How to Get Products Recommended by AI Agents

Key Takeaways

• AI shopping agents such as ChatGPT, Gemini, and Amazon’s Rufus read structured product feeds and schema markup first. They fall back to scraping a web page only when structured data is missing, and thin data usually means the agent moves on rather than guessing. 

• A single missing GTIN can be enough for an agent to drop a product from consideration entirely. 

• Three protocols are reshaping how agents reach your catalog: MCP (how agents access data), UCP (Google’s cross-platform commerce layer), and ACP (OpenAI’s merchant-facing feed standard). All three are still rolling out through 2026. 

• Consistency matters as much as completeness. The same GTIN, price, and spec values need to match across your website, your Google Shopping feed, and any agent-facing feed, or agents treat the mismatch as a trust problem. 

• A PIM is what keeps this data complete, consistent, and fresh across every channel automatically, instead of someone patching schema by hand each time a price changes. 

• Optimizing for AI shopping agents is a data discipline, not a per-protocol scramble. The same fixes that satisfy MCP, UCP, and ACP today will satisfy whatever comes next. 

How AI Shopping Agents Actually Pick Products 

In plain terms: an AI shopping agent recommends only the products whose data it can verify in milliseconds. In fact, every agent, regardless of which company built it, follows a similar pattern. It queries a structured product feed or a page’s schema markup first. Only when that data is thin or absent does it fall back to reading the page like a person would, and even then, it is reading for confirmation, not discovery. A page with no structured data is not invisible to an agent. It is just slower to trust, and slower usually means skipped in favor of a competitor whose data answered the query on the first pass. 

Three protocols now govern how agents reach that data: 

• MCP (Model Context Protocol) standardizes how an AI agent connects to an external system to pull product data in the first place. 

• UCP (Universal Commerce Protocol), co-developed by Google and Shopify, lets agents discover products, build carts, and complete checkout across participating merchants. Google has extended UCP-powered checkout into AI Mode in Search and Gemini, currently available to select merchants in the US, Canada, and Australia. 

• ACP (Agentic Commerce Protocol) is OpenAI’s merchant-facing feed standard. OpenAI has shifted its own approach here: rather than owning checkout end to end, it now prioritizes product discovery and comparison inside ChatGPT, while checkout increasingly happens through the merchant’s own experience, as with Walmart’s in-ChatGPT app. 

Still, none of these are fully mature yet. All three are mid-rollout through 2026, which means the smart move is treating “agent-ready” data as table stakes now, not a project to revisit once a protocol finalizes. MCP and UCP sit on different layers of the same stack, built to work together rather than compete, but both run on the same checklist below. 

The Agent-Ready Data Checklist 

This is the part that actually moves the needle. Before you touch schema markup or worry about which protocol to prioritize, get these six things right. Agents check most of them before anything else. 

1. Identifiers on every SKU

A GTIN, UPC, or EAN, plus an MPN, is how an agent cross-references your listing against a known product database. Skip this field and you are not deprioritized. Instead, you are often left out entirely. One analysis of AI shopping agent behavior found that a single missing GTIN is enough for an agent to skip a product, because the agent has no reliable way to confirm what the item actually is. 

2. Core attributes, fully populated

Specifically, brand, price with a currency code, current availability, and accurate parent-child variant relationships for size and color. For example, an agent comparing three water bottles needs to know instantly which one is in stock in the size a shopper asked for. If that relationship is broken or missing, the agent drops the variant rather than guessing.

3. Structured specs, not paragraph specs

Labeled specification rows with consistent units, mapped to a named attribute rather than buried inside a description paragraph. “Weighs approximately 2 lbs” in a block of marketing copy is nearly useless to an agent. “Weight: 2 lb” as a labeled attribute is immediately usable. 

4. Descriptive language that matches how people search

Natural phrases like “waterproof,” “vegan leather,” or “BPA-free” need to live in your attributes and titles, not just as keyword-stuffed filler. After all, agents and shoppers use the same everyday language. Your data should speak it too.

5. Trust signals an agent can weigh

Finally, aggregate ratings, review counts, and a machine-readable return policy and shipping terms. When two products are otherwise similar, this is often the tiebreaker an agent uses to decide which one gets recommended.

6. Freshness that matches how fast agents move

Inventory updated within minutes, not hours, and pricing refreshed at least every 24 hours. Stale availability is one of the fastest ways to get quietly dropped from a recommendation set, because an agent that recommends an out-of-stock item erodes the platform’s own credibility along with yours.

What Schema Markup Does for Your Products 

Schema.org markup is how a product page tells an agent, in a language it can parse without guessing, exactly what it is looking at. You do not need to write this by hand or become fluent in JSON-LD to use it well. You need to know what to ask your developer or your PIM team for. 

The piece that matters most is the Offer block: price, priceCurrency, availability, and priceValidUntil. Miss any of these and an agent often will not surface the product at all, because it cannot confirm the offer is current. Beyond the offer block, aggregateRating carries your review data into a format agents can read directly, and hasVariant maps the relationship between a parent product and its size or color options, the same relationship flagged in the checklist above. 

Two free tools tell you where you stand today. Google Merchant Center’s feed diagnostics flags missing or malformed required fields in your product feed. Google’s Rich Results Test shows you exactly how a given page’s schema markup renders, field by field, so you can see what an agent sees before it does. 

