How AI Agents Actually Discover Your Products: A Technical Guide to Agent Cards and Feeds
A technical guide to how AI agents discover products, covering Agent Cards, product feeds, and optimization strategies for agentic commerce visibility.
The Discovery Problem: From SEO to GEO
For the last twenty years, e-commerce discoverability lived in one world: getting humans to find you. Google's algorithm indexed your pages. Amazon's search ranked your listings. You optimized for keywords because keywords drove human behavior.
The infrastructure for agent-led product discovery is maturing fast. On April 27, 2026, Feedonomics launched Agentic Catalog Exports (ACE), an enterprise service designed to syndicate product catalogs to AI-powered shopping environments including OpenAI and Google Gemini. The shift from traditional search feeds to structured, agent-ready catalog exports is accelerating. In one production audit of a US Shopify catalog, AI shopping assistants ignored over 40 percent of the inventory simply because the product feed lacked structured attributes and stable identifiers.
The demand side confirms the trend. Thirty-four percent of US online shoppers have already used an AI agent to assist with a purchase decision, up from 9 percent in 2024. Shopify merchants with Agentic Storefronts see their products automatically surfaced in ChatGPT, Microsoft Copilot, Google AI Mode, and Gemini. Bloomreach launched Loomi Connect, an MCP integration that connects their e-commerce search intelligence directly into ChatGPT apps on the OpenAI marketplace.
That world is breaking. AI agents now handle discovery. When a consumer asks ChatGPT to find the best wireless earbuds for running, they are not clicking through Google results. They are asking an AI to evaluate thousands of products and recommend three. The AI reads product feeds, compares specifications, weighs reviews, and synthesizes an answer. Your SEO strategy does not reach that AI.
This shift requires a new discipline: GEO, or Generative Engine Optimization. If SEO was about getting humans to find you through text search, GEO is about getting AI agents to find, evaluate, and recommend you through structured feeds, APIs, and purpose-built product formats.
But how do these agents actually work? What data do they read? How do they decide whether to recommend your product over a competitor's? This article answers those questions with technical depth that translates into actionable strategy.
Three Paths to AI Agent Product Discovery
AI shopping agents discover products through three distinct mechanisms. Most retailers use all three, often without knowing it.
Path One: Web Crawling and Extraction
The oldest path is also the most familiar. Bots crawl your website, just like Googlebot does. But instead of indexing pages for ranking, they extract structured data to feed into LLM reasoning.
Perplexity, Bing AI, and even Claude retrieve product data this way. They hit your website, parse the HTML, extract product names, prices, descriptions, and images, then feed that data into their generative models alongside user queries. If your product page contains clear specifications and customer reviews, the agent can reason about it. If your page is thin or poorly structured, the agent may skip you entirely.
This path requires no special setup from you. Your existing product pages are being crawled right now. But there is a critical catch: crawlers cannot deeply reason about product intent if your markup is unstructured. A human can read "waterproof running watch with GPS and heart rate" and understand the use case. An LLM extracting plain text from a product page may misinterpret whether this watch is better for trails or roads, whether it suits beginners or advanced athletes.
Path Two: API Integration
The second path is structured APIs. Your product data lives in a catalog system. Instead of scrapers pulling from your website, authorized AI assistants connect directly to your API endpoints through ACP (Agent Commerce Protocol) or UCP (Unified Commerce Protocol).
This path offers precision. Your API returns structured product objects with all attributes in consistent formats. Price is always a number. Size options are always an array. Availability is a boolean. The AI agent consumes this data without the parsing ambiguity that comes with web crawling.
Google AI Shopping uses this path for retailers in their integration network. Klarna's Agent Mode connects through APIs to pull real-time inventory and pricing. Querytail's Design Partners leverage Querytail OS to make their catalogs API-readable for any LLM-powered shopping assistant.
But API integration requires technical setup and ongoing maintenance. You must expose your catalog through a standards-based interface. You must handle authentication. You must keep data synchronized. This path is worth it for high-value retailers, but not all brands have the resources.
Path Three: Agent Cards
The third path is purpose-built. Agent Cards are pre-packaged product summaries designed specifically for LLM consumption. They sit between your crawlable website and your API. They are lighter than APIs but richer than raw HTML.
An Agent Card is not a web page. It is a structured, queryable product summary that includes everything an AI agent needs to recommend your product. It contains the product's identity, attributes, reasoning context, constraints, and trust signals, all formatted for maximum clarity to a language model.
