ChatGPT Shopping: How Products Actually Get Recommended

By Muntazir · Published 2 October 2026

On 29 September 2025, OpenAI turned on a buy button inside ChatGPT. Etsy’s US sellers went first, one item at a time. More than a million Shopify merchants — Skims and Glossier among the names OpenAI mentioned — were told to expect it next (Marketing Dive). None of those retailers earned the slot with better product copy. They earned it by filing a feed: a plain list of what they sell, what it costs, and whether it’s still in stock, built to a format OpenAI’s own commerce documentation spells out field by field.

What actually decides whether a product gets shown?

OpenAI publishes the exact shape a merchant has to send. A product needs, at minimum, a stable ID and at least one variant, and that variant needs its own ID and a title. From there, the fields that decide whether ChatGPT can do anything useful with the listing: a price and a list price, each carrying an amount and a three-letter ISO 4217 currency code, and an availability object stating whether the item is in stock, backordered, preordered, out of stock or discontinued (OpenAI developer documentation). Leave any of those out and there’s nothing for the system reading the feed to put a price tag or a stock badge on — which is most of what makes a shopping answer usable at all.

Does a retailer have to sign up for Instant Checkout to appear?

No. Instant Checkout and a separate feature called Shopping Research answer different questions. Instant Checkout needs the merchant feed above, plus a way to hand off orders and payments through the Agentic Commerce Protocol OpenAI built with Stripe; Etsy, Shopify, Walmart and Target are the retailers running it so far. Shopping Research, which launched two months later on 24 November 2025, is a comparison and recommendation tool: a version of GPT-5 mini, trained specifically on shopping tasks, that reads reviews and product pages across the open web rather than a submitted feed. OpenAI’s own published comparison puts that model at 64% accuracy on shopping questions, against 56% for general-purpose GPT-5 Thinking and 37% for ordinary ChatGPT search (Digital Commerce 360) — a gap worth noticing in both directions: better than a plain search, and still wrong more than a third of the time even in its own best-performing mode. A retailer with no feed and no checkout integration can still turn up in a Shopping Research answer on the strength of its pages and reviews. It just won’t carry a price, a stock badge, or a one-click purchase.

What does a check that reads structured data actually find on a product page?

A structured-data check inside one AI-visibility audit engine spends part of every scan on this same layer, separately from anything OpenAI reads. On a page carrying schema.org Product markup, it asks whether the offer states a price, whether that price carries a currency, and whether the product declares availability — a missing price is flagged at medium severity, because a price-less listing can’t appear in a shopping surface at all. A fourth question is scored more lightly: whether the product carries a global identifier, an SKU, GTIN or MPN. Miss that one and the severity drops to low; the listing still shows a price and a stock status, it just can’t be matched against the same item sold somewhere else.

That ordering lines up with OpenAI’s own feed spec in one place and parts ways with it in another. Both require the same two things above everything else: a stable identifier and a price with its currency attached. Where they diverge is the global identifier itself. Google’s own merchant guidance treats a GTIN as close to mandatory for a new listing, limiting visibility for anything submitted without one once the manufacturer has assigned it. OpenAI’s commerce spec never names a GTIN, MPN or SKU as a requirement — it offers an optional barcode field on each variant and stops there. A retailer who spent years hardening a feed against Google Merchant Center’s identifier rules has already over-satisfied the one requirement OpenAI actually enforces.

Does any of this apply to a business that isn’t selling physical products?

Most of what that same audit engine checks sits somewhere else entirely — a plumber’s service area, a dentist’s opening hours, a solicitor’s accreditations: the pages a local or service business runs, not a product catalogue. None of its checks read a merchant feed, and it has no visibility into Instant Checkout or Shopping Research specifically. The overlap exists only because the underlying question is identical on both sides: can a machine read, without guessing, what’s on offer and whether it’s actually available. A schema checker answers that question for whatever structured data a page already carries, catalogue or not, and a free scan covers the same ground alongside the checks a service business is more likely to be judged on. The feed in this piece belongs to OpenAI, built for retailers; the wider mechanics of how a model reads a page are the same ones underneath it, regardless of what’s being sold.

Skims and Glossier did not reach the front of OpenAI’s merchant queue with sharper brand pages. A feed stating a stable ID, a price, a currency and a stock status is a smaller, duller piece of writing than most of what a marketing team produces — and it’s the one document ChatGPT actually reads before it decides to show anything at all.