How AI agents shop: the discovery problem OpenAI just made visible

On March 25, 2026, OpenAI launched product discovery in ChatGPT. Visual product grids. Side-by-side comparison tables. Image-based search. You describe what you want, refine it in conversation, and browse products without leaving the chat.

This is a real product solving a real problem. Shopping on the web means jumping between tabs, reading the same "best of" lists, trying to piece together what to buy. ChatGPT as an AI shopping assistant now handles that for you, at least for products from merchants who participate.

That qualifier matters. It defines the boundary of what AI shopping agents can see today, and what they can't.

The agent shopping flow

Before analyzing what OpenAI built, it helps to map the full shopping flow that an AI agent needs to support.

"I need trail running shoes under $150"
  
  1  DISCOVER      What exists? Search, filter, browse.
        ↓
  2  COMPARE       Same shoe on three sites? Which is cheapest?
        ↓
  3  DECIDE        Price history, reviews, availability.
        ↓
  4  TRANSACT      Checkout, payment, fulfillment.

Each step requires different data and different capabilities. Transaction is solved. Both ACP (OpenAI + Stripe) and UCP (Google + Shopify + 20 retailers) handle checkout. Stripe endorses both protocols. The transaction layer is standardized and competitive.

Discovery and comparison are where it gets interesting.

What OpenAI shipped

ChatGPT's shopping experience is powered by the Agentic Commerce Protocol, extended from checkout into discovery. Here is what it includes.

Merchant feeds. Retailers submit structured product data (CSV, TSV, XML, or JSON) to OpenAI's endpoint. Updates are accepted as often as every 15 minutes. Required fields include product identifiers (GTIN, UPC, MPN), variant details, pricing, availability, and logistics data.

Shopify Catalog. All Shopify merchants are automatically included through Shopify's Global Catalog. Shopify's MCP server indexes billions of products and deduplicates them using a Universal Product ID (UPID). If five stores sell the same coffee mug, an agent sees one consolidated listing. No extra work required from individual merchants.

Big retailer partnerships. Target, Sephora, Nordstrom, Lowe's, Best Buy, Home Depot, and Wayfair have integrated directly into ACP for discovery. Walmart gets a native in-ChatGPT app with account linking, loyalty integration, and Walmart payments.

Visual shopping interface. Products are shown as image grids, not text lists. Users can upload images as inspiration, compare products side by side with key specs, and refine results conversationally.

A strategic pivot. OpenAI pulled back its Instant Checkout feature to focus on discovery. In their words, Instant Checkout "did not offer the level of flexibility that we aspire to provide." Merchants now use their own checkout experiences while OpenAI concentrates on the pre-purchase experience. Discovery is the bet.

This is significant. The largest AI platform just decided that helping people figure out what to buy is more valuable than handling the payment.

What merchant feeds see

ACP discovery is built on merchant-submitted data. The merchant decides what products to list, at what price, with what description and images. This is the same model as Google Shopping or any product listing ad: the merchant controls the feed.

For the products and merchants it covers, this works well. A shopper looking for a Dyson vacuum at Best Buy gets accurate, current data because Best Buy submitted it. For known brands at major retailers, feed-based discovery is fast and reliable.

But feeds have structural limitations that matter for certain use cases.

No cross-seller verification. Each merchant submits their own feed independently. ChatGPT can show you the Salomon Speedcross 6 at Nordstrom for $140 and at REI for $130, but only if both submit feeds, and only if the system can match them as the same product across two different feeds with different titles, images, and identifiers. Merchant feeds are not designed for cross-seller comparison. They are designed to represent one merchant's catalog.

Coverage is opt-in. Shopify merchants are auto-included, which covers millions of stores. But WooCommerce sites, Magento stores, custom-built ecommerce platforms, and regional retailers are not. If a merchant doesn't submit a feed or isn't on Shopify, their products don't exist in ChatGPT's shopping experience.

Amazon is absent. The largest ecommerce platform in the US is not part of ACP or UCP. Amazon has chosen to keep its catalog inside its own ecosystem, serving it through Rufus (300 million users, reportedly $12 billion in incremental sales in 2025) rather than sharing it with competing AI platforms. This is a deliberate business decision, not a technical limitation. It means the single largest product catalog in the US is invisible to ChatGPT shopping.

Data is self-reported. The merchant controls what the agent sees. If a retailer inflates a "was" price, omits a product, or delays updating availability, the feed reflects the merchant's version of reality. There is no independent verification.

