Agent-Ready Product Data

Activate when: user says 'make our products buyable by AI shopping agents', 'is our product data agent-ready?', 'agentic commerce', 'the agent failed to place the order', asks 'why doesn't ChatGPT / a shopping agent recommend or buy our products?', is preparing a product feed for an AI assistant or an agent checkout protocol, or wants a catalog that AI agents can read, compare and purchase from. Do NOT activate when: the goal is only how often the brand is mentioned inside AI answers (use geo-generative-engine-optimization); there is no online purchase channel; the business is offline-only or quote-based B2B; the category may not legally be sold through automated channels. More: deciqai.com/c/agent-ready-product-data

Overview

An AI shopping agent does not respond to photography, mood or persuasive copy. It reads structured attributes, unambiguous variants, and machine-readable price, stock, shipping and returns. What it cannot parse, it skips; what it parses wrongly, it buys wrongly — and the merchant pays for the return and the complaint.

The core mechanism: the agent funnel is Seen (retrieved) → Understood (attributes and variants are unambiguous) → Bought (price, stock, policies and checkout have no blockers), and each layer is measured and repaired separately. A product that is perfectly described but invisible, or visible but uncheckoutable, loses the sale to the competitor whose data is more complete.

Compose with geo-generative-engine-optimization (getting recommended) · checklist (field completeness as a hard checklist) · theory-of-constraints (fix the narrowest funnel layer first).

When to Use

• Products appear in AI answers, but with stale information, the wrong variant, or the wrong price
• Test orders through an agent stall at variant selection, shipping cost, forced login or a CAPTCHA
• Preparing to submit a product feed, adopt an agent checkout protocol, or list in a platform catalog read by AI assistants

When NOT to use: no online transaction · products require a custom quote · categories where regulation bars automated sale · the problem is brand mentions, not catalog data.

Coaching Novices (Adaptive Front Door)

• Engine mode: user has concrete SKUs and a channel → run The Process directly.
• Coach mode: user is unfamiliar → guide one step at a time.

In Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.

1. What it is: "An agent only trusts fields it can parse. One missing field can cost one order."
2. Check fit: "Can a customer buy your product online, directly, without a quote?"
3. Elicit the real case: "Pick 3 best-selling SKUs. Send the product-page links and the attributes you most want an agent to understand."

[WAIT — do not advance until user responds]

4. Run The Process: "Run an agent through find → compare → add to cart → checkout on each SKU and record every point where it stalls."

[WAIT — do not advance until user responds]

5. Name the insight: "Your loss sits in one layer — Seen, Understood or Bought. Fix that layer first."

The Process

Output: SKU field audit + measured agent funnel + repair list

1. Field audit. For every SKU: title, brand, GTIN/MPN, category, material, dimensions, compatibility, intended user. Check against the target channel's published feed specification and schema.org Product. Gate: required-field missing rate is zero → otherwise fill fields first.
2. Disambiguate variants. Every color / size / bundle is its own identifiable variant with its own ID, price, stock status and image; attribute names and values are standardized, with units. Gate: any single variant can be located uniquely → otherwise split it.
3. Machine-readable transaction data. Price (with currency and promotion window), availability, shipping cost and time, and return policy are expressed in the feed and in structured data (Offer, OfferShippingDetails, MerchantReturnPolicy), and match the page. Gate: feed, schema and page agree → otherwise regenerate all three from one source of truth.
4. Policy and trust pages. Returns, warranty, privacy, terms and contact pages are public, crawlable, explicit, and need no login.
5. Cut checkout friction. Allow guest checkout and standard payment methods; remove forced registration and unnecessary verification; evaluate the agent checkout path the channel or protocol offers, and treat it as replaceable — channel checkout models change. Gate: never bypass fraud controls or payment security for speed.
6. Test with real agents. Use 3+ shopping agents or assistants, prompted the way buyers prompt, through the full flow. For each SKU record: Seen? Understood (attributes and price correct)? Bought? Gate: ≥3 runs per layer → otherwise no conclusion.
7. Repair and re-test. Fix the layer with the largest loss first; re-test in 4 weeks.
8. Stop-rule: data is complete and the product is still not Seen → the problem is no longer catalog data; move to brand and third-party evidence work (geo-generative-engine-optimization).

Output template: SKU | missing fields | variant issues | price/stock/shipping/returns consistent? | Seen rate | Understood rate | Bought rate | blocker | fix | re-test date

→ Method in Action: ChatGPT shopping and the Agentic Commerce Protocol (2025–2026)

Applying It Well

1. One source of truth. Generate page, feed and schema from the same record so they cannot drift apart.
2. Write attributes for machines. Standard units and enumerated values, not "extra large" or "slimming".
3. Unknown is not zero. Omit a value you do not know rather than filling a placeholder; an omitted shipping charge reads as unknown, not free.
4. Test beats inference. Re-run the agent flow after every catalog or site change.
5. Separate the durable layers from the volatile one. Clean attributes and variants serve every channel; a specific checkout integration may be withdrawn.

→ Primary sources: references/sources.md

Common Rationalizations

[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.

Fake move	Reality
[D] "If a human can read the page, that's enough"	Agents parse fields; they do not read marketing copy or text baked into images.
[D] "Variants live in one dropdown on one page"	A variant without its own ID, price and stock is a variant the agent will pick wrongly.
[D] "Shipping is calculated at checkout"	An unknowable total pushes the agent toward a competitor whose total is transparent.
[D] "We installed a schema plugin, so we're done"	Plugins routinely miss fields or disagree with the page. Audit and test.
[D] "We integrated the channel's checkout, so we're agent-ready"	Checkout is the most volatile layer. If attributes and variants are weak, the product is still not Understood.
→ Add [O] entries here after each real use — paste the actual failure pattern	What went wrong and why

Red Flags

• No GTIN/MPN · feed price differs from page price · return policy behind a login · forced account creation at checkout
• Placeholder values ("n/a", "unknown") in required fields · variants sharing one ID · never tested with a real agent

Verification

• Required-field missing rate is zero
• Every variant is uniquely locatable (own ID, price, stock, image)
• Feed, structured data and page agree on price, stock, shipping and returns
• Seen / Understood / Bought rates measured with ≥3 runs per layer
• Re-test date scheduled

---

Part of deciqAI Knowledge Skills — 253 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. See it run → https://www.deciqai.com/c/agent-ready-product-data · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.

Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/agent-ready-product-data.json