How to get your products into ChatGPT Shopping

ChatGPT Shopping went from announcement to functioning sales channel faster than any commerce surface in recent memory. Instant Checkout launched on September 29, 2025 for US users across Free, Plus and Pro plans, with US Etsy sellers live on day one and OpenAI announcing more than a million Shopify merchants to follow, names like Glossier, Vuori, Spanx and SKIMS among them. The market's read on the stakes was immediate: CNBC reported Etsy's stock rose 16 percent the day of the announcement.

The traffic data since then justifies the reaction. Adobe Analytics, analyzing over a trillion visits to US retail sites, measured AI-driven traffic up 393 percent year over year in Q1 2026, and by March 2026 those AI-referred shoppers were converting better than visitors from traditional channels while producing roughly 37 percent more revenue per visit. A shopper who arrives from a ChatGPT recommendation has already done the comparison work inside the conversation. Your job is to be the product that conversation lands on.

This playbook covers the full stack in order: crawler access, the product feed, Instant Checkout, and then the layer most merchants skip, which is winning the recommendation itself. The steps are sequenced deliberately. Feed plumbing without answer visibility produces a beautifully integrated catalog nobody gets shown. If you want the broader context on where this surface fits, start with our agentic commerce primer and come back.

Before feeds and checkout, ChatGPT has to be able to see your site. OpenAI's search crawler, OAI-SearchBot, fetches pages that ChatGPT cites and links in shopping and search answers. If your robots.txt blocks it, whether deliberately or through a blanket rule a previous agency added, you have removed yourself from consideration for every answer that relies on live retrieval. Check the file today; it is a two-minute audit with channel-level consequences.

While you are in there, distinguish between OpenAI's crawlers. OAI-SearchBot powers search and shopping retrieval, GPTBot gathers training data, and ChatGPT-User fetches pages a user asks about in real time. Some brands block training crawlers while allowing search crawlers, which is a coherent policy. Blocking all three is how a brand disappears from the channel entirely. The full taxonomy, including the other engines' crawlers, is documented in AI crawlers explained.

Crawlability also means renderability. Adobe's AI Content Visibility Checker found retail product detail pages scored just 66 percent on machine readability on average, the lowest of any retail page type, largely because critical content hides behind JavaScript, tabs and interactive widgets that crawlers never execute. A crawler that reaches your page but cannot parse the price, specs or availability leaves with nothing usable.

The product feed is how ChatGPT Shopping knows your catalog authoritatively rather than by scraping. There are two routes in. Direct integration runs through OpenAI's merchant program at chatgpt.com/merchants: you apply, get approved, and receive delivery details for pushing feed files to OpenAI. The platform route is simpler for most brands: Shopify and other commerce platforms syndicate merchant catalogs into the program, which is how a million-plus merchants connected without individual integrations.

The spec itself, published in OpenAI's Agentic Commerce documentation, accepts JSONL, CSV or TSV compressed with gzip, or Parquet with zstd, all UTF-8. It is a push model rather than a crawl: you deliver files to the endpoint provided at onboarding, and you can update as often as every 15 minutes. That cadence is a hint about what matters. OpenAI evaluates merchants on availability, price, quality and whether they are the maker or primary seller. Stale inventory and mismatched pricing are ranking problems, which makes feed operations a growth function rather than a back-office chore.

Content quality inside the feed deserves equal attention. The model matches products to conversational intents, so a title like "Trail Runner Pro 2, waterproof trail running shoe for rocky terrain, wide toe box" gives it far more to match against than a SKU-style name. Fill every applicable attribute field: materials, dimensions, compatibility, care. Every blank attribute is a buyer constraint you can never satisfy, because the agent filters on data it has, and absence reads as failure.

Appearing in ChatGPT Shopping and transacting inside it are separate commitments. With visibility alone, the buyer taps through to your site and checks out there, and you keep your full funnel, upsells included. With Instant Checkout, the purchase completes inside the conversation through the Agentic Commerce Protocol, the open standard Stripe and OpenAI co-developed. Payment flows through a shared payment token, so the agent never handles raw card credentials, and you remain the merchant of record: your fulfillment, your customer relationship, your returns process.

Enabling it is explicit in the feed: set enable_checkout to true, and supply the policy URLs the spec requires so buyers can read shipping, returns and terms without leaving ChatGPT. The launch scope was single-item purchases from US sellers, with multi-item carts and more regions on OpenAI's roadmap, so model the channel as high-intent single-product orders for now. Where ACP sits relative to Google's UCP and the payment trust protocols underneath is mapped in our protocol explainer.

