How to test an "agent-ready" claim

Vendor Watch exists to test claims rather than vendors, and the claim in front of every retail buying committee this year is “agent-ready”. It is unusually testable, because the specifications are public. Below are the questions I would put to any platform, PIM vendor, marketplace tool or implementation partner making the claim, and what a poor answer indicates.

No vendor commissioned, reviewed or paid for this piece. Vendor Watch cannot be sponsored — see the editorial and advertising policy.

1. Which protocol, by name and version? There is no single standard. The Agentic Commerce Protocol from OpenAI and Stripe covers merchant product feeds and checkout. Google’s Agent Payments Protocol, launched in September 2025, addresses agent-initiated payment authorisation and was donated to the FIDO Alliance in April 2026. Visa’s Trusted Agent Protocol handles agent identity so merchants can distinguish legitimate agents from bots. Mastercard’s Agent Pay for Machines, launched 10 June 2026 with more than thirty partners, targets programmatic agent-to-agent payment. These solve different problems. “Agent-ready” without a named protocol is not an answer.

2. Can it produce the full required feed today, including the awkward fields? The Stripe and OpenAI product feed spec requires id, title, description, a link that resolves with HTTP 200, brand, image_link, availability, price, product category and shipping. It requires mpn where there is no GTIN, and item_group_id wherever variants exist. Ask specifically about variant grouping and product_warning, the compliance field described as mandatory for items with regulatory warning requirements. Those are the two that reveal whether an integration is real.

3. What is the update latency, end to end? The spec states the system accepts updates every fifteen minutes, and OpenAI’s guidance recommends a full feed at least daily plus API updates through the day. The question is not what the channel accepts but what the vendor’s own pipeline delivers from a price change in your ERP to a published feed. Answers measured in hours are common and are worth knowing before signature.

4. Are the recommended commercial signals supported, and where do they come from? The spec recommends popularity_score, return_rate, product_review_count and product_review_rating. Return rate in particular has to be calculated from your own transaction data over a defined window. Ask who computes it, from which system, and whether you control the window. A vendor that has not thought about this has not built it.

5. Can it show measured machine-readability, not a demo? Adobe’s content visibility scoring found retail sectors ranging from 63% down to 47% machine-readable on high-value page content. Ask for a before-and-after on a real catalogue, with the method stated. A screenshot of a chat assistant naming the client’s product is a marketing artefact, not evidence.

6. What happens to agent traffic at your edge? Visa built the Trusted Agent Protocol precisely because merchant bot-detection was blocking legitimate shopping agents. Feed compliance is wasted if your WAF or bot manager refuses the agent at the door. This question usually belongs to a different vendor than the one answering the other five, which is the point of asking it.

One caution on the evidence. Vendor analyses of which product attributes win AI recommendations are proliferating, and some are interesting — Profound’s analysis of holiday shopper prompts, for instance, reports that product research and comparison prompts made up 68% of the holiday set. But as of August 2026 there is no independent peer-reviewed study establishing which specific attributes cause a model to recommend one SKU over another. The defensible ground is narrower and better: the platforms publish the fields they ingest and rank on. Buy against the published specification, not against the inference.