How to Get Your Products Recommended by ChatGPT Shopping

The March Reset Nobody Updated Their Advice For

Most guides to ChatGPT commerce were written for a world that lasted five months. The timeline matters, because it tells you where the durable asset is.

Shopping arrived in ChatGPT on April 28, 2025, for every tier including logged-out users, with OpenAI stating plainly that product results are not ads and are chosen independently. In September 2025, OpenAI and Stripe published the Agentic Commerce Protocol and Instant Checkout let US shoppers buy without leaving the chat, with Shopify merchants reportedly paying a 4% fee on those sales.

Then reality reported in. On March 4, 2026, OpenAI pulled back Instant Checkout, saying it is transitioning to apps. Shopify's president said only around a dozen of its millions of merchants were actively selling through AI agents. Retailers who tried it told a consistent story: Walmart saw roughly 3x worse conversion in-chat than when shoppers clicked out to its own site. Buying inside the chat was the part nobody wanted. Three weeks later, Shopify switched on Agentic Storefronts by default for eligible stores: discovery happens in ChatGPT, and the purchase lands on the merchant's own store, with no fees beyond standard processing.

Read the whole sequence and the lesson is unambiguous. The transaction layer keeps moving: in-chat, then apps, then your own checkout, and on Google's side a parallel standard, the Universal Commerce Protocol, arrived in January 2026. What never moved is where the decision gets made. McKinsey sizes agentic commerce at $3 trillion to $5 trillion orchestrated globally by 2030, and "orchestrated" is the operative word: value the agent influences, whoever runs the checkout. Discovery is the durable asset. That is what the four surfaces below are for.

The Four Surfaces

When someone says "my product shows up in ChatGPT," they could mean four different things, produced by four different retrieval paths. Everything else in this post hangs off this table.

The surfaces share inputs, which is why they get conflated. Clean product data helps all four. But the lever that moves the carousel barely touches the organic answer, and the work that wins organic answers is invisible to the feed pipeline. Budget accordingly.

Surface 1: The Shopping Carousel

The carousel is the row of product cards, with images, prices, and ratings, that ChatGPT renders inside a shopping answer. It is the surface most guides mean, and the one with the most concrete, documented inputs.

An application, not an upload

The single most repeated piece of advice, "submit your product feed to OpenAI," skips the part where you cannot simply do that. Feed onboarding is available to approved partners who apply through OpenAI's merchant form and clear a prohibited-products policy. If you sell on Shopify you likely do not need to: eligible stores are syndicated by default through Agentic Storefronts. And carousel products can also arrive through crawled structured metadata with no feed at all, which is how the surface worked at launch.

The fields that carry signal

OpenAI's product feed spec requires the basics: id, title, description, URL, brand, image, price, availability, and two flags, one of which, is_eligible_search, decides whether the product can be surfaced at all. The interesting fields are optional: star_rating, review_count, full review content, Q&A pairs, popularity_score, and return_rate. OpenAI is explicitly willing to ingest performance and reputation signals from merchants. If your feed sends none of them, you are asking to be ranked on the thinnest possible record.

What OpenAI actually says about ranking

One sentence. Recommended attributes, "like rich media, reviews, and performance signals," improve ranking, relevance, and user trust. That is the entire official disclosure. Nobody outside OpenAI has the weightings, so anyone quoting you a percentage lift per field invented it. Operationally, the docs are clearer: full snapshots at least daily over SFTP, intraday price and availability changes through the API. Stale price and availability data is the one failure mode entirely within your control.

Surface 2: Organic Answer Mentions

Ask ChatGPT "what running shoes should I buy for flat feet" in an ordinary conversation and you often get prose, not cards: a few recommended models with reasoning, sometimes with source links. No feed controls this surface. It is assembled from live web search plus what the model already believes, which makes it the surface where the classic visibility playbook applies.

The crawler that decides

OpenAI is unusually specific here: "OAI-SearchBot is used to surface websites in search results in ChatGPT's search features," per its bots documentation. Block it and you have removed your own pages from consideration; GPTBot, the training crawler, is a separate decision. Changes to robots.txt take roughly 24 hours to register. This is a two-minute check that a surprising number of retailers fail, usually because a CDN default made the decision for them.

Consensus, not copy

When OpenAI described how shopping selections work, the language was about the web's opinion of you: understanding how people are reviewing this, what the pros and cons are. Your product page states a claim; the answer is built from whether reviews, comparisons, and category lists agree with it. That is off-site work: review volume on the platforms your category trusts, presence on the comparison pages the engine keeps citing, and product pages whose facts survive being read without JavaScript. We covered the on-site half in the e-commerce playbook and the engine's general behavior in the ChatGPT spoke.