Agent Cards live in feeds that agents discover through standard channels: Sitemaps, feeds submitted to platforms like Google Merchant Center, or direct crawl. A platform like Klarna or Mirakl can aggregate Agent Cards from dozens of retailers and serve them to their agentic shopping assistants. A brand-owned Agentic Client Advisor can query Agent Cards to power recommendations on the brand's own website.
This path is scalable. You create Agent Cards once, and they can be consumed by multiple agents across multiple distribution channels. You do not need a separate API for each agent. You do not need to rebuild your product pages.
Anatomy of an Agent Card
A well-designed Agent Card contains five layers of information, each serving a distinct purpose in the agent's decision-making process.
Layer One: Product Identity
The foundation of an Agent Card is unambiguous product identity. This includes the product name, brand, SKU, UPC, and category. Identity must be precise because agents cross-reference products across feeds and platforms. If your Agent Card says "Nike Air Max 270" but your shopping page says "Air Max 270", the agent may treat these as different products.
Identity also includes locale-specific information: currency, region, language. An agent serving a French customer should see prices in EUR and descriptions in French. An Agent Card must signal these variants clearly so the agent can serve the right version to the right shopper.
Layer Two: Attributes
The second layer is structured attributes. These are key-value pairs that describe the product's physical and functional properties. Size, color, material, weight, dimensions, power consumption, warranty period, and so on.
Attributes are critical for comparison. When an agent evaluates three running watches, it compares their attributes side by side. It checks battery life, waterproof rating, available colors, price, and return policy. If your Agent Card omits battery life but a competitor's includes it, the agent may recommend the competitor because it can reason about that product more fully.
Attributes must be structured consistently across your entire feed. If one product lists size as "small, medium, large" and another uses "XS, S, M, L", the agent cannot reliably compare them. Your Agent Cards should define a taxonomy of attributes and enforce it across all products.
Layer Three: Reasoning Context
Structured attributes are necessary but not sufficient. They tell the agent what a product is. Reasoning context tells the agent why it exists.
This layer includes natural language descriptions that answer fundamental questions: What problem does this product solve? Who is it for? What situations is it ideal for? How does it differ from similar products? What is the intended use case?
A generic description might say "Lightweight running shoe with responsive cushioning". Reasoning context says "Designed for distance runners who prioritize forward momentum over maximum cushioning. Best for road races over 10 miles. The minimal drop and low stack height make it ideal for runners transitioning from traditional shoes, but beginners should start with a more cushioned model".
Reasoning context helps the agent match products to specific customer needs. When a shopper asks for "the best running shoe for my first marathon", the agent can read the reasoning context and understand which shoe is built for that journey.
Layer Four: Constraints
The fourth layer defines what cannot be done with the product. This includes availability, shipping restrictions, return policies, warranty exclusions, and geographic limitations.
Constraints prevent the agent from making promises you cannot keep. If a product is out of stock, the Agent Card says so. If you cannot ship to international addresses, the Agent Card flags it. If returns are not accepted on clearance items, the Agent Card specifies it.
Without constraint layers, agents may recommend products that cannot actually be delivered to the customer. This damages trust and creates support burden.
Layer Five: Trust Signals
The final layer is trust signals. These are facts that make an agent confident in recommending your product. Customer review summary and rating, certifications (FTC verified review badge, third-party testing, sustainability certifications), guarantees, awards, and notable endorsements.
Trust signals are particularly important in ecommerce because agents operate without visual trust cues. A human shopping in a store might trust a brand because of how it looks on the shelf. An AI has no such signal. It relies on aggregated customer opinion, third-party validation, and explicit guarantees.
Agent Cards should emphasize trust signals. If your product is top-rated on three independent review sites, say so. If it carries a ten-year warranty, highlight it. If it is certified for a standard that matters (LEED, Fair Trade, Energy Star), make that discoverable.
Feed Optimization for AI Readability
Your existing product feed, whether Google Shopping XML or a Facebook catalog, is optimized for human review and algorithmic ranking. AI agents need a different optimization strategy.
Natural Language That Answers Questions
Traditional product descriptions list specifications. "10-hour battery life. Bluetooth 5.2. Water resistant to 50 meters." These specifications are useful, but they do not answer the questions agents ask on behalf of customers.
AI-optimized feeds include natural language that answers common questions: "Will this headphone work in the rain? Will the battery last my morning commute? Can I use this for swimming?" The more questions your description anticipates and answers, the more confidently an agent can recommend it.
