Get products into ChatGPT Shopping by submitting a structured product feed via SFTP. The required fields, validation, refresh cadence, and how to verify agents can reach your pages.
You get products into ChatGPT Shopping by submitting a structured product feed, then keeping it accurate. There is no ad auction to win and no placement to buy. ChatGPT reads a file you provide, matches its contents against what a shopper asked for, and shows the products that fit. This guide walks the whole path a merchant can act on this week: who can list today, the access checklist, building the feed field by field, submitting it, keeping it fresh, the shortcut for Shopify and Etsy sellers, and how to confirm agents are actually reaching your pages.
ChatGPT Shopping is the product-surfacing layer inside ChatGPT. When someone asks for a recommendation ("a quiet under-desk treadmill under $500," "merino base layers for winter hiking"), ChatGPT can show actual products with images, prices, and links rather than just describing categories. The results are organic. OpenAI has been explicit that placement is not paid, so you cannot bid your way to the top. Relevance and data quality decide what shows.
Who can list today is the part where precision matters. The merchant feed program exists and merchants can submit catalogs, but OpenAI has not published a fully open self-serve portal with the same maturity as, say, Google Merchant Center. Access, regional availability, and onboarding timelines are rolling out and are not entirely public. So treat the program as real and live, but expect the edges (exact approval windows, which countries, any fees) to be less documented than a mature ad platform. Where this guide states a mechanism, it is the mechanism. Where a number would be guesswork, this guide says so instead of inventing one.
The honest framing: if you sell physical products with a clean catalog, you are in the target group. If you sell services, digital subscriptions, or anything without a SKU and a price, the product feed is not built for you yet.
Run this before you touch a feed file. Half of feed problems are actually page problems.
Your product pages must be reachable by AI crawlers. ChatGPT uses user agents like OAI-SearchBot and ChatGPT-User to fetch and verify pages. If your robots.txt blocks them, or a firewall rule challenges them, the products in your feed point to pages the system cannot confirm. Allow these agents explicitly.
Prices and availability must exist in the server-rendered HTML. A feed says a product costs $79 and is in stock. If the live page renders that price only after JavaScript runs, the verification step can disagree with your feed, and disagreement erodes trust in your data. Server-render the facts that the feed asserts.
Your link URLs must be stable and canonical. Each feed item points to one product URL. If that URL redirects through three hops or changes per session, matching gets unreliable. Use the canonical product URL.
You need a way to deliver a file over SFTP. The feed is not pasted into a web form. It is a file dropped on a server. Confirm you can generate it and push it on a schedule before you start formatting fields.
If you run a Shopify or Etsy store, skip ahead to the platform section first. You may not need to build any of this by hand.
The feed is a structured catalog file. ChatGPT's merchant feed accepts JSONL (one JSON object per line) or CSV, delivered over SFTP. After you submit, OpenAI validates the feed against a sample batch before ingesting the full catalog, so formatting errors get caught early rather than silently dropping products. Once accepted, the feed can refresh frequently, which is what keeps price and stock honest.
Here are the fields that carry the weight. Anyone who has built a Google Merchant Center feed will recognize the shape, because the product-data conventions overlap heavily.
id — a unique, stable identifier per product or variant. Do not recycle IDs between different products; the system treats the ID as the product's identity.title — the product name as a shopper would search it. "Patagonia Men's Capilene Cool Lightweight Crew" beats "PCCLC-Blue-M." Front-load brand and product type.description — what it is, what it's for, key specs. This is matched against natural-language queries, so write for a human reading an answer, not for a keyword scanner.link — the canonical product page URL. One stable URL, no session parameters.price — the current price with currency. Keep it equal to the page. A feed/page mismatch is a credibility problem.availability — in stock, out of stock, preorder. This field decides whether you even belong in a "buy now" answer.image_link — a direct URL to a clean primary product image. Agents and the UI both use it; a broken or watermarked image hurts the listing.brand — the manufacturer or brand. Critical for branded queries ("Anker power bank").gtin — the global trade item number (UPC, EAN, ISBN). This is the field that lets the system disambiguate your product from a near-identical one. If you have it, include it.product_category — where the item sits in a category taxonomy. It helps the matcher understand the product type beyond what the title says.One pattern holds across every merchant feed system: completeness improves match rate. A product with a real GTIN, an accurate category, true stock status, and a description written in plain language gets matched to more relevant queries than a sparse record with a title and a price. Sparse feeds technically validate. They just lose to complete ones when a shopper's question is specific. Treat empty optional fields as missed matches, not saved effort.
Two failure modes to avoid. First, titles built from internal SKU codes; the matcher reads titles as language. Second, stale prices baked into a feed you regenerate monthly while your store changes prices weekly. Both are common, both are fixable in the build step.
Submission happens through OpenAI's merchant onboarding, not by emailing a spreadsheet around. The high-level flow: you register as a merchant, you configure an SFTP destination where your feed file lives, and OpenAI pulls and validates it. The validation pass runs your file against a sample batch, flags malformed rows or missing required fields, and confirms the linked pages resolve before the full catalog is ingested.
Be precise about what is and isn't public here. The existence of the merchant feed program and the SFTP-plus-validation mechanism is established. The exact onboarding screens, any eligibility review, approval timelines, and whether there is a fee are not fully disclosed and have shifted as the program rolls out. If you read a blog post quoting an exact "approved in N days" number or a specific dollar fee, treat it with suspicion unless it links to OpenAI's own documentation. The defensible statement is: you submit a feed over SFTP, it gets validated, and accepted products become eligible to surface. Timelines and costs, plan for a range, not a promise.
