AI Shopping Assistants: The 2026 Brand Playbook for Product Feeds, Schema and Agentic Checkout

Classic e-commerce SEO aims at a ranking on a results page. An AI shopping assistant produces something else: a short answer that names a few specific products, explains why each fits the shopper’s question, and links (or, in some cases, sells) straight to them. The brands that show up in those answers are not doing anything exotic. They ship clean product data to every program the assistants read, keep it in sync with the live store, and give the assistant real text and images to work with. This playbook covers how the assistants choose, what each one actually reads as of October 2026, how to engineer the feed and the schema, and what the agentic checkout protocols mean for a brand that is not a giant retailer.

How AI shopping assistants choose products

No assistant publishes its ranking, but the shape of the process is visible in what they document and in how retrieval-augmented systems generally work. On a query like “running shoes for flat feet under $150,” an assistant typically:

1. Parses the intent into a category, a use case, hard constraints (price, size, ingredient to avoid) and context (gift, replacement, first purchase).
2. Fans out into several sub-searches (query fan-out) against its product index — Google’s Shopping Graph, Amazon’s catalog, the feeds merchants submit to OpenAI or PayPal — plus the open web for reviews and editorial coverage.
3. Filters on the hard constraints. A product whose price, size range or availability the system cannot read cleanly simply drops out.
4. Ranks the survivors on fit to the use case, using descriptions, attributes, review content and ratings.
5. Writes the answer: a handful of products, a sentence of rationale each, images, prices and links.

Three practical consequences follow, and the rest of this page is built on them:

• Structured data is the entry ticket. If your category, attributes, price and stock are missing or inconsistent in the feed, you are not in the candidate set, however good the product page reads.
• Text that names the use case supplies the rationale. The “why this one” sentence has to come from somewhere: your description, your highlights, your FAQ, your reviews, or a roundup that covered you.
• Images are part of the answer. Most assistants now show product images inline, and the shopper decides whether to click (or approve a purchase) while looking at them.

The assistants, one by one (status as of October 2026)

Each assistant gets its product data from a different place, so “optimize for AI shopping” really means getting into several separate programs.

ChatGPT shopping

ChatGPT shopping reads a merchant product feed that OpenAI ingests and indexes, alongside web results. OpenAI publishes the product feed spec with nine required fields for a discovery feed (item_id, title, description, url, brand, seller_name, image_url, availability, price), variant fields, optional shipping, returns and review fields, and two OpenAI flags: is_eligible_search and is_eligible_checkout. The spec says eligibility does not guarantee display, and that many capabilities need setup during onboarding with OpenAI; it also accepts a Google-compatible delimited feed once OpenAI confirms it for your registered feed.

On checkout, the story has already turned once. OpenAI launched Instant Checkout in September 2025 on the Agentic Commerce Protocol it built with Stripe, starting with US Etsy sellers. On March 24, 2026, OpenAI said that first version “did not offer the level of flexibility” it wanted and that it would let merchants use their own checkout while it focuses on product discovery (TechCrunch, CNBC). For a brand that means the ChatGPT job today is mostly discovery: a clean feed and a fast, honest landing page. Larger retailers can also build their own apps inside ChatGPT.

Google AI Mode, AI Overviews and Gemini

Google’s shopping answers draw on the Shopping Graph, which Google describes as more than 50 billion product listings with 2 billion updated every hour (Google, November 2025). Merchants feed it through Merchant Center and through the product structured data on their pages. Google’s own guidance on AI features is blunt: there are no extra requirements or special schema to appear in AI Overviews or AI Mode; pages need to be indexed and snippet-eligible, and Merchant Center and Business Profile data should be up to date. In other words, the work is ordinary feed and SEO hygiene done thoroughly.

Google is also the most active on checkout. In November 2025 it rolled out agentic checkout in the US: a shopper tracks a price, and when it drops, Google can offer to buy the item on the merchant’s site with Google Pay after the shopper confirms. In January 2026 it announced the Universal Commerce Protocol and checkout directly in AI Mode and Gemini (details below), plus Business Agent, a branded chat for retailers inside Search.

Amazon: Rufus, now Alexa for Shopping

Amazon’s assistant launched as Rufus, which Amazon said was trained on its catalog, customer reviews, community Q&A and information from the web (Amazon). In May 2026 Amazon brought Rufus and Alexa+ together as Alexa for Shopping, which compares features, prices and reviews, writes AI overviews on search and product pages, and adds agentic features such as Buy for Me (purchases from other stores on the shopper’s behalf for eligible products) and Auto Buying at a target price.

