AI shopping answers
ChatGPT Shopping Optimization: How AI Picks the Products It Shows
Ask ChatGPT for a carry-on under $200 that fits a small regional jet and it hands back four bags, a photo of each, a star rating and a row of stores. Yours was not among them. Nobody rejected it. The machine never had enough facts to consider it.
October 29, 2026 · 12 min read

The short answer
ChatGPT shopping optimization means supplying the product facts AI shopping answers choose from. ChatGPT builds product results from structured metadata such as prices, descriptions and reviews, plus a feed program for approved merchants. Google AI Mode shops from its Shopping Graph, filled by Merchant Center. Matching feeds, crawlable product pages, real reviews and complete attributes raise the odds.
On this page
- The short shelf that replaced the results page
- From one sentence to a shortlist: the stages in between
- OpenAI's feed spec, read as a wish list
- Merchant Center's newest fields were written for AI Mode
- Product stars, store stars and the GTIN that joins them
- Product markup is the page's sworn statement, not a lever
- Testing your catalog in ChatGPT and AI Mode
- A repair order for a store with more SKUs than hours
The short shelf that replaced the results page
A shopping answer is a shelf with room for a handful of products. ChatGPT got one in April 2025, when TechCrunch reported that ChatGPT search would answer product questions with a few recommendations, images, reviews and links to buy from retailers' pages, starting with fashion, beauty, home goods and electronics. OpenAI said the picks were made independently and were not ads. In late November 2025 it added shopping research, which Retail Dive describes as a buyer's guide built after ChatGPT asks follow-up questions about budget and other criteria.
Google's shelf lives in AI Mode. Its May 20, 2025 announcement paired Gemini with the Shopping Graph, whose listings carry details such as reviews, prices, color options and availability; Google put the count above 60 billion listings in May 2026. Ask for a travel bag, mention a trip to Portland in May, and AI Mode runs several searches at once and starts suggesting waterproof bags with easy pockets. That habit is query fan-out. The ChatGPT and AI Mode guides cover the rest of each engine.
From one sentence to a shortlist: the stages in between
Neither company publishes a ranking formula, and anyone who claims to have one is guessing in a nice font. What they do publish describes a pipeline clear enough to work with. One stage below is our reading, and marked so.
| What happens | What it reads | Source | |
|---|---|---|---|
| 1. The request becomes conditions | Budget, size, use and occasion turn into requirements | The shopper's words; signed-in ChatGPT memory can add preferences on Plus and Pro | Google; TechCrunch |
| 2. The candidate pool forms | Only products the engine holds data on can be considered at all | Shopping Graph listings from Merchant Center; for ChatGPT, third-party metadata, approved feeds and crawled pages | Google's AI optimization guide; TechCrunch; OpenAI's commerce docs |
| 3. Hard filters apply | Products that cannot confirm the stated price, size or delivery need fall away | Price, stock, variants, shipping cost, return terms | Our reading of how conditional questions behave, not a published rule |
| 4. Survivors are compared | The shortlist is weighed on specs and on what buyers say | Attributes, ratings, review text, comparisons elsewhere on the web | TechCrunch on ChatGPT; Google's AI optimization guide |
| 5. The product meets a store | Each pick is matched to sellers with a price and a link | Merchant feeds and product pages | Both engines link to retailers; neither fully publishes how stores are ordered |
Stage three is where small stores lose without knowing it. A shopper who types 'under $150, here by Friday, free returns' has written three filters. A product whose data leaves delivery times and return terms blank cannot confirm the last two. OpenAI's own feed specification says an omitted shipping price means unknown, not free. A machine that cannot confirm a condition has an easy alternative: a product that confirms it.
An AI shopping answer is not persuaded. It is satisfied, one checked box at a time.
