How ChatGPT Finds Product Information for Shopping
How ChatGPT Finds Product Information for Shopping
Learn how ChatGPT can use model knowledge, Search, public product pages and merchant feeds to find, check and surface ecommerce product information.
Haritha Kadapa
How Does ChatGPT Find, Verify, and Cite Product Information?
ChatGPT does not rely on one master copy of your product information. Depending on the experience being used, a product fact can come from knowledge learned during model training, live Search retrieval, public retailer and product pages used during shopping research, or structured product data supplied by merchants.
For an ecommerce team, those paths should be treated separately. Training affects what the model may already know; Search affects which webpages can be retrieved and cited; shopping research can read current retail information; and merchant feeds give ChatGPT structured product data such as price, availability and variants.
That distinction becomes important the moment two sources disagree.
If your PDP says a serum costs $32, an old retailer page says $38, and a feed still reports the product out of stock, “What does ChatGPT know about the product?” no longer has one simple answer.
One product can reach ChatGPT through several routes
Consider a shopper asking:
“Is this serum suitable for sensitive skin, what does it cost, and where can I buy it?”
Different parts of the eventual answer may require different information systems.
Information path | What it is useful for | Ecommerce implication |
Model knowledge | General product, brand or category knowledge learned during model development | Useful context, but not a reliable source for current price or stock |
ChatGPT Search | Retrieving current public webpages and linking or citing them in answers | Important PDPs and supporting pages need to be eligible for Search retrieval |
Shopping research / public retail sources | Reading product and retailer information while comparing products | Public product facts need to be current and consistent across channels |
Merchant product data | Structured product identity, descriptions, price, availability, variants and seller information | Catalog and feed quality become part of AI-shopping readiness |
These mechanisms overlap, but they should not be collapsed into “ChatGPT crawled my website.”
Training data is not the same thing as a live product lookup
OpenAI says its foundation models are developed using information from three broad sources: publicly available internet information, data accessed through third-party partnerships, and information provided or generated by users, trainers and researchers. During training, models learn patterns by adjusting their internal parameters; OpenAI explicitly says the model does not keep copies of the training material as a searchable database.
For ecommerce, that distinction matters.
Suppose a long-established brand has been discussed publicly for years. A model may already have broad associations with the brand or product category before any current webpage is retrieved.
That does not make model knowledge a dependable source for changing facts.
Price, stock, packaging, product names, formulations, sizes and retailer availability can change. Even when a model knows the product exists, current shopping information may need to come from a more recent source.
This is one reason the old idea of “GPT has a frozen copy of my website” is misleading.
The model has learned patterns from training information. That is different from looking up the latest state of a PDP.
ChatGPT Search creates a separate retrieval path
When ChatGPT uses Search, public webpages can be surfaced as current sources.
OpenAI uses OAI-SearchBot for its search products. Websites that want their content to be eligible for inclusion in ChatGPT Search should allow OAI-SearchBot; OpenAI says blocking it prevents the site's content from appearing in ChatGPT search answers, although links may still appear in limited navigational contexts.
This is distinct from GPTBot, which OpenAI uses for content that may be used in training foundation models.
The two permissions are independent.
An ecommerce company can therefore make a deliberate choice such as:
allow OAI-SearchBot for live Search visibility, while disallowing GPTBot for model-training use.
That is a much more useful technical distinction than simply asking whether “OpenAI is allowed in robots.txt.”
OpenAI also documents ChatGPT-User, a separate user-triggered agent that may access a webpage when somebody asks ChatGPT or a Custom GPT to interact with it. It is not the crawler used to determine Search inclusion.
For technical teams, the current OpenAI access model therefore looks roughly like this:
OpenAI agent | Primary role | What the site owner should know |
OAI-SearchBot | Search discovery | Controls eligibility for inclusion in ChatGPT Search |
GPTBot | Training-related crawling | Can be independently allowed or disallowed |
ChatGPT-User | User-triggered page access | Not the control for Search indexing |
That separation should be reflected in robots.txt decisions, CDN rules and log analysis.
Shopping research can go directly to current retail information
Shopping creates another route.
OpenAI says ChatGPT shopping research may use merchant product data, publicly available product information and other relevant retail sources. During research, it can read product pages, compare information such as price, features and reviews, and run additional searches as the shopper adds constraints.
This is closer to the problem an ecommerce team actually cares about.
A shopper may ask:
“I need a carry-on under $200 that fits European airline cabin limits.”
ChatGPT is not merely looking for a webpage containing “carry-on luggage.”
It may need to establish dimensions, price, product availability, retailer information and possibly review evidence before deciding which products deserve consideration.
If dimensions are explicit on Brand.com but wrong on a retailer page, or the correct variant is missing from the catalog feed, the product now has a product-truth consistency problem, not simply a content problem.
OpenAI also warns users that current shopping information can still be wrong or stale and advises checking the retailer's site for final price, availability, shipping and related purchase information.
Grounding improves recency and traceability.
It does not guarantee that every retrieved source is correct.
Direct product feeds create a structured catalog path
For merchants, OpenAI now supports structured product feeds specifically for discovery in ChatGPT.
