How do AI shopping agents work, and how does your store get into them?

By Andrej Ruckij · October 30, 2026 · updated November 1, 2026

TL;DR: An AI shopping agent turns a request (“running shoes under $120 for flat feet”) into a shortlist from product feeds, compares price, stock, reviews and policies, and then either sends the buyer to your checkout or completes the order through a protocol where you stay the seller. In November 2026 that runs mostly on feeds: Shopify stores are in ChatGPT, Copilot and Google’s AI surfaces by default, Google’s in-chat Buy button needs a separate integration, and Amazon keeps its own Alexa agent and blocks outside ones. The volume is still small. In a study of 973 ecommerce sites covering August 2024 to July 2025, before ChatGPT’s shopping checkout launched, ChatGPT sent under 0.2% of their traffic, and OpenAI pulled back from checkout inside ChatGPT in March. My read: the work that pays now is clean product data, reviews and measurement, split into an Amazon plan and a plan for everyone else. Checkout integrations can wait unless you’re on Shopify, where they’re already on and cost nothing extra.

Primores helps brands get their products found in AI search and shopping answers: see how we work with brands.

What changed

On Amazon’s earnings call on October 29, 2026, Andy Jassy said few customers buy through outside AI shopping agents because the experience is “still not great, nor more convenient than coming to a retailer directly,” Fortune reports. A month earlier Amazon blocked Meta’s Muse agent from its store, saying apps that buy for customers “should operate openly and respect service provider decisions” about whether to take part. Shopify, Walmart and Best Buy went the other way and welcomed Muse. Amazon’s own agent is big: over 400 million customers used Alexa for shopping in the past year, interactions grew 570%, and US customers who use it spend 50% more per order.

Two older changes matter as much for a store owner, and the first version of this article underplayed them:

OpenAI stepped back from checkout inside ChatGPT. Instant Checkout launched in September 2025 with Etsy, Walmart and Shopify as partners, per Retail Insight Network. In March 2026 an OpenAI spokesperson said “Instant Checkout is moving to Apps” and that it is “prioritizing making ChatGPT search and product discovery great.” Shopify stores are now discovered in ChatGPT and the buyer pays on the store’s own checkout.
The courts haven’t settled whether a site can shut agents out. Amazon won an injunction against Perplexity’s Comet browser in March 2026, and in August the Ninth Circuit vacated it, finding that on the current record “Perplexity itself did not access Amazon’s computer systems” because the agent acts on the user’s instructions. The case is back in district court.

My read: “AI shopping” isn’t one channel. Amazon’s agent works mainly from Amazon listings and reaches other stores only on Amazon’s terms. The open agents (Google, ChatGPT, Meta, Copilot) work from feeds and send buyers to your store. Each needs its own plan, and neither needs a big budget yet.

How an AI shopping agent works, step by step

It understands the request. The buyer describes a need in plain words, often with constraints: budget, size, use, delivery date.
It pulls candidates from product data. Google’s agents draw on its Shopping Graph of “60+ billion continually refreshed product listings.” ChatGPT now leans on merchant feeds: Profound, an AI-visibility vendor, measured the share of ChatGPT Shopping picks coming from feeds jumping from 8.26% to 61.54% on July 10, 2026, and about 65% by early September, across its own tracked prompts (OpenAI hasn’t confirmed the change).
It compares. Alexa for Shopping builds side-by-side comparisons and custom guides for big purchases covering “features, prices and reviews across Amazon and other websites,” with a year of price history.
It waits, if asked. Google’s Universal Cart, launched in the US on October 15, 2026, saves items from several retailers and sends alerts on price drops and restocks. Alexa can add an item to your cart when it hits a target price.
It hands over or checks out. This is where the agents differ most:
- ChatGPT mostly refers. For Shopify stores, the buyer completes the purchase on the store’s own checkout via an in-app browser in ChatGPT on mobile, per Shopify, and in a new tab on desktop.
- Google shows a Buy button in AI Mode and Gemini for merchants on the Universal Commerce Protocol (UCP). Payment goes through Google Pay and “You will remain the seller of record.”
- Meta’s Muse opens a browser, fills in forms and pays with Link by Stripe, which “generates a one-time-use card so your real card details stay hidden.” It “checks with the person before sensitive actions like… making a purchase.”
- Amazon’s Buy for Me completes the purchase on a brand’s own website with the customer’s encrypted name, address and payment details, and “Delivery, returns and exchanges, and customer service are managed by the brand store.”