None of this is a criticism of OpenAI's approach. Merchant feeds are the right starting point. They are structured, standardized, and available now. But they represent one category of product data: cooperative, opt-in, merchant-controlled.

The independent discovery gap

There's another category of product data: what's actually on the page.

A few scenarios where merchant feeds don't help.

Price monitoring. A brand wants to know what competitors are charging for similar products. Competitors are not going to submit their pricing data to a feed that helps the brand undercut them. The only way to get competitive pricing is to extract it from the source.

MAP compliance. A manufacturer needs to verify that authorized resellers are not advertising below the minimum advertised price. Resellers violating MAP are not going to self-report the violation through a product feed. You need to see what is on the page, not what the reseller claims. (We wrote about how MAP monitoring works in detail.)

Long-tail discovery. A buyer is looking for a specialty product from a niche manufacturer running WooCommerce. No ACP feed. Not on Shopify. The product exists on the web but is invisible to any feed-based system.

Cross-platform comparison. You want the best price on a specific running shoe across every retailer that carries it. Feed-based discovery covers the retailers that opted in. Independent extraction covers the ones that didn't.

These are not edge cases. Price monitoring alone is a multi-billion dollar market. MAP enforcement is a legal requirement for many brands. And cross-platform price comparison is the single most requested feature from online shoppers.

Merchant-controlled discovery tells you what sellers want you to see. Independent discovery tells you what's actually there. Different problems, different data sources.

The emerging agentic commerce stack

The stack is becoming clearer. It has layers, and discovery isn't a single layer.

  Consumer
    ↓
  AI Agent (ChatGPT, Gemini, Claude, custom)
    │
    ├── Merchant Discovery            ACP feeds, Shopify Catalog
    │   opt-in, merchant-controlled
    │
    ├── Independent Discovery         Web extraction, cross-seller matching
    │   ground truth, any site
    │
    ├── Protocol Layer                ACP checkout, UCP checkout
    │
    └── Transaction Layer             Stripe, PayPal, Adyen

Merchant discovery is what OpenAI just built. Merchants share their catalogs. The agent shows products from participating merchants. This works well for known brands at major retailers. It is fast, structured, and the merchant maintains control over their listing.

Independent discovery covers everything else. Product data extracted from the open web, normalized to a consistent schema, matched across sellers. The sites that don't submit feeds. Cross-seller comparison. Ground truth against self-reported claims.

Different use cases need different points on the spectrum.

The analogy to web search is useful here. Google does not rely solely on structured data that websites submit (schema.org, sitemaps). It also crawls the web independently. Both signals matter. Structured data from the source is clean and reliable. Crawled data provides coverage and independence. Web search got good when both signals combined.

Agentic commerce is heading in the same direction.

What the protocols cover (and don't)

Three protocols are shaping the stack. Each makes specific architectural choices about discovery.

ACP (OpenAI + Stripe). As of March 2026, ACP covers both discovery and checkout. Merchants push structured feeds to OpenAI. Products appear in ChatGPT. The checkout flow redirects to the merchant's own experience. ACP is live in ChatGPT for all users. Discovery is limited to feed participants and Shopify Catalog.

UCP (Google + Shopify + 20 retailers). Announced January 2026 at NRF. UCP covers checkout, identity linking, and order management. Product discovery is on UCP's roadmap but is currently delegated to external layers, primarily Shopify Catalog MCP for Shopify merchants and Google's own shopping infrastructure for others. UCP explicitly leaves discovery to external sources.

Shopify Catalog MCP. The most complete standardized discovery interface today. Two tools: search_global_products and get_global_product_details. Covers billions of products from Shopify merchants. Deduplicates using Universal Product IDs (UPIDs). Available to any MCP-compatible agent. But limited to Shopify's ecosystem.

The gap is consistent across all three: coverage stops at the boundary of participating merchants. Everything outside that boundary, the non-Shopify DTC stores, the international retailers, the marketplaces that keep their data proprietary, remains uncovered.

The numbers behind the shift

The market is moving faster than the infrastructure.

AI-driven traffic to US retail sites grew 693% year over year during the 2025 holiday season. Shoppers arriving from AI assistants converted 31% more than other traffic sources. AI referrals outperformed every other channel on bounce rate, with 33% fewer immediate exits.