This does not mean verbose product descriptions. It means precise, question-anticipating language. "Rain-resistant design handles typical weather but not submersion" is more useful to an agent than "water resistant".
Comparative Context
Agents frequently evaluate your product alongside competitors. They ask "How does this compare to the market leader? Is this the budget option or the premium choice? What trade-offs does it make?" If your Agent Card does not address these questions, the agent must infer them from price and generic attributes, which often leads to wrong conclusions.
Your feed should include comparative context: "This headphone prioritizes battery life over noise cancellation, unlike the premium model. It is ideal for travelers who value longevity over active sound blocking." This context helps the agent place your product correctly in the market.
Use-Case Mapping
Every product serves multiple use cases, but not all. A waterproof phone case works for beach trips and rainy commutes but not for deep-water diving. An agent needs to understand this mapping.
Use-case mapping in your feed looks like structured data that links products to scenarios. "Ideal for: beach trips, snorkeling, poolside use. Not recommended for: deep diving, extended underwater use, industrial environments." This helps agents make precise recommendations based on customer needs.
Structured FAQ Per Product
Customers ask the same questions repeatedly. Agents ask them too. If you have accumulated FAQ content for a product, format it as structured data in your Agent Card. Agents can then reference this FAQ when qualifying a recommendation.
A structured FAQ might include: "Is the band adjustable?" "What is included in the box?" "Do you offer international shipping?" "What is the return window?" Agents consume this data to confirm they can actually deliver on a recommendation.
The AI Readability Test
Here is a simple quality check: Can GPT-4 recommend the right product from your feed alone, without visiting your website? Give your product feed to ChatGPT. Ask it a realistic customer question: "I run marathons on road courses and want a lightweight shoe under $150 with at least 8 millimeters of stack height." Can the model pick the right product from your feed and explain why?
If it cannot, your feed is not AI-optimized. The model either lacks the context to understand your products or lacks the reasoning context to match products to needs. Iterate your feed until the AI readability test passes.
Distribution Channels for Agent Cards
Once you have built Agent Cards, you need to distribute them to agents. There are five primary channels.
| Channel |
Use Case |
Setup Complexity |
Real-Time Data |
| ACP Endpoints |
ChatGPT, Claude, Gemini, and other LLM assistants |
High |
Yes |
| Google Merchant Center + UCP |
Google AI Shopping and Google Ads |
Medium |
Yes |
| Direct Crawl |
Perplexity, Bing AI, other independent crawlers |
Low |
Delayed |
| Klarna APP Format |
Klarna Agent Mode shopping assistant |
High |
Yes |
| Mirakl Nexus |
Aggregator for multi-brand marketplace Agent Cards |
Medium |
Yes |
ACP Endpoints for LLM Assistants
ACP (Agent Commerce Protocol) is an emerging standard for direct API integration between retailers and LLM assistants. If you implement an ACP endpoint, assistants like ChatGPT and Claude can query your product catalog in real time.
This requires building an API that conforms to the ACP schema and registering it with platforms that support ACP. The upside is maximum control over how your products are presented. The downside is significant development effort and ongoing maintenance.
Google Merchant Center + UCP
Google Merchant Center is where you have submitted product feeds for years. UCP (Unified Commerce Protocol) extends Google's infrastructure to support real-time catalog queries for Google AI Shopping and Google Ads.
If you already use Google Merchant Center, UCP is relatively straightforward to implement. Google provides tooling and documentation. This is the path of least resistance for most retailers who already rely on Google Shopping.
Direct Crawl
Perplexity, Bing AI, and other independent agents crawl your website without special agreements. If your Agent Cards are published on your website as structured data (Schema.org markup in your HTML, JSON-LD in your page headers), these crawlers will find and use them.
This requires no partnership negotiation. But it means discovery is delayed. Crawlers may not visit your site frequently, so new products or price changes take time to propagate. And you have no control over how the crawler interprets your data.
Klarna Agent Mode
Klarna's shopping agent accepts product feeds in their proprietary APP format. If you are a Klarna partner, publishing Agent Cards through the APP channel makes your products available to Klarna's assistant when customers ask it to shop.
Klarna handles most of the integration work. But you are locked into their format and their distribution channel. Klarna Agent Mode is valuable where Klarna has market presence, but it is not universal.