The common thread: the agent is a buyer that works from data. Attributes, prices, stock, reviews and policies are what it compares, so they decide whether you’re on the shortlist.

The checkout routes, and how a store joins each

Agent	How your products get in	How the order is paid	Who can use it (Nov 2026)
ChatGPT, Shopify store	Automatic through Shopify’s Agentic Storefronts, on by default since late March 2026	On your own Shopify checkout, inside ChatGPT’s in-app browser	Stores selling to US customers, with terms, privacy and return policies filled in (per Craftshift)
ChatGPT, other platforms	Apply to OpenAI for direct feed access (waitlisted, per Search Engine Journal’s report of Profound’s research), or list through Shopify’s Agentic plan or a platform on Stripe’s Agentic Commerce Suite	On your site; ChatGPT acts as a referrer	Feed access is gated; OpenAI plans self-serve later in 2026
Google AI Mode, Gemini, Universal Cart	Merchant Center feed for discovery; for the Buy button, an interest form plus a UCP integration and the `native_commerce(checkout_eligibility)` attribute	Google Pay, with your payment provider behind it; you stay seller of record	Buy button: US, Canada and Australia for participating merchants (Google’s UCP help page); Universal Cart launched in the US only
Meta Muse	Launched with “the entire Shopify catalogue” plus named retailers (Walmart, Best Buy, Sephora, Ulta, Wayfair and others)	Link by Stripe one-time card; Shop Pay and PayPal being added	Rolling out in the US; I found no sign-up route for stores outside those two paths
Microsoft Copilot	Through Shopify: “Thousands of merchants are already selling in Microsoft Copilot”	Direct checkout switch in Shopify	Shopify stores
Amazon Alexa for Shopping	Your Amazon listings; outside stores only through Shop Direct and Buy for Me, which Amazon chooses	Amazon buys on your site with the customer’s details	You can’t apply; you can opt out by emailing [email protected]

Three protocol names come up behind these routes. OpenAI’s Agentic Commerce Protocol defines the product feed and checkout calls (“OpenAI is not the merchant of record”). Google’s UCP, co-developed with Shopify, “allow[s] AI systems to connect to merchants’ checkout flows.” Google’s Agent Payments Protocol (AP2), announced in September 2025 with more than 60 companies, adds signed “mandates”: an Intent Mandate when the shopper asks and a Cart Mandate the shopper signs to lock item and price, which together form what Google calls a “non-repudiable audit trail.” Shopify’s Harley Finkelstein said each of its merchants is “UCP-ready”; off Shopify, ask your platform or payment provider which of these it supports before building anything yourself.

On Shopify the controls sit under Sales channels > Agentic, per a guide from Craftshift. “Allow Shopify to manage for me” is on by default; switching it off shows a toggle per channel, and switching a channel off can take up to 7 days to stop sharing. The Sources tab shows how many products are eligible and why others aren’t.

Not on Shopify: Shopify’s Agentic plan lists products from other platforms in Shopify Catalog. A September 2026 guide on eCommerce Fastlane says it needs an external URL for each product, a verified domain, published policies, identity and business verification, and a sales history, and that “ChatGPT never uses direct checkout on this plan; it acts as a referrer.” Stripe’s Agentic Commerce Suite (December 2025) promises one integration across agents and names Wix, WooCommerce, BigCommerce, Squarespace, Etsy and commercetools as platforms that will use it.