OpenAI's feed spec, read as a wish list
OpenAI's commerce getting-started guide says feed onboarding is limited to approved partners, who apply through chatgpt.com/merchants, and recommends a full file once a day with API updates in between. Most small stores will not be in that program soon. Read the feed specification anyway. It is OpenAI's plainest statement of what ChatGPT wants to know about a product, and each field mirrors something your product page should say.
| Field | Status | What it tells a store |
|---|---|---|
| item_id, title, description, url, brand, seller_name, image_url, availability, price | Required | The minimum for discovery. Titles up to 150 characters and descriptions up to 5,000, in plain text, with direct image URLs anyone can open. |
| review_count, star_rating | Optional | Product reviews only. The spec says to leave seller and store reviews out of these two. |
| store_review_count, store_star_rating | Optional, separate setup | The store's own reputation, kept apart from the product's. |
| shipping_price | Optional | Left empty, it reads as unknown, not free. |
| accepts_returns, return_deadline_in_days | Optional | Returns stated as data. A policy URL alone does not set the flag. |
| gtin, mpn | Optional | The identifiers that let one product's facts line up across sources. |
| is_eligible_search | On unless switched off | The spec adds that eligibility does not guarantee display. |
The spec also accepts a Google-compatible product file once OpenAI confirms it for a registered feed, so a store with a tidy Merchant Center feed starts ahead. OpenAI's program excludes adult content, age-restricted goods and prescription-only medicines.
Stores outside the program have two other doors. One is the platform: Shopify says its Agentic Storefronts channel is enabled automatically for eligible stores and places products in ChatGPT, Gemini and other assistants (see Shopify stores). The other is the open web. OpenAI says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers, so check robots.txt against the AI crawlers table first.
Merchant Center's newest fields were written for AI Mode
Google's AI optimization guide says Merchant Center feeds help products appear in its AI responses. When Google launched the Universal Commerce Protocol in January 2026, it also promised feed attributes built for conversational shopping, naming answers to common product questions, compatible accessories and substitutes. They now exist. Google's help page on conversational attributes lists six, and the individual pages say they are mainly meant for experiences such as AI Mode.
- question_and_answer: up to 30 pairs per product, each side up to 1,000 characters, from users, the merchant or the manufacturer. Product facts only: no prices, dates or keyword lists.
- document_link: up to five PDF links per product, such as manuals, assembly guides and package inserts, each under 50 MB, public, stable and open to Googlebot.
- related_product: up to 30 links to other items, each labeled as an accessory, a required part, a substitute, part of a set, often bought with, or a different brand.
- item_group_title and variant_option: a shared name for a product family, and the size or color that sets each variant apart.
- popularity_rank: where a product sits in your own catalog by popularity, as a percentage of inventory.
All six are optional, do not affect product approval, and are best added through a supplemental data source, Google says. Google does not say they change rankings, and neither will we. The case is simpler. When a shopper asks AI Mode whether a tent sleeps two adults and a retriever, the answer comes from somewhere, and a pair you wrote beats a stranger's forum guess.
Product stars, store stars and the GTIN that joins them
Both engines ask two questions of reviews: is the product good, and is the store safe. OpenAI keeps the counts in separate fields, Google in separate programs. Its Product Ratings program shows aggregated stars in ads and free listings once a store has at least 50 reviews across its products, compiled from merchants, review aggregators, review sites and Google users. Reviews arrive through a supported aggregator, a reviews feed in Merchant Center, or the free Google Customer Reviews program.
The detail that decides whether any of it counts is the identifier. Google matches reviews to products mainly by GTIN, and says fallbacks such as SKU, brand plus MPN or the product URL usually match fewer. Leave GTINs out of the feed and reviews may never attach to the product they praise. Our practice, not either company's rule: own-brand stores should register GTINs too.
- Ask every buyer after delivery, and never trade a discount for a sentiment; the FTC's review rule bans it.
- Show review text on the product page in the HTML, not only inside a widget that loads later.
- Reply to complaints in public. How engines weigh replies is covered in reviews and AI recommendations.
Product markup is the page's sworn statement, not a lever
Google's checklist for AI features asks owners to keep Merchant Center current and to make structured data match the visible text, and says no special markup exists for its AI features. So Product markup is a signed copy of facts already on the page. When the feed says $129 and the page says $139, an engine has two witnesses and a reason to trust neither. The fields are covered in product page SEO; the AI angle sits in structured data for AI.
