Deciding whether your product catalog can support AI shopping agents starts with testing it against real buying questions, not counting how many fields are filled in. A catalog built for keyword search and human browsing can still fail an agent that needs to reason about suitability, relationships and current availability.
A practical checklist to run before claiming your catalog is agent-ready, companion to our Product Intelligence guide.
- Agent-readiness is a set of testable behaviors, not a single yes-or-no property of your catalog.
- Test identity resolution, verified attributes, relationships, current information, missing-data handling and permission boundaries separately.
- A catalog can pass some checks and fail others; address the gaps that matter to your actual shopper requests first.
Why isn't a complete product feed enough?
A product feed with every field populated tells you that data exists. It does not tell you whether an AI agent can correctly identify the right variant, explain why a product suits a stated need, or recognize when a fact is missing rather than guessing one. Readiness is about behavior under real questions, not completeness of a spreadsheet.
Run each of the following checks against your actual catalog before concluding you are ready, and treat failures as a prioritized backlog rather than a single pass-fail grade.
Checklist item 1: Product identity resolution
Ask a system built on your catalog to find a specific product by a vague or partial description, not just its exact name or SKU. Then ask about a similar but different product and confirm it does not conflate the two.
- Does it correctly distinguish between similar variants (size, color, model year)?
- Does it avoid returning a near-match when the exact requested item exists?
- Can it explain which variant it selected and why?
Checklist item 2: Verified attributes, not inferred ones
Ask about a specific material, dimension or ingredient fact. Then ask about a fact your catalog does not actually record.
- Does it answer accurately from the source record when the fact exists?
- Does it say the information is unavailable, rather than inventing a plausible answer, when the fact does not exist?
- Can you trace an answer back to the specific source field?
Checklist item 3: Relationships and suitability
Ask for a complementary or compatible product, and for an alternative to something out of stock or over budget.
- Are complementary suggestions grounded in an actual stated relationship, not just frequently-bought-together statistics?
- Does a suggested alternative genuinely meet the original requirement, or just resemble the original product?
- Does the system distinguish between a substitute and a complement when explaining its suggestion?
Checklist item 4: Current information
Ask about price, stock level or an active promotion, then change the underlying data and ask again.
- Does the answer reflect the current state, not a cached or outdated version?
- Is there a clear, known refresh interval for the data sources the system draws on?
- Does an out-of-stock item get correctly excluded from availability, not just flagged after the fact?
Checklist item 5: Permission boundaries
Ask the system to perform an action beyond its intended scope, such as applying a discount it has no authority to create, or accessing account information without proper authorization.
- Does it decline clearly, explaining the boundary, rather than attempting the action or inventing a workaround?
- Is the distinction between discovery, inventory lookup and purchase-related actions enforced, not just described in documentation?
- Who is responsible for reviewing an attempted boundary violation when it happens?
Checklist item 6: Journey continuity
Ask a multi-part question, then add or change a requirement partway through, as a real shopper might.
- Does the system retain the original requirements it should keep, while updating the ones that changed?
- Does it ask a clarifying question when a change creates ambiguity, instead of guessing?
- Does context survive moving from a general question into a specific product's details?
How should you score and prioritize the results?
Treat each checklist item as pass, partial or fail, and attach it to a concrete example from your own catalog, not an abstract judgment. A partial result on relationships might matter far less than a fail on permission boundaries, depending on whether you plan to support any purchase-related actions for external agents at all.
Prioritize fixes according to the stages (discovery, lookup, cart, checkout) you actually intend to support, described in our introductory guide to agent-to-agent commerce. A retailer only supporting discovery and lookup can deprioritize deep permission-boundary testing around checkout, for now, while still testing it for inventory and account-adjacent requests.
Bayezon AI's Product Intelligence and merchandising governance are built to address exactly these checks, connecting verified product facts, relationships and retailer rules so they hold up under this kind of testing rather than only looking complete on paper.
Run this checklist against your own catalog
Test identity resolution, relationships, current data and permission boundaries using real questions from your retail category.
FAQ
Is a complete product feed the same as an agent-ready catalog?
No. A complete feed means fields are populated. Agent-readiness means the system built on that feed behaves correctly under real buying questions, including ones involving missing data and permission boundaries.
What is the single most important check to run first?
It depends on what you intend to support. If you plan any purchase-related external-agent access, permission boundaries deserve priority. If you are only supporting discovery, relationships and current information matter more.
Can a catalog be partially agent-ready?
Yes, and this is the normal state for most retailers starting out. Treat the checklist as a backlog, not a single pass-fail test.
How often should this checklist be re-run?
Re-run it after any significant catalog restructuring, a new product category launch, or before expanding which agents or actions you authorize.
One last thing
The most revealing test is usually the simplest: ask about a fact your catalog genuinely does not contain, and see whether the system admits it does not know.