Mirakl Nexus
Mirakl Nexus is an aggregation layer for marketplace operators. If you sell through a marketplace that uses Mirakl (fashion retailers, electronics, home goods), your products can be exposed through Nexus to multiple Agent Cards and shopping assistants without individual integration.
This is valuable for sellers who already operate through marketplaces. You submit products once, and Mirakl handles distribution to agents.
Measuring AI Discoverability
Once your Agent Cards are live, you need to measure whether agents are actually finding and recommending your products. This requires new measurement approaches because AI referral traffic looks different from human referral traffic.
AI Referral Traffic Tracking
When an agent recommends your product and a customer clicks through, that click should be tracked as a referral. Use UTM parameters in your Agent Cards to tag traffic from specific agents and channels. If your Agent Card is distributed through Klarna, use a Klarna-specific UTM parameter. If it is distributed through Google, use a Google-specific parameter.
You can also detect agent-generated traffic by analyzing user-agent headers. When ChatGPT sends a customer your way, the user-agent identifies the origin. Track these by agent to understand which assistants drive traffic.
Agent Cards Impression and Click-Through Metrics
Some platforms (Google, Klarna, Perplexity) publish metrics on how many times your Agent Cards were presented to customers and how many times customers clicked through. These metrics are the AI-era equivalent of impressions and clicks in Google Search.
Monitor these metrics like you would monitor Google Search impressions. If your Agent Card is showing up 100 times per day but getting clicked 0 times, your Agent Card copy, images, or positioning is weak. If another product in your category is getting 10 times more impressions, you may need to improve your reasoning context or trust signals.
Citation Tracking in AI Responses
When an AI agent recommends your product, it sometimes includes a citation: "Recommended by Klarna AI" or "Per GQ's review". Some platforms let you track mentions of your products in AI-generated content. This is still emerging, but it offers valuable signal about whether agents are reasoning about your products favorably.
FAQ
No. Agent Cards are complementary. Your Google Shopping feed continues to serve human shoppers on Google's platform. Agent Cards serve AI agents. Many retailers maintain both and continuously evolve both formats. Over time, as agents handle more commerce, Agent Cards may become your primary visibility channel. But for now, you need both.
How often should I update Agent Cards?
As frequently as your product catalog changes. If you update prices or availability daily, your Agent Cards should reflect that. If you release new products weekly, Agent Cards should be published within 24 hours of the product becoming available for sale. Real-time updates require API-based distribution. Periodic (daily or weekly) batch updates are acceptable for direct crawl channels.
Can I control what the AI agent says about my product?
You can control the data you provide, not the agent's interpretation. Your Agent Card describes the product. The agent then reasons about it, compares it to alternatives, and synthesizes a recommendation. If your reasoning context is clear and complete, the agent is more likely to recommend accurately. But you cannot dictate the exact language the agent uses.
What is the minimum product data needed for an Agent Card?
Product name, description, price, and availability are the baseline. Add structured attributes (size, color, material), use-case context, and at least one trust signal (review rating, certification, guarantee). The more complete the Agent Card, the more confidently the agent can recommend it.
How do I test AI discoverability?
Submit your Agent Card feed to a test agent. Give GPT-4 or Claude your product feed and ask realistic customer questions. Can the agent find and recommend your products? Does it understand the use cases? Does it compare your products to alternatives accurately? This test illuminates gaps in your reasoning context and attributes.
Conclusion: GEO as a Strategic Imperative
AI agents are not the future of commerce discovery. They are the present. Billions of shoppers are asking questions to ChatGPT, Gemini, and Klarna instead of typing into Google. Your products are being evaluated by AI, not algorithms. Your competitive advantage lives in how well your Agent Cards answer the questions agents ask on behalf of customers.
This is not about adding a new marketing channel alongside email and social. This is about fundamental shifts in how commerce discovery works. The retailers who build comprehensive, reasoning-rich Agent Cards first will own the agents' recommendations. The retailers who wait will be relegated to the long tail.
Start with one product category. Build Agent Cards with rich reasoning context and complete attributes. Test them with GPT-4. Measure AI referral traffic. Learn what works. Expand to your entire catalog. This is the path to AI discoverability.
Ready to explore how Querytail can help? Request a demo to see the platform in action, or contact our team with any questions. If you are a brand looking for early access, apply for the Design Partner program.
Ready to Build Your Agentic Catalog?
Agent Cards are not just a technical format. They are a new language for commerce. Let us help you speak it fluently.
Book a Demo