Ads are a different surface, on purpose

Since February 2026, US users on the Free and Go tiers see labeled ads at the bottom of some answers. Ads are bought; the answer above them is not. Treating ad placement as a substitute for organic presence misreads the layout: the recommendation shoppers actually asked for still cannot be purchased from OpenAI.

Surface 3: Shopping Research

Launched November 24, 2025 for all logged-in tiers, shopping research is the long-form variant: ChatGPT interviews the shopper, then assembles a structured buyer's guide, built on a model post-trained for shopping from GPT-5-Thinking-mini. OpenAI says results are organic, drawn from public retail sites, and that merchants who want visibility go through an allowlisting process.

The practical implication is about depth rather than presence. A research report compares specs, surfaces trade-offs, and answers objections, which means it rewards product pages that contain spec-level detail, honest comparison context, and real Q&A. A thin page with a hero image and three adjectives gives the model nothing to build your section of the report from. This is the surface where the difference between a catalog entry and a genuinely informative product page pays most directly.

Surface 4: Checkout and Storefronts

The buy path is where 2026 has been most chaotic, and where the merchant's job is mostly eligibility hygiene rather than optimization.

Shopify: on by default

Since March 24, 2026, eligible Shopify stores are discoverable in ChatGPT automatically. The eligibility list is mundane and checkable in your admin: products eligible for Shopify Catalog, completed terms, privacy, and return policies, agreement to the supplemental terms, selling to US customers, and any legal disclosures within the first 6,000 characters of the product description. Shopify's own product-data guidance for these surfaces is worth taking literally: group real variants under one parent, use the most specific category available, keep marketing copy out of data fields, and assume nothing rendered by JavaScript gets read. On desktop, the purchase lands on your own store in a new tab, which means your checkout, not a protocol, is the conversion surface.

The protocols underneath

Two standards now share the layer. ACP, from OpenAI and Stripe, handles delegated checkout where it still exists; its production FAQ is blunt that the merchant remains merchant of record and owns refunds and chargebacks. UCP, from Google and Shopify with Etsy, Wayfair, Target, and Walmart co-developing, does discovery-to-payment on Google surfaces; a merchant's concrete artifact is a capability manifest at /.well-known/ucp, and UCP-powered checkout on AI Mode and Gemini is currently select merchants, three countries, gated on a Merchant Center attribute and Google Pay. Perplexity runs the same play with PayPal, which keeps the retailer as merchant of record. The pattern across all three: platforms want the transaction to happen, and none of them wants to own your customer service.

What You Can Verify Today

This category has a folklore problem: ranking factors nobody outside the platform can know, sync cadences that contradict the documentation, and lift percentages with no study behind them. The antidote is to sort claims by whether you can check them yourself.

The machine-readability check deserves urgency: Adobe found roughly a third of product pages cannot be properly read by AI at all. You can test your own in under a minute with our free AI crawler inspector, which fetches a URL exactly as 13 AI bots see it.

Measuring It Honestly

Two facts make shopping-surface measurement harder than most guides admit, and both are structural rather than fixable.

First, the answers are probabilistic. The same shopping prompt does not return the same products every time, so a screenshot of your product in a carousel is one draw from a distribution, not a ranking. The only measurement that means anything is an appearance rate: the same prompts, sampled on a schedule, tracked as a series. We have written about why one measurement misleads; shopping surfaces are the extreme case.

Second, the traffic is partially dark. GA4 added an AI Assistant channel group in May 2026, which helps, but it can only classify sessions that arrive with a referrer, and Google has not published the full list it recognizes; referrer-less visits from apps and in-app browsers still land in Direct. Visibility and measured referral traffic have decoupled, a problem we unpacked in the dark AI traffic post. The honest setup tracks both independently: appearance rate on the answer side, and AI-referred sessions as a floor, never a total, on the analytics side.

How Ranqo Helps

Ranqo measures the surface that drives the others: the answer layer. It runs your buyer prompts against ChatGPT and five other engines on a weekly cadence, records whether your brand and products are named, at what position, with what sentiment, and which sources the answer leaned on, so "are we recommended for this" becomes a trend line instead of a screenshot. The page audit checks the machine-readability problems this post keeps returning to, from crawler access to content that only exists in JavaScript.

The carousel and shopping research are app surfaces without a stable public API, so no tool can sample them the way the conversational layer can be sampled; the off-site consensus that feeds them, and the answers themselves, are measurable today, and that is where we point the instrument.