For brands that sell on Amazon, the inputs are the listing itself: title, bullets, description, the structured attributes in the category template, A+ Content (see Amazon product photography), images, reviews and Q&A. Answers to factual questions (“is it dishwasher safe,” “does it fit a 15-inch laptop”) can only be as good as the attributes you filled in, and reviews will surface what your listing left out, good or bad.

Perplexity

Perplexity answers shopping questions from the open web and from merchant catalogs. In November 2025 PayPal and Perplexity launched Instant Buy for US users: PayPal’s “store sync” makes a merchant’s catalog discoverable in Perplexity, “agent ready” lets the merchant accept payment there with its existing PayPal setup, and the retailer stays merchant of record. Because Perplexity also cites web pages heavily, your own product pages and editorial coverage matter here as much as the catalog.

Microsoft Copilot

For Microsoft Copilot, Microsoft announced Copilot Checkout in January 2026, live first on Copilot.com in the US with PayPal, Shopify and Stripe; Shopify merchants were to be enrolled automatically after an opt-out window, and merchants remain merchant of record. Copilot’s product data comes through Microsoft Merchant Center and Bing’s index, so the same feed and Bing indexing hygiene applies.

Product feed engineering

One well-built source feed, exported in each program’s format, beats five feeds maintained by hand. Treat your product information system (or your Shopify catalog) as the single source of truth, and generate the Google, Microsoft, OpenAI and marketplace feeds from it. Think of the fields in groups, because each group answers a different question the assistant asks.

Identity: is this the same product everywhere?

gtin, mpn and brand let systems match your product across retailers and reviews. Google’s product data specification requires brand for new products, strongly recommends gtin, needs mpn when there is no GTIN, and gives you identifier_exists = no for custom or handmade items that genuinely have neither. Never invent a GTIN to fill the field. Variants need a shared group ID (item_group_id in Google, group_id plus variant_dict in OpenAI’s spec) so a size or color resolves to the right SKU.

Description: written to be extracted

Titles and descriptions now feed a language model rather than a keyword matcher. Write them so any one sentence can be lifted out and still be true: short sentences, one claim each, concrete nouns (“merino wool, 210 gsm, machine washable cold”) instead of adjectives (“premium, luxurious”). Google gives you two fields made for this: product_highlight (2 to 100 short highlights) and product_detail (section, attribute and value triples for specs). Fill the category attributes that distinguish variants — material, pattern, size, color, age_group, gender — and both google_product_category and your own product_type path. These are exactly what an assistant filters on when the shopper says “for a toddler” or “in linen.”

Offer: price, stock, shipping and returns that match the store

Price, availability, shipping cost and return policy are the hard filters. OpenAI’s spec says to update the feed whenever prices or availability change — a sale starting or ending, an item selling out — and notes that date fields do not schedule changes for you. Google’s merchant listing docs warn that a priceValidUntil in the past can stop a listing from showing. The rule underneath both: the feed is a copy of the store, so update it when the store changes, and audit regularly for drift between feed price and checkout price.

Trust: ratings and reviews

OpenAI’s feed takes review_count and star_rating for product reviews only (not store reviews), and says to omit the rating when there are no reviews. On your own pages, review markup must follow Google’s review snippet guidelines: mark up reviews that are visible on the page, from real reviewers. Do not pad either with seeded or AI-written reviews.

Field map across programs

Structured data on your product pages

The feed and the page should say the same thing. Google’s merchant listing structured data requires a Product with name, image and an Offer (not an AggregateOffer) carrying price and priceCurrency. It recommends availability, gtin, mpn, sku, brand, itemCondition, aggregateRating and review, and suggests declaring return and shipping policies once under your Organization markup, using the offer-level hasMerchantReturnPolicy and shippingDetails only for exceptions. Add BreadcrumbList to show the category path.

Three habits prevent most problems:

• Generate markup from the same data as the feed, so a price change cannot update one and not the other.
• Mark up only what is visible. Google asks that structured data match the page text; ratings in markup should come from the review widget the shopper can see.
• Validate after every theme or app change. Shopify apps and theme updates are a common source of duplicate or broken Product blocks. Run the Rich Results Test on a sample of product pages, and watch Merchant Center diagnostics.