The current product specification works at the purchasable product or variant level and requires core information including an identifier, title, description, product URL, image, availability, price and brand. Optional fields can describe additional attributes, variants, shipping and returns.
OpenAI recommends that full catalog feeds be refreshed at least daily, and its documentation specifically tells merchants to update price and availability when those values change.
This creates an important difference between a normal webpage and a commerce feed.
A PDP is written partly for shoppers and may contain rich explanation, imagery, FAQs and merchandising.
A product feed provides structured product facts at scale.
For a catalog team, both matter.
The PDP may answer why a product fits a need.
The feed can make its identity, price, availability and variant information much easier to ingest consistently.
How does ChatGPT decide which product information to use?
There is no documented rule saying, for example:
feed beats PDP, PDP beats retailer, retailer beats review.
OpenAI's shopping documentation describes a mixed information environment.
For product selection, ChatGPT can consider structured metadata from first- and third-party providers, other third-party content, the shopper's query and context, and model-generated responses produced before new search results are considered.
That means ecommerce teams should avoid thinking of the process as a simple source hierarchy.
The more practical problem is consistency.
If the same SKU is represented as:
$32 on the PDP;
$38 at one retailer;
out of stock in an old feed;
30 ml on Brand.com;
50 ml on a marketplace listing;
there is no content-formatting tactic that makes those facts collectively clean.
Someone has to fix the underlying product information.
“Verification” is not the same as proving a fact is true
This is where the useful part of the older grounding discussion belongs.
When ChatGPT retrieves live material, the generated answer can be grounded in external sources rather than relying only on learned model knowledge.
That improves the system's ability to use recent, inspectable information.
But grounding should not be described as a truth engine.
A retailer can publish the wrong dimensions.
An old product page can remain indexed.
A feed can fail to refresh.
Two legitimate merchants can show different prices at the same time.
Even OpenAI's own shopping documentation tells users to verify final purchase information on the merchant site because prices, stock and discounts can change quickly.
For ecommerce, the useful objective is therefore consistent, current product truth across the sources the system can access.
A citation and a product-data input are not the same thing
Another common mistake is assuming that every piece of product information influencing ChatGPT must appear as a visible citation.
Search answers can cite webpages.
Shopping research explicitly reads public retail sites and cites sources during the research process.
But ordinary shopping results can also use structured metadata supplied by merchants or third-party providers to construct product titles, descriptions, price displays and other product information. OpenAI notes that ChatGPT may even simplify product titles and descriptions because different merchants describe the same item differently.
So an ecommerce team needs to monitor two related but different things:
Which sources does ChatGPT visibly cite?
and
Which product information is ChatGPT actually showing?
A citation audit cannot tell you everything about the catalog layer.
A practical product-information test
The cleanest way to diagnose this is to choose one priority SKU and follow one fact through every representation.
Suppose the product is a waterproof jacket.
Pick facts that matter to real buying questions:
waterproof rating, price, available sizes, insulation type and stock.
Then compare:
Location | Waterproof | Price | Size / variant | Stock |
Brand PDP | Record | Record | Record | Record |
Product feed | Record | Record | Record | Record |
Product schema | Record | Record | Record | Record |
Major retailer A | Record | Record | Record | Record |
Major retailer B | Record | Record | Record | Record |
ChatGPT result | Record what appears | Record | Record | Record |
Sources / merchants surfaced | Record URLs | — | — | — |
Run the same shopper question several times under the same market and test conditions.
Do not start by asking whether the brand “ranks.”
Look for disagreements.
If ChatGPT reports the wrong price, which accessible source still contains it?
If the correct jacket appears but one size is missing, is the feed incomplete?
If a competitor appears for “lightweight waterproof jacket” while your product does not, is weight clearly represented anywhere in the catalog?
That is the type of technical test that can turn a vague AI-visibility problem into a fixable catalog issue.
For publication, this section should be completed with one actual Gravton-controlled SKU test rather than fabricated results.
What should technical and catalog teams check?
Start with access.
If ChatGPT Search visibility matters, make sure OAI-SearchBot is not unintentionally blocked by robots.txt, CDN rules or security infrastructure. Treat GPTBot training permission as a separate policy decision.
Then inspect the product-information chain.
The PDP, product schema, product feed and major retailer listings should agree on identity, variant, price, availability and the attributes that determine whether the SKU fits important shopper questions.
For merchants eligible to provide product data directly to OpenAI, feed freshness should be operational rather than occasional. OpenAI recommends at least daily full-feed updates and also provides APIs for feed and product updates.
Finally, test the output.
Use stable shopper questions. Record the products ChatGPT surfaces, the product facts it presents, public sources it cites where visible, and the merchants it shows.
A technically accessible catalog can still lose because the product does not match the need or because competitors have stronger evidence.
But without accurate product truth, the system is being asked to make the decision from contradictory inputs.
Find your AI search visibility gaps
See which questions surface your brand, which ones send buyers to competitors, and where your content has visibility gaps.
VISIBILITY & CONTENT STRATEGY