This is not a hypothetical concern for the people running these catalogs today. In primary research Pimberly commissioned for its State of Product Data 2026 study, a global PIM product owner at a US-headquartered equipment parts distributor put it plainly: “How is AI reading our pages, and how do we have the right schema so it knows how to pull the data accurately?” He described agentic AI and agentic commerce as the clearest bet for how the sector changes over the next few years, not a buzzword to wait out. 

This is precisely the kind of work a PIM exists to do. Schema markup is only as good as the underlying data feeding it, and keeping attributes complete and consistent enough to generate clean feeds automatically, across every channel, without someone hand-patching schema per page, is the actual job. Optimizing a storefront for AI search starts with that same foundation, which is exactly why PIM has become the foundation of AI commerce rather than an optional layer on top of it.

One Set of Numbers, Every Channel 

Completeness is only half the problem. The same GTIN, title, price, and spec values need to match across your website, your Google Shopping feed, and any agent-facing feed. When they do not, an agent does not read the discrepancy as a minor sync delay. It reads it as a trust signal failure, the same way a shopper would if a price changed between a search result and the product page. 

This is where manual, spreadsheet-and-export workflows tend to break down first. A price update pushed to the website but not the Shopping feed, a spec value corrected in one export but not another, a GTIN typo that only shows up on one channel. None of these are dramatic failures on their own. Even so, they compound as SKU count grows, and they are exactly the kind of inconsistency that starts at the supplier level long before it reaches your storefront. Our post on supplier data quality covers why bad inputs upstream break good catalogs downstream, agent-facing feeds included. 

The Agent-Ready Audit: Run This Against Your Catalog 

Pull up a handful of your highest-traffic SKUs and check each one against this list. Any “no” answer marks where to start.

1. Does every SKU have a GTIN or UPC, plus an MPN? 
2. Is price, currency, and current availability populated and accurate for every variant? 
3. Are specs stored as labeled attributes with consistent units, not buried in description text? 
4. Does your product copy use the natural language shoppers actually search with? 
5. Do you have visible aggregate ratings, review counts, and a machine-readable return policy? 
6. Is inventory updated within minutes and pricing within 24 hours? 
7. Do your website, Shopping feed, and any agent-facing feed all show the same GTIN, price, and specs right now? 

If you answered yes to all seven, you are ahead of most of the SERP. If you answered no to more than two, that is your priority list for this quarter. 

Don’t Wait for the Protocols to Finish Rolling Out 

ACP and UCP are both mid-rollout through 2026, and it would be premature to call either one fully live or universal. What is not premature is the data standard underneath both of them. The clean, governed, agentic commerce product data that makes a listing agent-ready is the same data a well-run catalog should already have, checkout protocol or not. Waiting for a standard to finalize before fixing your feed is waiting for a reason to keep doing nothing. 

FAQs

Q: What is an AI shopping agent?
A: An AI shopping agent is software, built into a chat interface like ChatGPT or Gemini, or a shopping tool like Amazon’s Rufus, that searches, compares, and recommends products on a shopper’s behalf. It queries structured product data rather than browsing pages the way a person does. 

Q: How do AI agents decide which products to recommend?
A: Agents query structured feeds and schema markup first, ranking products on completeness, accuracy, and freshness of fields like identifiers, price, availability, and specs. When data is thin or inconsistent, the agent typically drops the product rather than guessing at missing details. 

Q: What is a GTIN and why does it matter for AI commerce?
A: A GTIN (Global Trade Item Number) is a standardized product identifier, such as a UPC or EAN, that lets an agent cross-reference a listing against known product databases. Without one, an agent has no reliable way to confirm what a product actually is, and a single missing GTIN can be enough to get a product skipped. 

Q: Do I need Schema.org markup for AI agents to find my products?
A: Yes. Schema.org Product markup, especially the Offer block with price, priceCurrency, availability, and priceValidUntil, is how a page communicates product details in a format agents can parse directly, rather than inferring them from unstructured text. 

Q: What’s the difference between MCP, UCP, and ACP?
A: MCP standardizes how an AI agent connects to a system to access data. UCP, built by Google and Shopify, governs product discovery, cart building, and checkout across participating merchants. ACP is OpenAI’s merchant-facing feed standard for product discovery and comparison inside ChatGPT. 

Q: Can a PIM make my product data “AI agent ready” automatically?
A: A PIM centralizes and governs product data so that identifiers, attributes, specs, and pricing stay complete and consistent across every channel. As a result, that consistency is what lets clean feeds and accurate schema markup generate automatically, rather than being patched by hand per channel every time something changes. 

Govern the Data, Win the Recommendation

Every item on the checklist above points back to the same root cause. Products get skipped by AI agents when product data is incomplete, inconsistent, or stale, and every one of those is a data governance problem before it is a protocol problem. Pimberly exists to solve exactly that: a single source of governed product data that feeds your website, your Shopping feed, and every agent-facing channel from the same accurate, current record, instead of three different exports that quietly drift apart. If you want to see what that looks like against your own catalog, how PIM has become the foundation of AI commerce is where that case gets made in full, and Pimberly’s own AI features close these gaps faster than a manual audit ever could.