A brand that has spent a year earning mentions in AI answers asks a reasonable question. All that work, and why don’t our products show up when somebody asks ChatGPT what to buy?
Because the shopping results and the answer text come from two different systems. One reads a product feed you submit. The other reads the web.
Getting cited in an answer takes editorial work. Getting into the shopping results takes a catalog integration. Doing one well moves you nowhere on the other.
This is what each mechanism is, what OpenAI’s own documentation says about the feed route, and how to tell which one your business needs.
What ChatGPT Shopping Is
It’s a product discovery surface inside ChatGPT, and it runs on catalog data that merchants hand over deliberately.
OpenAI’s merchant page describes it as sharing a product feed so products can reach shoppers as they explore options, compare, and decide what to buy. The products appear in rich results carrying images, pricing and key details.
The Purchase Finishes on Your Site
By default the shopper finishes on your own site or app, and OpenAI states there are no fees for purchases completed there.
That matters for how you evaluate the channel. It behaves like a referral surface rather than a marketplace that takes a cut, at least in the shape described today.
What It Is Not
It isn’t the part of ChatGPT that names brands in a written answer. When someone asks which project management tool suits a small team, the reply is assembled from retrieved sources, not from anybody’s product feed.
Those two surfaces sit next to each other in the same interface, which is exactly why teams assume one investment feeds both.
The Feed Is the Entry Point
Everything about the shopping surface starts with structured catalog data, and OpenAI’s documentation is unusually direct about that.
The key concepts guide describes merchants providing a secure, regularly refreshed feed carrying identifiers, descriptions, pricing, inventory, media and fulfillment options.
What the Feed Carries
Required fields exist so price and availability display correctly. Beyond those, OpenAI’s own wording is that recommended attributes such as rich media, reviews and performance signals improve ranking, relevance and user trust.
Read that sentence twice if you own a catalog. The company is stating plainly that richer data changes placement, which makes feed quality a merchandising decision rather than a technical chore.
Two Delivery Routes
You can send the feed as an uploaded file or through an interface, and the documentation recommends combining them. The suggested pattern is the entire feed once a day as a file, with updates sent through the day.
One detail is worth knowing before you plan: promotions data can only be supplied through the interface route, not by file.
How It Gets Validated
Integration begins with an initial sample feed for validation, followed by daily snapshots. So the first thing OpenAI sees is a sample, and the thing that keeps you accurate afterwards is the refresh cadence.
If your prices move often, the refresh is the part that protects you. A stale feed misrepresents your own catalog to a shopper who’s mid-decision.
A Shortcut If Your Data Already Fits
Feeds built to Google’s product data format can be accepted in that shape, with OpenAI mapping the supported fields into its own schema. The documentation is careful that this covers a core profile rather than every attribute and market.
For most merchants with an existing shopping feed, that turns a rebuild into a mapping exercise.
Who Is Already In, and Who Applies
This is the detail most coverage skips, and it decides whether the channel is a project or a form.
If you sell through Shopify or Etsy, OpenAI states your catalog is already integrated and no application is needed. Nothing to build, nothing to submit.
Everyone else applies through the merchant page and joins a waitlist. The get-started guide describes feed onboarding as currently available to approved partners.
So for a large share of smaller merchants the work is already done. The remaining question is whether their product data is good enough to compete once it’s there.
The Protocol Underneath
The shopping surface is built on the Agentic Commerce Protocol, which OpenAI describes as open infrastructure for commerce inside ChatGPT and publishes publicly.
Today the documentation scopes it to product discovery and more accurate recommendations through shared product data. OpenAI states the intention of expanding it to the fuller shopping journey over time.
The Three Flows
Supporting checkout inside ChatGPT means implementing three separate things: sharing the product feed, handling orders and checkout, and handling delegated payment.
Most merchants only need the first one. Feed sharing is what makes products appear; the other two are what let a purchase complete without the shopper leaving.
The gap between them is bigger than it sounds. OpenAI’s production guidance asks for sandbox certification covering session creation, shipping updates, payment tokenisation, order completion and error handling, each demonstrated end to end with logs.