What it costs

Route	Fee as published
Shopify store in ChatGPT	”no transaction fees beyond standard processing rates” (Shopify)
Shopify Agentic plan	no monthly fee; Shopify’s direct checkout charges 2.9% + 30¢ in the US, per eCommerce Fastlane (not from Shopify’s own page)
OpenAI Instant Checkout (old route)	a reported 4% on top of payment processing (Shopifreaks, January 26, 2026, said Shopify confirmed it); the Shopify integration no longer uses Instant Checkout
Google UCP checkout	I found no published fee; Google’s help page doesn’t mention one
Stripe Agentic Commerce Suite	no price on the launch announcement
Amazon Buy for Me	I found no fee; Amazon buys at your own site’s price

So the honest cost today is staff time: feed work, review collection, and for Google’s Buy button, a developer project with your payment provider. Etsy, one of the first Instant Checkout partners, said it mainly gave it visibility of a transaction after the purchase was done (CNBC, via Retail Insight Network).

What agents read, and how to fix it

ChatGPT’s feed. OpenAI’s feed spec requires nine fields: item_id, title, description, url, brand, seller_name, image_url, availability and price. Optional fields include gtin, sale_price, shipping_price, return_deadline_in_days, review_count and star_rating. Prices and stock should be updated “when they change,” and “Eligibility does not guarantee display.” On Shopify, Catalog builds this for you; the quality of your titles, descriptions and variants is what decides whether you get picked.

Google’s feed. Google says there’s “no special schema.org structured data that you need to add” for its AI features, and no new “AI text files” either. It asks for Merchant Center and Business Profile data that’s up to date and structured data that matches the visible page. What’s new is a set of Merchant Center attributes, as listed in PPC Land’s May 2026 summary of Google’s documentation. Question and answer, document link, related product and popularity rank carry Google’s note that they are “primarily intended for use in conversational experiences such as AI Mode.” Product highlight and product detail sit alongside them, which Google’s documentation ties to discovery on AI-driven surfaces and to technical specifications. Google’s May 20, 2026 post says retailers globally can use “conversational attributes”; PPC Land dates the attributes themselves to January 2026.

Reviews and ratings. In the lab study below, small differences in rating and review count moved agent choices a lot, and so did rewritten descriptions, depending on the model. ChatGPT’s feed has fields for ratings and review counts, so send them.

Amazon. Alexa builds its comparisons from your listing’s features, prices and reviews. I found no published research on what moves Alexa’s picks, so treat listing work there as a sensible bet.

How agents choose between products

The best evidence is a Columbia and Yale study, “What Is Your AI Agent Buying?” (Allouah and co-authors, version of December 15, 2025), which put AI agents in a mock store and ran randomized trials:

Ratings move choices a lot. A product with a 10% baseline chance of being picked rose to 15.4% (Claude Sonnet 4), 20.3% (GPT-4.1) or 16.0% (Gemini 2.5 Flash) with a rating 0.1 stars higher.
Review counts buy pricing power. Doubling reviews let a product charge 19%, 37% or 17% more without losing share.
Labels cut both ways. A “Sponsored” tag dropped selection from 10% to 7.9-8.9%. An “Overall Pick” badge raised it to 19.9-42.6%.
Position matters per model. For Claude Sonnet 4, moving a product from the bottom-right slot to the top row raised selection about five times.
Rewriting descriptions helps unevenly. A seller-side agent rewriting a description gained 3.66 to 14.89 points of share depending on the buyer’s model, and 0.32 points (not significant) with Gemini 3.0 Pro. In 67% of category and model pairs there was no significant change.
Different models crown different winners. In the iPhone case category, GPT-4.1 picked Mikeke 62.6% of the time; GPT-5.1 picked ESR 95% of the time and Mikeke 5%.

