Skip to main content

Guide Machine-facing channels

What AI shopping agents need from your product data.

In short

AI shopping assistants and agents rely on product data they can read. They need unambiguous identity, complete attributes, current offers, evidence and clear policies in a form a machine can interpret. This Youzu guide lists those requirements and shows how to check a record against a named channel.

Published by Youzu · Editorial policy

Published
Reading time
9 min read

Key points

  • Start from the product feed you already maintain; agents read the same facts, more strictly.
  • Keep identity, typed attributes, variants and offers as separate, machine-readable fields.
  • Check one real record against one named channel before changing the whole catalog.
  • Map your fields to that channel's documented schema, then validate a sample.
  • Keep human approval for what is published; readiness is a description, not a promise.
In this guide

What agentic commerce means

Agentic commerce describes shopping in which software agents act for a person: they search, compare and, where a channel allows it, place an order. The agent is not reading a product page the way a person does. It reads product and offer data, and it can only use what that data states clearly.

That changes what a weak record costs. A person can work out that 'jkt olive' is an olive jacket and that the size is in the description. An agent comparing jackets across stores needs the color, the size, the material and the seller's terms in fields it can parse. When a fact is ambiguous, the agent guesses, asks the person or moves on to a product whose record is clearer.

None of this is a new kind of data. The facts agents need are the facts search filters, comparison tables and marketplace feeds have always needed. Agents simply read them more literally and with less tolerance for gaps.

It also changes who notices mistakes. Search engines and marketplace feeds show their results to a person, who can spot an odd title or a wrong photo. In an agent flow, the person may only see the agent's summary, so a wrong attribute or a stale offer can travel further before anyone catches it.

What named channels ask for

Requirements depend on the channel, so name the channel before changing anything. 2 efforts are worth reading directly, because they shape what agents will expect.

  • Universal Commerce Protocol (UCP). An open standard that aims to give agents, platforms, businesses and payment providers a common language for agentic commerce, so they can interact without one-off integrations.[2] Google's guide for merchants describes using existing Merchant Center data to reach shoppers on its AI surfaces, and invites businesses to join a waitlist, which marks an early rollout.[1]
  • Model Context Protocol (MCP). An open standard for connecting AI applications to external systems such as data sources and tools.[3] It is a way for an assistant to reach your data, not a product data format in itself.

Both point back to the same place: the structured product and offer data you already publish in feeds. If your feed is complete, consistent and current, you are most of the way to what these channels read. If it is not, no protocol will fill the gaps for you.

Product data specifications that channels already publish show the requirements that recur everywhere:[4]

  • A stable identifier for each product and variant, such as a GTIN (a globally unique product identifier of 8, 12, 13 or 14 digits issued under GS1 standards and printed as a barcode) where one exists.
  • A title and description that describe the product rather than a promotion.
  • Typed attributes for the facts buyers filter on: size, color, material and dimensions.
  • Current price and availability, held on the offer with its seller and condition.
  • Images that show the product itself, with text that describes them.
  • The policies a buyer needs, such as delivery and returns.

The 9 layers

The Agent Readiness checklist groups what a record needs into 9 layers in 3 bands: machine-readable, governed and channel-ready. Each layer is a question you can answer for one record and one named channel.

Reference

9 layers in 3 bands

The 9 layers of the Agent Readiness checklist
LayerBandQuestion
Accessible recordMachine-readableCan a machine reach the full product record without scraping a page?
Stable identityMachine-readableDoes each product keep one identifier across sources and updates?
Structured attributesMachine-readableAre the attributes a buyer filters on present, typed and consistent?
Supported variantsGovernedAre variants modeled explicitly rather than described in free text?
Current offerGovernedIs offer information current, and do you know how stale it can get?
Policy contextGovernedDo category, compliance and moderation rules travel with the record?
Usable mediaChannel-readyIs imagery consistent, rights-cleared and described in text?
Channel schemaChannel-readyHave you mapped your record to one named destination's schema?
Validated interfaceChannel-readyHas that interface been validated against a real sample, with errors reviewed?
Each layer is a question you answer for one record and one named channel. The answers describe what is in place; they are not a score.