Practical move while you wait on access: get the feed file built and validating against the spec now, so the moment your merchant account is live, you submit a clean catalog instead of debugging field formats under pressure.
A product feed is a living file, not a one-time upload. The two things that go stale fastest are price and availability, and those are exactly the two fields a shopper acts on. A feed that lists an out-of-stock item as available, or last month's price, produces a bad answer, and bad answers are how a merchant quietly loses trust in the system's eyes.
Refresh as often as your catalog actually changes. If prices move weekly and stock moves daily, your feed should reflect daily. The feed mechanism supports frequent refreshes precisely so the data can track reality. There is no advantage to a slow cadence.
The completeness point compounds over time, too. As you add GTINs to products that lacked them, fill in categories, and tighten descriptions, your match rate against specific queries improves. A merchant who treats the feed as infrastructure, monitored and updated, beats one who submitted once and forgot it. This is the same discipline that wins in our agentic checkout readiness checklist: the data layer is only as good as its freshness.
If you sell on Shopify or Etsy, you may not need to build a feed by hand at all. Those platforms have catalog integrations that surface eligible products into ChatGPT Shopping through the platform itself, rather than each merchant standing up their own SFTP feed. Your existing product data, the titles, prices, images, and stock you already maintain in the admin, becomes the feed.
That does not make the work disappear; it relocates it. On Shopify, your product data quality still decides your match rate, and your store still has to be reachable and parseable by AI crawlers. We covered the store-level fixes in depth in our guide to Shopify AI search visibility: structured product data, server-rendered prices, and clean titles do for the platform integration exactly what a well-built feed does for a manual one.
The takeaway for platform sellers: confirm your products are eligible and complete inside your admin, then spend your effort on page reachability and data quality rather than file formatting. For self-hosted and custom stores, the manual feed in Step 1 is the path.
Submitting a feed does not tell you whether agents are fetching the pages it points to. That signal lives in your server logs, under AI user agents, and it is the difference between assuming you are visible and knowing it.
Look for OAI-SearchBot and ChatGPT-User in your logs, and note which product URLs they request. Google Analytics filters most of this traffic out by design, so the tool everyone checks shows nothing. A ChatGPT-User hit on a product page in particular means a real shopper's question led ChatGPT to fetch that page live, the clearest sign your feed and your pages are connecting to actual demand.
If a feed item's linked page never shows an agent fetch, that is your diagnostic. Either the page is blocked, the link is wrong, or the product simply isn't matching queries yet. Crawlytics automates this read: it identifies AI crawlers and agents, shows which pages they touch and how often, and trends it over time, so you can see whether ChatGPT is reaching your catalog before purchase data ever shows up. The free Agent-Ready Grader gives you a one-minute baseline on whether your pages are even fetchable and parseable, which is the precondition for everything above. For the deeper picture on what "agent-ready" means once they arrive, our writeup on selling to AI agents end-to-end and the transactability audit both go further than a feed alone.
The order that wins: confirm pages are reachable, build a complete feed, submit and validate it, keep it fresh, then watch the logs to verify agents arrive. Skip the first step and the rest is optimizing a thing the system can't see.
Written by Crawlytics Team. Crawlytics tracks AI bots, generates llms.txt, and powers WebMCP commerce, all from one snippet on any stack. See how it works →
Product placement in ChatGPT Shopping is organic, not paid, so you do not bid for position the way you would with search ads. Whether OpenAI charges any merchant or onboarding fee for the feed program itself is not fully disclosed publicly and has shifted as the program rolls out, so treat any specific dollar figure you see quoted with caution unless it links to OpenAI's own documentation. What is clear: results are ranked by relevance and data quality, not by spend. The reliable way to "pay" for better placement is a more complete, more accurate feed.
There is no publicly confirmed standard timeline, so plan for a range rather than a promise. After you submit a feed over SFTP, OpenAI validates it against a sample batch, and accepted products become eligible to surface; appearing in an actual answer then depends on a shopper asking something your products match. Approval and onboarding windows are not fully documented and have changed as the program rolls out. Anyone quoting an exact day count without citing OpenAI's docs is guessing. The faster route is shipping a clean, complete feed so you clear validation on the first pass.
No. Self-hosted and custom stores get in by building and submitting a product feed directly, the field-by-field path in Step 1. Shopify and Etsy are shortcuts, not requirements: their catalog integrations can surface eligible products through the platform, so you maintain product data in your normal admin instead of generating an SFTP file. Either way, the same fundamentals decide your results, accurate prices and stock, clean titles, real images, and product pages that AI crawlers can actually fetch. The platform changes how the data gets in, not what makes it perform.
The load-bearing fields are id, title, description, link, price, availability, and image_link, with brand, gtin, and product_category sharply improving how well you match specific queries. Title and description matter because they are read as natural language against what shoppers ask. GTIN matters because it disambiguates your product from near-identical ones. Price and availability matter because they decide whether you belong in a buy-now answer and whether your feed agrees with your live page. Completeness beats minimalism: a fully populated record matches more queries than a sparse one that merely validates.
Getting listed in ChatGPT Shopping is about discovery, the product showing up in an answer, which is separate from the agent completing checkout. Whether a purchase finishes inside ChatGPT depends on the payment and checkout integrations in play, and that landscape is moving fast. The safe mental model: the feed gets you found and clicked; completing the transaction is a downstream step that depends on whether your checkout is reachable and parseable by an agent. We cover that side in our pieces on agent transactions and the agentic checkout readiness checklist. Optimize discovery first, then make sure the checkout an agent reaches can actually be completed.
This page is part of Crawlytics.app. View all pages: llms.txt · llms-full.txt