The feed is a data exercise and checkout is an engineering programme. Decide on them separately, because bundling the two is how the whole thing stalls.
What Stays on Your Side
The division of labour is spelled out, and it’s more favourable than people expect. ChatGPT collects buyer, fulfillment and payment details and passes them to the merchant’s endpoints.
The merchant then validates the order, determines fulfillment options, calculates and charges sales tax, runs its own payment and risk analysis, and charges through its existing processor. The merchant accepts or declines.
You keep the commercial decisions, the risk model and the payment relationship. That is a meaningfully different arrangement from listing on a marketplace.
What Good Product Data Looks Like
OpenAI publishes best-practice guidance for feeds, and most of it is merchandising advice rather than engineering advice.
Write Factual Descriptions
The guidance asks for concise, factual copy that helps a user understand the product, and it accepts plain text or bullet-style text equally.
That rules out the register most catalogs default to. A description written to persuade reads worse here than one written to inform.
Use Optional Fields on Purpose
Optional fields can improve answer quality without being required for ingestion. The guidance says plainly that a field needing brittle transforms should be left out until the underlying data is stable.
An empty optional field costs you less than a wrong one. That’s a useful licence if your catalog data is uneven across categories.
Model Variants at Row Level
Each purchasable option needs its own identifier sitting under a stable parent product. Title, address, description, media, availability and price stay variant-specific wherever they differ.
Shared media belongs at product level only when it genuinely applies to every variant. Getting this wrong is how a shopper sees one colour at another colour’s price.
Keep Links Valid and Consistent
Product and media addresses need to be valid and properly encoded, spaces included. Seller links should be durable public pages, and the seller name should be the one a shopper expects to see on the listing.
How to Tell Whether It’s Working
You measure it by tagging the feed, because the traffic lands on your own site and looks like any other referral otherwise.
The best-practice guidance suggests adding attribution parameters to the product address so feed clicks are identifiable, and keeping those parameters consistent across snapshots.
Consistency is the part teams miss. A tracking parameter that shifts between daily snapshots splits one channel across multiple rows in your reporting, and nobody notices until the quarter is being explained.
Tag the feed before you launch it, not after the first good month. Retrofitting attribution leaves the early data unattributable.
Where Merchants Get This Wrong
Treating the Feed as an Export
A feed generated once and forgotten drifts away from the catalog behind it. Price and availability are the two fields a shopper acts on, and both change without anyone deciding they should.
Waiting for the Application to Clear
Approval is outside your control and data quality isn’t. The catalog work is the same either way, so it belongs in front of the queue rather than behind it.
Expecting the Feed to Answer Category Questions
A feed describes products you already sell. It says nothing when a shopper is still deciding what kind of thing to buy, and that conversation happens on the other surface entirely.
Why Citation Work Doesn’t Put You in the Carousel
It doesn’t because the two surfaces read different inputs, and no amount of quality on one input reaches the other.
Earning a mention in a written answer depends on what published sources say about you. A model retrieves pages, reads passages and repeats what it finds, which is why editorial placements and reviews move that needle.
Appearing in shopping results depends on a feed you submitted, containing fields you control. There’s no editorial step, no retrieval of third-party pages, and no way for a great roundup mention to insert your product into it.
| Question | Answer surface | Shopping surface |
|---|---|---|
| What does it read? | Published pages it retrieves | The product feed you submit |
| Who controls the input? | Publishers and reviewers | You, directly |
| What improves it? | Editorial coverage and mentions | Feed completeness and accuracy |
| How fast does it change? | Weeks to months | As fast as your refresh |
| Who can use it? | Any brand | Merchants with a catalog |
The row that surprises people is the fourth one. Feed data updates on your schedule, while increasing brand mentions in AI search runs on the pace of publishing and retrieval.
What Still Matters Outside the Feed
None of this makes editorial work optional for a merchant, and the reason is the question a shopper asks first.
Category Questions Come Before Product Questions
Somebody deciding between categories, brands or approaches gets a written answer built from retrieved sources. Your feed has no part in that conversation, and it’s the conversation that frames the purchase.