For the page-level content rules that help with citations on informational queries, see how to get cited by AI search engines and our AI Overviews guide.

Product images for AI answers

Assistants show images inline, and a shopper weighs a recommendation by its picture. The platforms publish only a few hard rules, and they are worth knowing exactly:

• Google’s main image_link must show the product and must not carry promotional text, watermarks or borders. You can add up to 10 additional_image_link images, which may show staging and the product in use. Google has announced a minimum of 500 × 500 pixels for all product images, enforced from January 31, 2027.
• For page markup, Google recommends several high-resolution images of at least 50K pixels in 16:9, 4:3 and 1:1 aspect ratios, crawlable and indexable.
• OpenAI’s image_url must be a public, direct link to a JPEG or PNG; additional_image_urls takes the rest.
• Merchant Center requires that AI-generated images keep their IPTC metadata (the trainedAlgorithmicMedia digital source type) in image, additional image and lifestyle image links. Check the hosted file, not just the original: compression, background-removal and CDN steps can strip it.

Beyond those rules, our working heuristic is to cover four kinds of shot per product, because they answer the four questions shoppers ask before they buy:

Make the in-use shot match the use cases your copy claims: if the attributes and highlights say “for travel,” show it packed in a carry-on, not on a kitchen counter. Keep a consistent look across the catalog so it reads as one brand.

Where AI imagery fits, and where it must not

The hero is usually a studio photo or a clean cutout of the real product; in-hand, in-use and lifestyle shots are where a shoot per SKU gets expensive, and where AI UGC is practical. ppl.studio’s e-commerce product photos, Shopify product pack and product UGC photos workflows use your uploaded product photo as the prop in each scene, and AI Experts keep the same model across a catalog. The limits matter more than the speed:

• The product in every image must be your product as it ships: same color, shape, label and size. Check every output against the real item, and regenerate or drop anything that drifts. An image that misrepresents the product is a policy problem and a returns problem.
• Never present AI images as customer photos, in review galleries or in “real results” before-and-after claims.
• Keep the AI-generation metadata where a platform requires it, and follow each marketplace’s own image policy (Amazon’s main-image rules are stricter than Google’s).

Content and reviews the assistant can quote

The rationale sentence in an AI answer has to come from text somewhere. Give it good text to use.

• A “Who it’s for” block on each product page that names specific buyers and situations (“for runners with flat feet who want a firm midsole”), plus a “Best for / not for” list. The “not for” line costs you nothing with the wrong buyer and makes the rest more credible.
• A short FAQ per product that answers the questions shoppers actually ask support: sizing, care, compatibility, what’s in the box. Mirror the answers in product_detail and, on Amazon, in the attributes and Q&A.
• Review prompts that ask about the use case. “How did these hold up on long runs?” produces quotable, specific reviews in a way “Leave a review” does not. Filtering reviews by use case on the page helps shoppers too.
• Honest comparison pages (“our model A vs model B,” “us vs the category leader”) with a plain table of specs and who each option suits.
• Editorial coverage. Assistants that cite the web lean on “best X for Y” roundups. Classic product PR — samples, a clear angle, good images — still works, and the result now gets reused in AI answers every time that roundup is cited.

Make sure your brand itself is unambiguous too: consistent Organization markup with sameAs links to your real profiles, the same brand name in every feed, and a claimed Google Business Profile if you have physical locations.

Agentic checkout and the commerce protocols

Agentic checkout is a purchase the assistant completes for the shopper, after the shopper approves, without sending them to browse your site. Every version has the same three parts: a checkout the agent can call programmatically (create a cart, get the final price with shipping and tax, submit the order); a delegated payment credential — a scoped token instead of a raw card number; and order-status updates back to the assistant. In every program described here, the brand stays merchant of record: it keeps the payment processing, the fulfilment and the customer.

Agentic Commerce Protocol (ACP)

The Agentic Commerce Protocol is an open standard developed by Stripe and OpenAI under the Apache 2.0 license (agenticcommerce.dev). It covers the checkout session and passing payment credentials without exposing them, can be implemented over REST or MCP against an existing commerce backend, and launched with ChatGPT as the first AI platform and Stripe’s Shared Payment Token as the first payment method. It was the protocol behind Instant Checkout; after OpenAI’s March 2026 change, it still describes the merchant data ChatGPT shopping runs on.

