helmet under €40, head circumference 52 to 56 cm, with an LED light. The shop has 90 helmets, 54 of them under €40. Head circumference is filled in for 30 of those, and 12 match the 52 to 56 cm range. The LED light is listed as an attribute on 4 helmets; on another 5 it is only visible in the photo. So 4 products enter the selection and 5 qualifying ones drop out, because their feature is not readable in the data.
How a product enters the selection
The model answers an intent, not a keyword. When someone writes that they want a quiet dishwasher for a flat, under a certain price, the noise level, the dimensions and the price all have to be readable from structured data. A missing attribute means the product does not match the condition and drops out, even when it actually meets it.
Descriptive properties beat keywords
A classic feed is built for parametric search. A conversational query asks about use: who it is for, what it suits, what is in the box. That is exactly what most catalogs lack.
Price contradictions eliminate the product
When the price on the site differs from the feed or from a marketplace, the model cannot verify it and offers a different shop instead.
From our own practice
The most effective preparation we currently run in client accounts is conversational attributes in Google Merchant Center. It is the same work: adding to the product data the properties people ask about in sentences.
The attributes are filled in as sentences rather than parameters, and on large catalogs they are generated by rules over the feed instead of by hand. In the accounts we manage, the most common reason a product fails to show is not a low bid, it is a missing or inconsistent attribute.
Common mistakes
- Waiting for an ad format. There is nothing to buy yet, only a data layer to prepare.
- Copying the manufacturer's description. The same text as twenty other shops is no reason to pick you.
- Leaving products without reviews. Ratings are one of the few quality signals a model can work with.
- Blocking crawlers. Product data has to be reachable or it never enters the selection.
Related terms
See also agentic commerce, product feed, AI Mode and AI visibility.
Frequently asked questions
Can you advertise in ChatGPT Shopping?
Paid formats are only being introduced and differ by market. The product selection inside an answer cannot be bought.
Do I need a separate feed for it?
Not necessarily. The base is the feed you already use for Shopping and comparison sites, extended with descriptive properties. OpenAI also accepts a product feed in its own specification, which reuses the same data.
How do I find out whether my products show up?
By manually testing a sample of model queries and watching sessions from chatgpt.com in analytics. There is no dedicated report yet.
How we can help
Feeds and product data are daily work for us, marketplace layer included. Details are on the E-commerce marketing agency page.