Agentic commerce

Product feeds for AI agents: what they read, and why yours has to agree with itself

A product feed used to be a shopping-ads chore. In 2026 it is the input that decides whether AI agents can find, compare, and buy from you at all. The catch is that your catalog now exists in three places at once, and agents check them against each other.

Updated August 26, 2026 · 12 min read

A product feed is your catalog expressed as structured data: one record per item, with typed fields for title, description, price, availability, images, shipping, and returns. It has existed for years as the thing that powers shopping ads. What changed in 2026 is its job. Google’s own guidance to advertisers, presented at NRF 2026, is direct about it: the attributes you submit determine your eligibility for the new AI surfaces. The feed stopped being a channel-specific chore and became the front door.

The part almost nobody explains properly is that your catalog now exists in three places at once. There is the merchant feed you submit to a platform, the protocol feed that agentic checkout runs on, and the structured data on your own product pages. Agents read all three, and more importantly they read them against each other. A price that disagrees across those three is not a cosmetic problem; it is the most common reason a merchant quietly stops being recommended.

This guide covers what agents actually read, the three expressions of your catalog and how they differ, why consistency is a hard requirement rather than a nice-to-have, the attributes that decide eligibility, and what to do if you sell services and have no SKUs at all.

Why the feed became the front door

When a shopper types four constraints into an assistant ("waterproof hiking boots, wide fit, under $180, in stock in a 10"), the assistant is not reading your marketing copy and forming an impression. It is filtering typed fields. Every constraint it can check against a structured attribute is a constraint you can win on. Every constraint your data does not express is one you are automatically excluded from.

The scale of the machinery underneath makes the point. Google’s Shopping Graph refreshes on the order of two billion listings per hour, which is why feed freshness now matters more for AI surfaces than it ever did for traditional shopping ads. Stale data does not just underperform, it actively disqualifies, because an assistant that recommends a sold-out item has embarrassed itself in front of its user and will discount that source next time.

There is also a gating effect worth knowing. As Google rolled UCP access out in phases through 2026, priority went to merchants with clean, complete Merchant Center feeds in good standing. Feed quality was not merely how you performed once you were in; it was how you got in.

The three expressions of your catalog

These are not three copies of the same file. They are three different jobs, and confusing them is why merchants under-invest in one and wonder why the others underperform.

The relationship between the first two is the one that trips people up. Merchant Center is the catalog data layer. UCP is the conversational and checkout layer that sits on top of it, adding cart, checkout, payment, and post-purchase order handling. A retailer with a clean Merchant Center feed gets discovered in AI Mode. A retailer with full UCP support also gets agentic checkout. The feed is necessary for both, and sufficient for neither the whole way. The UCP merchant guide covers the protocol layer in detail.

The third expression is the one merchants skip, and it is doing quiet, important work. Product, MerchantReturnPolicy, OfferShippingDetails, and AggregateRating markup on your own pages let an agent verify your product information independently of any feed you submit. That independence is the entire point. A feed is a claim you make about yourself; on-site structured data is the corroboration. The JSON-LD guide has copy-paste templates.

Consistency is a hard requirement

Google continuously compares what you submit in Merchant Center against what is actually on your website. When the two disagree, trust drops and visibility suffers. That mechanism predates agentic commerce, but its consequences got sharper, because the tolerance on the other side collapsed.

The practical implication is architectural rather than editorial. If your feed, your protocol catalog, and your page markup are maintained as three separate artifacts by three separate processes, they will drift, and drift is the failure mode. The durable answer is one canonical source of truth that generates all three, so a price change is one edit that propagates everywhere rather than three edits someone has to remember.

The attributes that decide eligibility

Google’s NRF 2026 guidance was specific about what a feed needs to carry to be useful to agents, and the shape of the advice is consistent across every surface: richer, more complete, more explicitly stated. In rough order of how often their absence costs a recommendation:

• Long, descriptive titles. Not brand-plus-SKU. The attributes a shopper would name (material, fit, size, capacity, color, use case) belong in the title where they can be matched.
• Rich descriptions that answer questions rather than sell. Agents extract facts from these, so a paragraph of adjectives contributes nothing while a paragraph of specifics contributes everything.
• Conversational attributes. Google added Merchant Center fields specifically for agentic surfaces, covering answers to common product questions, compatible accessories, and substitutes. These are the fields that let an agent handle "will this fit my..." without leaving the conversation.
• Multiple quality images. Multimodal assistants genuinely use them, and a single low-resolution shot limits how confidently you can be surfaced.
• Complete commercial terms. Shipping cost and timeframe, return window and conditions, promotional pricing. An agent comparing two similar items on total landed cost cannot include you if it cannot compute yours.
• Accurate, current availability. The single fastest way to lose standing, because it produces a visibly wrong recommendation.

The blunt version, and it is worth stating this way: every missing attribute is a lost recommendation opportunity, not a slightly weaker one. Filters exclude; they do not rank down.

Which feed do you actually need?

This depends less on strategy than on what you already run.

1. On Shopify: much of it is handled. Agentic Storefronts activated UCP and native MCP servers by default in the Winter 2026 Edition, so your work is data quality rather than integration.
2. With an existing Merchant Center feed: you are most of the way there. UCP layers onto the same catalog pipeline, and the simplified Merchant Center onboarding exists precisely for this case. Enrich what you already submit before building anything new.
3. On another platform: check whether your provider has shipped UCP support before writing code. Several committed through 2026, and waiting a quarter often beats a custom integration you then maintain forever.
4. Custom or headless: build the Merchant Center feed first for discovery, then the protocol layer for transaction. Discovery is the cheaper half and it is where the volume is.
5. Selling through ChatGPT: feed data drives product discovery there, but note that in-chat checkout is gone. OpenAI retired Instant Checkout in March 2026, so the feed earns you the recommendation and your own checkout closes it.

What if you do not have products?

Feeds are SKU-shaped, and that is a real limitation rather than something to talk around. If you sell appointments, consults, or classes, today’s feed specifications model shipping and returns and variants, none of which describe a Tuesday afternoon slot. Google has formed domain councils for verticals including lodging and food ordering, so this is moving, but it has not arrived.

In the meantime the equivalent work is the same work with a different output format: named, priced offers as structured data on their own pages, real availability exposed through a callable endpoint, and a booking action an agent can actually invoke. That path runs through an MCP server rather than a feed, and it is covered in how AI agents book service businesses. The underlying discipline is identical: state what you sell in typed fields, keep it true, and make it verifiable.

The checklist

If you do nothing else this quarter, do these in order:

1. Audit for disagreement first. Pick twenty products and compare price, availability, shipping, and returns across your feed, your protocol catalog if you have one, and your live page markup. Fix every mismatch before enriching anything, because enrichment on top of drift just propagates the drift.
2. Enrich titles and descriptions. This is the highest-return editing work and it does not require engineering.
3. Fill in commercial terms completely. Shipping, returns, and promotional pricing are frequently blank and frequently decisive.
4. Add or complete on-site structured data so your claims verify independently of the feed.
5. Fix your refresh cadence. Availability and price should propagate in minutes, not overnight, and certainly not weekly.
6. Generate rather than maintain. If three artifacts are edited by hand, budget for the day they disagree. One source of truth removes the whole class of failure.

The through-line across every agentic surface is unglamorous and consistent. Agents do not reward persuasion, they reward verifiability. A catalog that says exactly what it is, everywhere, in the same way, is the entire competitive advantage available here, and it is available to a small merchant on equal terms with a large one.