A shopper who arrives at the comparison already believing your brand is the sensible option behaves differently from one meeting you cold in a list.
Reviews Feed Both
Reviews are one of the few inputs that count on both sides. They’re a recommended feed attribute in OpenAI’s own documentation, and they’re also the third-party evidence a model leans on when it writes about your category.
Your Own Pages Still Get Retrieved
Product and category pages are still fetched and read for the answer surface, which is what AI search optimization for ecommerce covers in practical terms.
An Open Protocol Changes the Maths
The protocol is published openly, with public documentation and a public repository, rather than kept as an internal interface.
That matters for how you value the work. A feed built to a proprietary spec is a bet on one company. A feed built to a published protocol is closer to infrastructure other assistants could read.
Nobody should promise you portability that hasn’t happened yet. What’s fair to say is that the format is documented in the open, which is a different risk profile from a closed integration.
Treat that as a reason to keep your catalog data clean in general, rather than as a reason to expect a second surface to arrive on a schedule.
If You Sell Software or Services
You have no catalog to submit, so the shopping surface isn’t a channel you can enter, and that’s worth saying plainly rather than working around.
The buying conversation in your category happens entirely in written answers. Somebody asking which tool fits a twelve-person team gets a reply assembled from retrieved pages, and every input to it is editorial.
What Replaces the Feed for You
Coverage on pages an assistant retrieves does the job the feed does for a merchant. Category roundups, comparison posts and review pages are the equivalent inventory, and you earn a place in them rather than uploading one.
The mechanism is covered in niche edits for AI search, where the wording around a placement is what an assistant can repeat back.
The One Thing That Carries Across
Reviews count on both sides. For a merchant they’re a recommended feed attribute; for a software business they’re third-party evidence a model reads when it writes about your category.
So a review programme is the rare investment that pays whichever surface your buyers use, which makes it the sensible first move when you’re unsure.
How to Decide Which One You Need
You decide by asking what a buyer types before they buy from you, then matching the surface to the question rather than to the channel that sounds newest.
If You Sell Physical Products at Volume
Do the feed. If you’re on Shopify or Etsy it costs you nothing but attention to data quality. If you’re not, the application is cheap to submit while you keep working on everything else.
If You Sell Software or Services
The shopping surface isn’t your channel, because there’s no catalog to submit. Everything you can influence sits on the answer side, which is where AI visibility for ecommerce brands and its B2B equivalents apply.
If You Sell Considered Purchases
Do both, in that order. The written answer shapes the shortlist, and the shopping result closes it, so the feed without the editorial work leaves you as an unfamiliar name in a comparison.
Frequently Asked Questions
Does ChatGPT Shopping cost anything?
OpenAI states there are no fees for purchases completed on your own site or app, which is the default flow. Treat that as the arrangement described today rather than a permanent guarantee, because the documentation says the protocol will expand over time.
Do I need to be on Shopify to take part?
No. Shopify and Etsy catalogs are described as already integrated with no application needed, and every other merchant applies through OpenAI’s merchant page. Feed onboarding is currently limited to approved partners.
Will good content get my products into shopping results?
No, and this is the most expensive misunderstanding in the area. Shopping results are built from submitted catalog data, so editorial coverage cannot place a product there no matter how strong it is.
Does the feed help me get mentioned in written answers?
Not directly. The answer surface reads published pages rather than merchant feeds, so mentions still come from coverage, reviews and your own retrievable pages.
How often should the feed refresh?
OpenAI’s guidance describes a full feed once a day with updates through the day, and integration starting with a sample feed for validation. Match the update frequency to how often your prices and stock move.
Is this the same as running ads in ChatGPT?
No. Advertising uses a separate feed schema with its own eligibility flag. An ads feed and a standard product feed are configured differently, even though they share a base structure.
Check Which Surface Your Buyers Use
Run the question your customers ask before they buy, once as a category question and once as a product question. Watch which of the two surfaces answers, and whether you appear in either.
That five-minute test tells you where the gap is. If you’re missing from the written answer, the work is editorial and our AI visibility programme for e-commerce is built for exactly that. If you’re missing from the shopping results, the work is your feed.