Agent Payments Protocol (AP2)

Google announced AP2 on September 16, 2025, with more than 60 organizations including Adyen, American Express, Coinbase, Mastercard, PayPal and Worldpay. It solves a narrower problem: proving that a human authorized the agent to spend. It does that with signed “mandates” — an Intent Mandate that records what the shopper asked for (or the rules of a delegated task, such as a price limit) and a Cart Mandate that records the exact items and price approved. It works as an extension of the A2A protocol and MCP and is payment-method agnostic.

Universal Commerce Protocol (UCP)

Google announced UCP on January 11, 2026, co-developed with Shopify, Etsy, Wayfair, Target and Walmart and endorsed by companies including Visa, Mastercard, Stripe and Adyen. It aims to cover discovery through post-purchase and is compatible with A2A, AP2 and MCP. According to Google’s Merchant Center help page, UCP-powered checkout in AI Mode and Gemini is in early access for participating merchants in the US, Canada and Australia; only listings with the native_commerce attribute show the Buy button; shoppers pay with Google Pay; and Merchant Center stays the hub for product data. Google also announced new Merchant Center attributes for conversational commerce, rolling out to a small group of retailers first.

Most brands get there through their platform

You do not need to implement a protocol to be buyable by an agent. Shopify and Etsy catalogs are integrated with ChatGPT shopping, Shopify co-developed UCP and partners on Copilot Checkout, and PayPal merchants can reach Perplexity through store sync. For a brand on one of these platforms, the work is keeping that integration healthy and the catalog data clean. Direct protocol work makes sense for merchants running their own checkout at volume.

The checkout handoff: the feed is a cache

The fragile moment is the checkout handoff, between “the agent picked your SKU” and “the order is confirmed.” At that point the agent needs live answers to three questions: is this exact variant in stock, is the price the one the shopper was shown, and what is the total with shipping and tax. A feed is a snapshot; the checkout is the truth. When they disagree, a careful agent stops rather than surprise the shopper, and you lose a sale you had already won. Make the checkout endpoint fast and authoritative, and keep the feed close enough to it that disagreements are rare.

Expect stricter rules for transactions than for discovery. Age-restricted goods, alcohol, supplements with health claims and financial products may be recommended but excluded from in-assistant purchase, and each program sets its own eligibility; check the rules for every program you join, and use the eligibility flags (is_eligible_checkout, native_commerce) deliberately rather than turning them on across the catalog.

A readiness checklist, in order

1. Audit the data you already have. Export your catalog and check every product for GTIN or MPN, brand, category, variant grouping, price, stock, shipping and return policy. Fix gaps at the source, not in one feed.
2. Join the programs. Google Merchant Center and Microsoft Merchant Center first; then OpenAI’s merchant program if you are not on Shopify or Etsy; PayPal store sync if you take PayPal; Amazon listing clean-up if you sell there.
3. Match page markup to the feed. Product and Offer markup generated from the same data, validated on a sample of pages.
4. Rewrite descriptions for extraction. Start with your best sellers: concrete specs, highlights, “who it’s for,” a product FAQ.
5. Fill the image set. A compliant hero plus in-hand, in-use and detail shots for each product, with AI metadata preserved where you use generated images.
6. Fix review prompts. Ask about the use case after purchase; show real reviews on the page and mark them up honestly.
7. Close feed-to-checkout drift. Update the feed when price or stock changes, and spot-check feed values against the live checkout.
8. Decide on checkout per program. Turn on platform integrations (Shopify, PayPal) where they exist; consider UCP or ACP directly only if you run your own checkout at scale.
9. Measure monthly, change one thing at a time. See below.

Measuring AI shopping visibility

• A fixed prompt panel. Write 20 to 50 questions a real buyer in your category would ask, run them in each assistant on a schedule, and record which products are named and what reason is given. It is crude, but it shows whether your changes land. Third-party AI visibility tools automate this (AI share of voice).
• Referral traffic. In GA4, group referrers such as chatgpt.com, perplexity.ai, copilot.microsoft.com and gemini.google.com into an AI assistants channel. Google’s own AI Overviews and AI Mode traffic is reported inside Search Console’s Performance report under the Web search type, not separately.
• Program diagnostics. Merchant Center disapprovals, OpenAI feed validation errors and Amazon listing quality warnings are the earliest sign that a product dropped out of the candidate set.

Frequently Asked Questions