Each band builds on the one before. Machine-readable layers decide whether a channel can read the record at all; governed layers decide whether what it reads is right and current; channel-ready layers decide whether it arrives in the form one channel expects. A record can be fully machine-readable and still fall short in the governed band, which is the most common pattern in a first check.

The order matters. A record that cannot be reached without scraping a page, or whose identifier changes on every import, cannot be fixed by better descriptions. Work from the first band down. The middle band is where most catalogs need effort: product variants (one purchasable version of a product that differs along defined dimensions, such as size, color or pack size) modeled explicitly, each seller's terms kept on the marketplace offer (one seller's listing of a product with its own price, stock and condition, attached to a shared product record) with a known age, and policy rules attached to the records they govern.

Checking one record

Start with one real record and one named channel, not with the whole catalog. The worked example below checks the jacket record used throughout these guides against the 9 layers: what 'in place' looks like, a typical gap and the first step to close it.

Worked example

Checking the jacket record against the 9 layers

The jacket record against the 9 layers
LayerIn place looks likeA typical gapFirst step
Accessible recordThe full jacket record can be read as structured data through an API or a feed, without scraping the product page.The record exists only as text on a page.Expose the record through a feed or API that your team can document.
Stable identityThe olive jacket in size M keeps one identifier in your feed, your PIM and your listing.The identifier changes when the supplier file is re-sent.Pick one identifier per variant and send it every time.
Structured attributesColor: Olive, Material: Cotton and Closure: Buttons are separate typed fields.Color and material sit inside the title.Move facts out of titles into the fields your category defines.
Supported variants3 colors and 3 sizes sit under one jacket as variants, not as unrelated listings.Sizes are written in the description.Model variants explicitly and link them to one product.
Current offerEach seller's price and availability sit on that seller's offer, with a known maximum age.Price is copied into the product record and goes stale.Keep offers apart from the product and record how old each can be.
Policy contextCategory rules and moderation rules travel with the record, so a restricted item is flagged before it reaches a channel.Rules live in a document, not on the record.Attach the rule set for each category to the records it covers.
Usable mediaImages are consistent, cleared for use and described in text.Images have no text description, or the rights are unclear.Describe each image in text and confirm you may use it.
Channel schemaYou mapped the jacket's fields to one named channel's documented schema.No mapping exists yet.Pick one channel and map its required fields to where they live.
Validated interfaceA sample of records passed that channel's validation and every error was reviewed.The interface was never tested on real records.Validate a sample and give every error an owner.
A description of one record for one named channel, layer by layer. It makes no prediction about ranking, traffic or sales.

3 gaps appear in almost every first check. Color and material sit in the title instead of typed fields. Sizes are written in the description instead of being modeled as variants. And price is copied onto the product record, where it goes stale as soon as one seller changes theirs. Each has a first step that a catalog team can take without waiting for any channel.

  1. Pick one record that sells well and one channel you care about.
  2. Answer the 9 questions for that record, in words, with evidence.
  3. Fix the first band before the second, and the second before the third.
  4. Map the record's fields to the channel's documented schema, one by one.
  5. Send a sample through the channel's validation and give every error an owner.

Write the result down in words for each layer: in place, partly in place or not started, with the evidence. The words matter more than any summary, because they tell the next person exactly what to fix and where.

What readiness does not guarantee

A ready record is a necessary condition, not a sufficient one. Channels decide what they show, to whom and in what order, using signals well beyond product data. A complete record makes it possible for an agent to understand your product; it does not make an agent choose it.

  • Readiness describes what is in place today for one named channel; it is not a score.
  • It makes no prediction about traffic, ranking or sales.
  • It does not replace human approval of what is published.
  • It needs re-checking whenever the channel's requirements or your catalog change.

It is also worth deciding what you want agents to do before any of them can. Some businesses prefer agents to send shoppers to their own checkout; others are open to agents completing a purchase. Those are commercial decisions about each channel's terms, separate from whether your data is ready.

Youzu does not make a catalog visible in AI assistants and does not let agents buy from you. No ranking is promised. What Youzu does is upstream: Youzu helps marketplaces, classifieds and retailers turn raw seller and supplier product data into structured, reviewable catalog records, the kind of records the layers above describe.