For ChatGPT, where the facts sit matters more than which schema type holds them. Vercel's December 2024 analysis found OpenAI's crawlers did not execute JavaScript, so a price, a stock badge or a block of reviews injected by a script after load may never reach the crawler behind ChatGPT search. Open a product page's source and search for the price. JavaScript and AI crawlers walks through the test.
Stated by Google or OpenAI
- Merchant Center feeds help products appear in Google's AI responses
- OpenAI's feed requires price, availability and a seller name on every product
- Blocking OAI-SearchBot keeps a site out of ChatGPT search answers
- ChatGPT's organic product picks are not ads, by OpenAI's account
Sold without evidence
- A special 'AI product schema' that lifts items into answers
- An llms.txt file listing the catalog for chatbots
- A guaranteed spot in the ChatGPT carousel
- FAQ markup as a shopping signal; FAQ rich results ended May 7, 2026
Testing your catalog in ChatGPT and AI Mode
- 1
Write questions with conditions attached
Twenty buying questions, each with a budget, size, use or deadline: 'rain jacket for bike commuting under $150', not 'rain jackets'.
- 2
Ask ChatGPT in fresh chats
Signed out and signed in, since memory shapes signed-in results. Note the products, stores and sources.
- 3
Ask AI Mode, then narrow it
Add a second condition, as shoppers do, and watch which products survive.
- 4
Open Merchant Center's AI performance tab
Under Analytics, then Products, the AI performance report shows your share of voice against your competitor set for English shopping queries in AI Mode and AI Overviews, grouped into Discovery, Evaluation and Ready to buy, and lists popular attributes that may be missing from your data.
- 5
Run the free check
Our AI visibility check puts a shopper's question to ChatGPT and other engines through their APIs and shows which stores were named. API answers can differ a little from the apps.
- 6
Find the visits
ChatGPT citation links often carry utm_source=chatgpt.com and land in GA4's AI Assistant channel; AI Mode clicks stay in Organic Search. The GA4 AI traffic guide has the setup.
Log the same questions monthly; one screenshot proves nothing about an engine that answers differently each run. Prompt tracking covers how many, and AI share of voice turns the log into a number. Merchant Center's report counts organic traffic only and runs a few days behind.
A repair order for a store with more SKUs than hours
- 1.Let the readers in. Allow OAI-SearchBot, Googlebot and Bingbot, and keep a Microsoft Merchant Center feed for Copilot.
- 2.Make three sources agree. Feed, markup and page show one price, one stock status, one shipping cost and one return window.
- 3.Fill the identifiers. GTIN, brand and MPN on every product that has them.
- 4.Turn policies into numbers. Shipping cost by region and return days, in the feed and on the page.
- 5.Answer what buyers email about. Q&A pairs and PDF manuals for the twenty most-asked-about products.
- 6.Grow reviews honestly, and join Product Ratings once you pass 50.
- 7.Then switch on the channels. Apply to OpenAI's merchant program or enable your platform's AI channel.
Checkout comes last on purpose. ChatGPT's Instant Checkout launched in September 2025, and press reports say OpenAI told merchants in March 2026 they can use their own checkout while it concentrates on discovery. Google's protocol checkout covers eligible US retailers. A buy button only appears on a product an engine already chose; the agentic search page follows the protocols, and the GEO hub maps every engine.
A small catalog with a clean Merchant Center account can do all of this alone over a few weekends. Help pays when the catalog runs to thousands of SKUs, the platform hides prices behind scripts, or the same three rivals keep taking the shelf. At DataCram, GEO for stores sits inside our monthly SEO programs. Programs start at $575 a month, depending on complexity and setup, there is no long contract, and the tier is set after a free GEO audit. Stores that need rebuilding start at $2,500, fixed, through e-commerce development. See GEO pricing, GEO for e-commerce and the e-commerce SEO guide.