McKinsey projects agentic commerce will reach $1 trillion in US B2C retail by 2030, with global estimates of $3-5 trillion. IBM reports that 45% of consumers already use AI for at least part of their buying journey.

But there is a gap between intent and execution. Only 22% of users who research products in AI commerce tools complete their purchase inside the AI interface. The rest leave to verify prices, check other sellers, or buy elsewhere. That behavior reveals a trust gap. Shoppers don't yet trust that the AI is showing them the complete picture, because in most cases, it isn't.

Independent discovery, the kind that covers the full web and enables real cross-seller comparison, is part of closing that gap.

Where this is heading

Where we think this goes.

Merchant-controlled and independent discovery will coexist. Just as web search uses both structured data and crawled content, agent commerce will use both merchant feeds and independent extraction. Feeds are fast and structured. Independent extraction is comprehensive and verifiable. Neither alone is sufficient.

Platform-locked discovery will lose to platform-agnostic. ACP feeds go to ChatGPT. Shopify Catalog is available via MCP to any agent, which is a better model. Product intelligence that serves any agent platform will be more valuable than data locked to a single chat interface. This seems obvious in retrospect, but most of the infrastructure being built right now is platform-specific.

Cross-seller product matching becomes critical infrastructure. Recognizing the same product across different sites, with different titles, images, and identifiers, is the foundation of real price comparison. Shopify's UPID does this within Shopify. Doing it across the entire open web is harder and more valuable.

What we're building

Extralt is the independent layer. We extract structured product data from the open web, normalize it to a consistent ecommerce schema, and match products across sellers.

The pipeline has four phases: Extract (crawl and structure), Enrich (classify, translate, match), Extend (build cross-seller relationships), and Explore (query the dataset via API and MCP server).

We don't replace ACP feeds or Shopify Catalog. We cover what feeds can't see: competitive pricing from non-cooperating merchants, products from sites that don't submit feeds, cross-seller price comparison across the full web, and ground truth against self-reported data.

One MCP server, any agent platform.

Join the waitlist if you're building on this stack.

Frequently asked questions

How do AI shopping agents find products?

AI shopping agents find products through two main channels: merchant-submitted product feeds (like ACP feeds sent to ChatGPT or Shopify Catalog data) and web extraction (crawling product pages directly). Most current implementations rely on merchant feeds, which means agents only see products from merchants who opt in. Independent extraction covers the rest of the web.

What is the Agentic Commerce Protocol (ACP)?

ACP is an open standard developed by OpenAI and Stripe that enables AI agents to discover products and complete purchases. Merchants submit structured product feeds (CSV, TSV, XML, or JSON) to make their catalogs visible to ChatGPT. As of March 2026, ACP supports product discovery with visual browsing, comparison tables, and conversational refinement. Leading retailers including Target, Sephora, Nordstrom, and Best Buy have integrated.

What is the difference between ACP and UCP?

ACP (Agentic Commerce Protocol) is backed by OpenAI and Stripe, focused on ChatGPT shopping. UCP (Universal Commerce Protocol) is backed by Google, Shopify, and 20+ retailers including Target, Walmart, and Wayfair, designed for Google AI Mode and Gemini. Both handle checkout. ACP now also handles discovery via merchant feeds. UCP delegates discovery to external layers like Shopify Catalog MCP.

Can AI agents compare prices across different stores?

With merchant feed-based discovery, AI agents can show products from multiple stores but cannot reliably confirm that two listings are the same product. Cross-seller price comparison requires independent product matching, identifying the same item across different sites regardless of how each merchant describes it. This is what Shopify's UPID does within Shopify. Doing it across the open web requires extraction and enrichment infrastructure.

Why is Amazon not part of ACP or UCP?

Amazon has built proprietary AI shopping agents (Rufus, Alexa+, Buy for Me) within its own ecosystem. Rufus serves 300 million users. Joining open protocols would expose Amazon's catalog to competing AI platforms, which conflicts with Amazon's advertising business model. Amazon's catalog, the largest in US ecommerce, is invisible to both ChatGPT shopping and Google's UCP-powered experiences.

What products can ChatGPT find when shopping?

ChatGPT can find products from merchants who submit ACP feeds (Target, Sephora, Nordstrom, Lowe's, Best Buy, Home Depot, Wayfair) plus all Shopify merchants via Shopify Catalog integration. Products from non-Shopify independent stores, Amazon, and merchants who don't submit feeds are not included. ChatGPT shopping is currently available in the US, with more regions planned.