Microsoft’s Magentic Marketplace study (November 2025) adds the failure side. Most models took the first “good enough” offer instead of comparing further, more listings led to fewer comparisons, and fake credentials, fake reviews and prompt injection could steer them; Gemini 2.5 Flash was generally resistant. Microsoft’s conclusion: “Agents should assist, not replace, human decision-making.”

Both are lab stores, and models change every few months. The finding that holds across them: ratings and reviews move picks, and agents can be gamed.

How much shopping goes through agents today

Very little, and growing fast. The numbers that measure share all point the same way:

ChatGPT referrals: under 0.2% of ecommerce traffic across 973 ecommerce sites from August 2024 to July 2025, about 50,000 transactions against 164 million from other channels (Kaiser and Schulze, reported by Practical Ecommerce).
Shopify, Q2 2026: AI-driven traffic and orders to Shopify stores tripled year on year, and new-buyer orders from AI came in at nearly twice the rate of other channels. Harley Finkelstein added that “the volume from agentic commerce is still small relative to our massive GMV,” and that 75% of AI-attributable orders came from outside Shopify’s 100 largest categories.
Amazon Buy for Me: a brand with 4,000 products listed through it got about 16 orders from November 2025 to early January 2026 (Modern Retail).
Meta Muse: more than 10 million users; over a quarter of engaged users who shopped through it had an average purchase above $200.
Adobe, US retail: AI-referred traffic up 693% year over year in the 2025 holidays against 2024, with Adobe noting “the base of users remains modest”; up 62% year on year in July 2026, with those visits converting 60% better than non-AI traffic and 53% more revenue per visit.

That last point has counter-evidence. In the 973-site study, ChatGPT traffic converted about twice as well as paid social, but organic search converted about 13% better than AI referrals, affiliate 86% better and paid search 45% better. The two measure different periods and sites, so I’d treat “AI visitors convert better” as unproven for your store until your own analytics show it.

And buyers aren’t handing over the card yet. In Accenture’s survey of 25,590 consumers in 16 countries (2026), 32% would let an agent make a purchase decision within set limits, with the shopper approving before payment, and 9% would let it start and complete a purchase.

Fraud, chargebacks and bot blocking

Disputes land on you. Chargeflow, which sells chargeback management (so it has a reason to stress the risk), notes that “no network rule assigns agent errors to the platform, and the merchant is the party answering the dispute.” There’s no agent-specific reason code at Visa or Mastercard: a wrong item goes in as Visa 13.3 or Mastercard 4853, an unrecognized order as Visa 10.4, and “No card network gives agent orders a separate allowance” against dispute-ratio thresholds such as Visa’s excessive-dispute threshold (1.5% and at least 1,500 monthly fraud and dispute records in most regions). Its advice: “Keep your own copy of the cart, price and shipping address the agent submitted.” AP2’s signed mandates and Muse’s one-time cards are built to give you that record; in Google’s and Shopify’s routes you stay seller of record, so refunds and returns are yours too.

Bot protection can block real buyers. VentureBeat reported that merchants’ bot detection can “now risk accidentally blocking legitimate AI shoppers along with bad actors.” The fix being built is identification: Visa’s Trusted Agent Protocol, built with Cloudflare on HTTP Message Signatures and aligned with Web Bot Auth, lets an approved agent sign its requests and say whether it’s browsing or buying. Mastercard is building Web Bot Auth into Agent Pay, and American Express will also use it. Agents are approved and onboarded through Visa’s Intelligent Commerce program, and the merchant decides which verified agents to let in. Amazon already requires agents to identify themselves with a text snippet in their web requests, which is one reason it gave for blocking Muse; it also said Meta never told it Muse would access the store and raised concerns about login credentials, while Meta says Muse can’t see customers’ passwords or payment information, according to Retail Gazette’s reporting.

