How does ChatGPT decide which products to recommend?

When a shopper asks for the best cordless vacuum for pet hair under a budget, ChatGPT classifies the request as shopping intent and searches rather than answering from the model alone. The candidates come from the web index built by OpenAI's search crawler, OAI-SearchBot, from pages fetched on the shopper's behalf by ChatGPT-User, and from product data merchants have supplied directly through OpenAI's product feed program.

The model then reads what came back and writes a recommendation. Its training data shapes the vocabulary and the general knowledge of the category, but the specific products, prices and availability in a shopping answer are retrieved. A brand that is not retrievable at that moment is not in the running, no matter how well known it is.

This matters for a store because it moves the question from reputation to readability. If your product pages cannot be crawled, if your prices are rendered only by JavaScript the fetcher does not execute, or if your feed is not submitted, the model has nothing to work with.

ChatGPT shows product cards with image, price, seller and review summaries alongside its prose. OpenAI has published a product feed specification that lets merchants supply that data directly, with fields that will look familiar to anyone who runs Merchant Center: id, title, description, link, image, price, availability, brand, GTIN, shipping and return information, and an enable flag for its checkout program.

Where a merchant has submitted a feed, the card can carry accurate price and stock. Where it has not, the card is assembled from whatever the crawler found, which can be stale or incomplete. Either way, the model cross-reads the page and the structured data. A feed price that differs from the page price is the kind of inconsistency that gets a product left out of a shortlist or shown with a wrong figure.

The practical rule is to treat the OpenAI feed like any other channel feed: generate it from the same source of truth as your Google feed, keep it refreshed on the same cadence, and check it against the live page.

A shopping answer is usually built from a mix of sources: the merchant's product pages, publisher round-ups, forum threads, review aggregators and marketplace listings. When the shopper asks which option is best, the model leans on pages that compare options, because they answer the comparative question directly. Your own product page rarely says your product is better than a rival's; a magazine test or a long forum thread does.

So a product's chance of being recommended depends heavily on whether it appears in the comparison content that gets retrieved for that category. That is earned by being reviewed, being stocked by retailers whose pages get crawled, being discussed in communities, and having product specifications that a writer can quote accurately.

You can see this yourself. Ask ChatGPT a category question and expand the sources. If every cited page is a publisher list and your brand is on none of them, the fix is not on your website.

Shopping prompts are long. They carry a budget, a use case, a location, sometimes an allergy or a pet or a body type, and often a rejection of something already tried. ChatGPT filters the retrieved candidates against those constraints and, where memory is enabled, against what it already knows about the person from earlier chats.

That means two shoppers asking a similar question get different lists, and there is no single ranking to win. A product is recommended when it is retrievable and when the retrievable information about it answers the constraints: ships to Canada, fits a wide foot, fragrance-free, works with a specific mount.

The implication for product content is to state the attributes shoppers filter by in plain text on the page and in the feed. A specification table with material, dimensions, compatibility, country availability and care instructions gives the model something to match against. A lifestyle paragraph does not.

There is no analytics report for this, so build a simple audit. Write the questions a buyer in your category would actually ask, in their words and with their constraints: five buying intents, each phrased five ways. Run each prompt a few times in fresh conversations with memory off, and record which products and stores appear, which sources are cited, and whether your own products show with correct price and availability.

Repeat monthly. Look for movement after you submit or fix a feed, after a product page rewrite, or after a new review lands on a publisher site. Because the outputs vary, judge on frequency across runs rather than on any single answer.

One caution on interpretation: appearing in a ChatGPT recommendation is not a promise of traffic or sales, and no one can promise a placement. Treat the audit as a readability check on your product data, and measure the commercial effect separately through referral traffic and revenue per session from chatgpt.com.