Boundaries

What Youzu does not do

  • Youzu does not make a catalog visible in AI assistants and does not let agents buy from you.
  • Youzu does not guarantee ranking or visibility in search engines or AI assistants. No ranking is promised.
  • Youzu does not replace your ERP, search or storefront; those systems stay authoritative for their own data. It can replace your PIM, or work alongside it.
How an evaluation runs

Checklist

One-record readiness check

The 9 layers as rows, each with its question and the first step to close a gap, plus empty columns for your channel, your evidence and the owner. It is a description, not a score.

Download the record check (CSV)

A CSV file of 9 rows, named youzu-agent-readiness-record-check.csv. No form to fill in first.

One-record readiness check, scrollable
BandLayerQuestionFirst step if not in placeYour channelEvidenceOwner
Machine-readableAccessible recordCan a machine reach the full product record without scraping a page?Expose the record through a feed or API that your team can document.
Machine-readableStable identityDoes each product keep one identifier across sources and updates?Pick one identifier per variant and send it every time.
Machine-readableStructured attributesAre the attributes a buyer filters on present, typed and consistent?Move facts out of titles into the fields your category defines.
GovernedSupported variantsAre variants modeled explicitly rather than described in free text?Model variants explicitly and link them to one product.
GovernedCurrent offerIs offer information current, and do you know how stale it can get?Keep offers apart from the product and record how old each can be.
GovernedPolicy contextDo category, compliance and moderation rules travel with the record?Attach the rule set for each category to the records it covers.
Channel-readyUsable mediaIs imagery consistent, rights-cleared and described in text?Describe each image in text and confirm you may use it.
Channel-readyChannel schemaHave you mapped your record to one named destination's schema?Pick one channel and map its required fields to where they live.
Channel-readyValidated interfaceHas that interface been validated against a real sample, with errors reviewed?Validate a sample and give every error an owner.

FAQ

Questions buyers ask

Short answers to the questions teams ask about this topic.

Start with identity and attributes. Give each product and variant one stable identifier, move facts such as color and material out of titles into typed fields, and keep price, stock and condition on the offer rather than the product. Clear, consistent product facts give search engines and shopping assistants better material to work with.

Usually not at first. The channels that agents read mostly start from the product data you already publish in feeds. Fix that data, map it to the named channel's documented schema and validate a sample. A separate integration becomes relevant only when a channel requires one, so read its current documentation before building anything.

No one can promise that. Channels decide what they show and in what order, using many signals beyond product data. A complete, consistent record makes it possible for an agent to understand your product; it does not make an agent choose it. No ranking is promised, and readiness is not a prediction of traffic or sales.

Agent Readiness is an explainer and checklist for whether product and offer information suits a named machine-facing channel; it gives no score and human approval stays. It groups the questions into 9 layers in 3 bands, and you can work through it for one record before involving anyone else.

No. Youzu does not make a catalog visible in AI assistants and does not let agents buy from you, and no ranking is promised. Youzu helps marketplaces, classifieds and retailers turn raw seller and supplier product data into structured, reviewable catalog records, which is the product data that machine-facing channels read.

Last reviewed

Sources

  1. Google Universal Commerce Protocol (UCP) Guide, Google for Developers. developers.google.com/merchant/ucp. Accessed . Back

  2. Universal Commerce Protocol (UCP), ucp.dev. ucp.dev. Accessed . Back

  3. What is the Model Context Protocol (MCP)?, Model Context Protocol documentation. modelcontextprotocol.io/docs/getting-started/intro. Accessed . Back

  4. Product data specification, Google Merchant Center Help. support.google.com/merchants/answer/7052112. Accessed . Back

About this guide

Published by Youzu · Editorial policy

Published

Every guide defines its terms, works through one example and cites the public sources it relies on, with the date each source was checked.

Guides explain methods. They do not rank Youzu above other approaches, and an example is never presented as a customer result.

More in all guides and the catalog glossary.

Next step

Automate Commerce. Make shopping personal.

Bring your catalog, listing flow or shopping idea. Let's find the right place to start.

Book a demo

Tell us about your catalog and we will get back to you.