Listings you didn’t approve. Modern Retail reported on January 6, 2026 that 145 brands told one founder’s survey that Amazon had listed their products through Buy for Me without consent, with errors like a vinyl sticker shown with a photo of pants and orders for out-of-stock items. Amazon calls Shop Direct and Buy for Me programs “we’re testing” and says brands “can opt out at any time” by emailing [email protected].

What to do with this

For: online stores selling physical products to US customers, on Shopify or another platform, with or without Amazon. Stores outside the US can still do steps 1, 3, 4, 5 and 7; Google’s conversational attributes are open worldwide, but the checkout routes above are mostly US-only (Google’s also covers Canada and Australia). Not for: services, local businesses without a catalogue, and B2B sales that close through a rep. Agents don’t buy those yet.

Split the plan in two (30 minutes). List which products sell on Amazon and which on your own store. Amazon products need listing work for Alexa; own-store products need feed work for Google, ChatGPT, Copilot and Muse. Check: every top product has an owner for one or both.
On Shopify, check what’s live (20 minutes). Go to Sales channels > Agentic, open Sources, and read the eligible count and the reasons next to “Ineligible.” Make sure your terms, privacy and return policies are filled in. Leave “Allow Shopify to manage for me” on unless you have a reason to opt out. Check: ask ChatGPT and Copilot for your product by name and by need, and see whether it shows with the right price.
Not on Shopify, pick one way in (an afternoon to decide). If you already run a Merchant Center feed, start there: Google’s checkout help page calls Merchant Center the hub for your product data. For ChatGPT, apply for direct feed access, or ask your platform (WooCommerce, BigCommerce, Wix, Squarespace) whether it supports Stripe’s Agentic Commerce Suite, or weigh Shopify’s Agentic plan. Check: one route chosen, with a date.
Fill the fields agents compare (an hour per category). In Merchant Center, add question-and-answer, product highlight, product detail and related product attributes to your top products. In any feed, send gtin, shipping_price, return_deadline_in_days, review_count and star_rating. Write descriptions with the use cases buyers type (“for flat feet”, “fits a 13-inch laptop”). Check: your top 20 products carry the new attributes in Merchant Center and the optional fields in your other feeds.
Grow reviews honestly (ongoing). Ask every buyer for a review after delivery. In the lab study, a 0.1-star gain lifted a product’s pick rate by 5.4 to 10.3 points depending on the model, and doubled reviews bought 17-37% price headroom; description rewrites moved picks too, unevenly. Check: review count per top product, monthly.
Prepare for disputes (an hour with your payment provider). Ask your processor how agent orders are flagged and keep the submitted cart, price and address for every order from an agent referral. Check: you can pull those three things for last month’s AI-referred orders.
Measure agent traffic on its own (30 minutes). Build a segment for referrals from chatgpt.com, gemini.google.com, copilot.microsoft.com and muse.ai and watch orders and conversion, not visits. Shopify’s Agentic plan doesn’t show ChatGPT order history in Shopify, so this report is your only view. Check: last month’s AI-referred orders in one report.
For Amazon, decide on Buy for Me (10 minutes). Search the Amazon Shopping app for your brand under “Shop brand sites directly.” If you’re listed and don’t want to be, email [email protected]. Check: the listing is gone or correct.
Leave Google’s Buy button for later unless you’re a US, Canadian or Australian merchant with a developer and a payment provider that supports Google Pay. The interest form is in Google’s UCP help page. My read: wait until agent orders show up in step 7.

What I’d do this week

Pick your three best sellers and ask Alexa, Google AI Mode and ChatGPT to recommend a product in that category. Write down who they pick and why they say they picked it. Then put your listing next to the winner’s: missing attributes, fewer reviews, a different price. Attributes and prices you can fix this week; reviews take longer.

More on how AI answers pick sources: seo/agentic-search-optimization, glossary/retrieval-vs-citation, ChatGPT shopping is mostly Google Shopping, and the protocol side in the AI checkout protocol war.

Sources

More of this every day on the channel: https://t.me/aidaily_business