Questions owners ask us about AI search
Make sure OAI-SearchBot can crawl your store, because OpenAI says sites that opt out will not appear in ChatGPT search answers. Put price, stock, shipping and returns in the page HTML, keep reviews visible, and give products their GTINs. Then apply to OpenAI's merchant feed program or turn on your platform's AI channel, such as Shopify's Agentic Storefronts. Nobody can promise placement; complete data raises the odds.
Not in the organic picks. OpenAI said when shopping results launched that products are chosen independently and are not ads. Press reports say it began testing ads in the US in February 2026 for Free and Go users, labeled sponsored and placed below the answer, and says ads do not influence ChatGPT's answers. The organic shelf is earned with data, which is what a GEO program works on.
Usually one well-kept catalog can serve both. Google reads Merchant Center. OpenAI's feed specification accepts a Google-compatible product data file once OpenAI confirms it for a registered feed, though onboarding is limited to approved partners. Microsoft says Microsoft Merchant Center feeds inform Copilot's shopping results. Many platforms, Shopify among them, syndicate the catalog to AI channels for you.
OpenAI has not published guidance on schema for shopping. Reporting on the launch says product results draw on structured metadata from third parties, such as prices, descriptions and reviews, and OpenAI's crawlers were found not to run JavaScript. So keep accurate Product markup in the server-rendered HTML, matching the visible page, and treat it as supporting evidence rather than a switch.
Your brand's share of AI impressions for shopping queries in AI Mode and AI Overviews, compared with a competitor set you define, plus how many of your products appear for top terms and which popular attributes your data lacks. It covers English queries for accounts in the US, Canada, Australia, India and New Zealand, counts organic traffic only, and updates daily with a few days' lag.
It is no longer the center of it. Instant Checkout launched in September 2025 with US Etsy sellers, and press reports say OpenAI announced in March 2026 that merchants can use their own checkout while OpenAI focuses on product discovery. For most stores the question that matters is whether ChatGPT picks the product, not where the payment happens.
It depends on the door. OpenAI says robots.txt changes reach its search systems in about 24 hours and recommends a full feed upload daily. Google said in 2025 that more than 2 billion Shopping Graph listings are refreshed every hour. Product Ratings can take up to two weeks to appear after onboarding. Whether the answers change is another matter, so test monthly.
Keep reading
- Retail and e-commerce GEOGEO for E-commerce: Getting Your Products Named in AI Shopping AnswersHow online stores get recommended by ChatGPT, Google AI Mode, Gemini, Copilot and Perplexity: feeds, product markup, reviews, rules and a 30-day plan.
- Product pagesProduct Page SEO: Descriptions, Product Markup and Reviews That Rank and SellHow to write product descriptions that rank and sell, add Google's Product markup, handle variants and reviews, and use AI without a catalog of twins.
- Agents and protocolsAgentic Search: When AI Agents Browse, Compare and Book for Your CustomersAgentic search explained: what AI agents already book and buy, the protocols behind them (MCP, UCP, ACP, AP2), and what your business must make readable.
- AI as your SEO assistantChatGPT for SEO: Prompts That Save Hours, and the Answers You Have to CheckHow to use ChatGPT for SEO: prompts for keyword ideas, outlines, title tags, schema and Search Console data, plus the numbers and facts it makes up.
- The math behind the dialAI Share of Voice: How to Measure It Without Fooling YourselfHow to calculate AI share of voice from a prompt set, which of three formulas to use, how many answers you need, and why one run of ChatGPT proves nothing.
- Engine guide: AI OverviewsHow to Get Cited in Google AI Overviews, the Summary That Sits Above Your RankingWhat decides which pages Google AI Overviews cite, what Google says does not matter, how to measure it in Search Console, and how to opt out if you must.
- Everything in Generative Engine OptimizationSee the section
Find out why the AI shelf skips your products
Send us your store and the questions your buyers ask. We run them through ChatGPT and AI Mode, check your feed, markup and reviews against what the engines read, and send a fix list with a fixed monthly quote.
Which of these sounds like you? (pick any)
Rather talk now? Call (210) 346-0848.