Should you enable it? If your margin tolerates another transaction path and your operations can honor in-chat promises about shipping and returns, yes. Friction removed at the moment of highest intent is rarely a bad trade. But it is step three for a reason: checkout only fires after your product is the one recommended, which brings us to the layer that actually decides revenue.

Here is the uncomfortable arithmetic. Thousands of merchants can have flawless feeds and checkout enabled, and a given answer still shows a handful of products. The feed gets you into the pool; the recommendation is won elsewhere. OpenAI is explicit that shopping results are organic and unsponsored, ranked on relevance, which means there is no budget line that buys placement. There is only the work of being the product the model concludes is the right answer.

That conclusion draws on more than your feed. For research-shaped questions, the kind that begin "what is the best," the model retrieves and weighs web sources: buying guides, comparison pages, review sites, community threads. If those sources consistently name your competitors and omit you, the answer inherits the omission no matter how clean your feed is. This is the same third-party battlefield we documented for brand recommendations in how ChatGPT recommends brands, now with a buy button attached to the outcome.

The work, concretely: identify the buying questions in your category, check which products the answers name and which sources they cite, then close the gaps. Publish answer-shaped comparison and use-case content on your own domain. Earn honest presence on the review sites and communities the engines keep citing. Keep review volume fresh, since quality signals feed the ranking. Reachroller compresses this loop into a product: it tracks your buying questions on a schedule, stores every raw answer and citation as receipts, scores your visibility honestly with branded questions excluded, and generates the publish-ready fix page for each question you lose.

Treat ChatGPT Shopping with the same measurement discipline as paid search or email. Tag and segment AI referral traffic in analytics so you can see conversion and revenue per visit for the channel; Adobe's data suggests you may be pleasantly surprised. Watch for ChatGPT-sourced orders once Instant Checkout is live. And track your recommendation presence over time, because it moves.

It moves because AI answers are probabilistic. The same question, asked twice, can surface different products in a different order. A single manual check that happens to show your product proves almost nothing, and a single absence proves equally little. The honest metric is a mention rate across repeated runs with the raw answers stored for audit, which is exactly how Reachroller reports it: every score links to the actual answers behind it, so you are never celebrating a coin flip. Why answers vary this much is covered in why AI gives a different answer every time.

Run the loop weekly: check the questions, note the losses, publish the fix, get it indexed, recheck. Merchants doing this now are compounding presence in a channel most competitors have not started measuring. The window where that is true will not stay open long.

The blanket crawler block. Somewhere in 2023, an agency added a rule denying all AI bots to protect content from training scrapes, and nobody has looked at robots.txt since. The rule now also blocks the retrieval crawler that decides whether your products can appear in answers. Audit the file, and separate your training-data policy from your search-visibility policy deliberately.

The set-and-forget feed. A feed uploaded once and refreshed quarterly drifts out of sync with reality within days, and OpenAI ranks on availability and price accuracy. Wire the feed to your inventory system, use the 15-minute update window the spec allows for volatile fields, and monitor rejects the way you monitor failed payments.

SEO instincts applied verbatim. Keyword-stuffed titles and thin doorway descriptions carry over badly: the Princeton GEO research measured keyword stuffing performing below baseline in generative answers. The model matches meaning rather than strings, so a title written for a human deciding between two products outperforms a title written for a 2019 crawler. The habits worth keeping and dropping are mapped in GEO vs SEO.

Declaring victory on one screenshot. A founder asks ChatGPT about their category, sees their product, and moves on. Probabilistic answers make single observations nearly worthless in both directions. Decide with mention rates over repeated runs or do not decide at all.

Days 1 to 7: audit. Check robots.txt for OAI-SearchBot. Load your top ten product pages with JavaScript disabled and note what disappears. Ask ChatGPT the ten buying questions that most plausibly precede your purchases, three times each, and record which products and sources appear. This baseline costs a day of focused work and reframes every decision that follows.

Days 8 to 15: plumbing. Connect the feed through your platform or apply at chatgpt.com/merchants. Fix the attribute gaps the audit exposed, starting with availability, price and identifiers. Rewrite your top twenty titles and descriptions for intent matching.

Days 16 to 30: the visibility loop. For each buying question where competitors get named and you do not, publish one answer-shaped page and earn or refresh one third-party mention. Get the pages indexed, then recheck the questions. From here the work is a weekly rhythm rather than a project, which is the point: Reachroller automates the tracking, the receipts and the fix drafts so the rhythm survives contact with a founder's calendar.