The 27 Steps to Validate a Product Catalogue for Agentic Commerce

AI agents do not browse your website. They read your product feed. If your feed is incomplete, broken, or out of sync, your products do not exist for them. This guide is the operational checklist we use at L'Atelier Growth when we audit a catalogue for agentic commerce readiness. Twenty-seven concrete steps grouped into six categories.

The cost of a missing field is no longer a lower CTR on Google Shopping. It is being skipped entirely by ChatGPT, Gemini, and Perplexity when the agent decides what to recommend.

Agentic commerce changes the rules of product visibility. The agent does not give a brand the benefit of the doubt. It compares structured fields across competitors and picks the cleanest, most complete, most trustworthy entry. A small merchant with a pristine feed can outrank a Walmart entry with sloppy data on the same query.

Three forces compound this. First, protocols like Google's Universal Commerce Protocol standardize what fields the agent expects. Second, OpenAI's Apps SDK does the same on the ChatGPT side. Third, merchants who skip these protocols become invisible to entire ecosystems.

According to Google's NRF 2026 announcements, agents will reach hundreds of millions of users by the end of 2026. The catalogues that are ready will scale. The ones that are not will lose share quietly without ever knowing why.

Each category builds on the previous one.

• Identification (1 to 5): decides whether an agent can even recognize your product exists.
• Content quality (6 to 11): decides whether the agent understands what it is.
• Pricing and availability (12 to 16): decides whether the agent trusts the offer.
• Shipping and delivery (17 to 20): filters out offers that cannot fulfil.
• Trust signals (21 to 24): determines how you rank against a competitor with a similar product.
• Protocol integration (25 to 27): exposes all of the above to the agents that actually matter.

Skipping a category earlier in the chain breaks everything downstream. A perfect protocol integration on a catalogue with missing GTINs still loses to a Walmart entry that has both. Fix in order. Do not treat the 27 as a grab bag.

Every product must be uniquely identifiable. Without clean identifiers, agents cannot deduplicate, compare, or reconcile inventory across surfaces.

1. 1.Every product has a unique, stable SKU that does not change across time.
2. 2.GTIN, UPC, EAN, or ISBN is filled when the product has one. Unidentified branded products are downranked.
3. 3.Brand field is populated and matches the official brand name exactly, including capitalization.
4. 4.MPN (manufacturer part number) is filled for branded items with no GTIN.
5. 5.Product type and Google product category are both set, with at least three levels of taxonomy depth.

Agents read titles and descriptions as natural language. Vague titles get skipped. Stuffed titles get penalized.

1. 6.Titles include brand, product type, and primary attribute (size, color, model) in a natural readable order.
2. 7.Titles stay under 150 characters and contain no marketing fluff like "best deal" or "limited offer".
3. 8.Descriptions are at least 300 characters of plain prose that explains the product, its use, and key benefits.
4. 9.Descriptions contain no HTML markup, no emoji, no all-caps shouting.
5. 10.Bullet points are stored as structured attributes, not as a wall of text inside the description.
6. 11.Each product has at least one high-resolution image (minimum 800x800 px) on a white or neutral background.

Stock and price accuracy are non-negotiable. An agent that recommends an out-of-stock product loses trust in the merchant immediately.

1. 12.Price field is populated with the current selling price, in the correct currency, including taxes where required.
2. 13.Sale price is filled when applicable, with a sale_price_effective_date when the discount has a deadline.
3. 14.Stock status is updated in real time or near real time (under 15 minutes lag).
4. 15.Quantity available is exposed for low-stock items so the agent can prioritize accurate offers.
5. 16.Variants (size, color) each have their own stock and price, not aggregated at the parent level.

Agents need to answer "can it arrive by Friday in Paris" without guessing. Structured shipping data turns this from a question into a filter.

1. 17.Shipping cost is exposed per zone, per weight, and per delivery speed, not just as a flat rate.
2. 18.Estimated delivery time is filled per zone, with min and max business days.
3. 19.Free shipping thresholds are exposed when applicable, with currency and zone.
4. 20.Click and collect availability is filled for merchants with physical stores.

Trust used to be a brand asset. Now it is a data field. Agents read it and rank with it.

1. 21.Customer reviews are exposed at the product level, with average rating and review count.
2. 22.Return policy is exposed in plain text per product, with return window and return cost.
3. 23.Warranty information is filled when relevant, with duration and scope.
4. 24.Certifications (organic, fair trade, energy class) are exposed as structured attributes.

The cleanest feed in the world stays invisible if it is not connected to the surfaces where agents look.

1. 25.Google Merchant Center feed is active, validated, and free of disapprovals.
2. 26.At least one agent protocol is integrated: UCP for Gemini, Apps SDK for ChatGPT, or MCP for Claude.
3. 27.A monitoring routine catches feed errors within 24 hours, with alerts wired to whoever owns the catalogue.

The catalogue becomes selectable across every major agent surface. ChatGPT can recommend it inside conversations. Gemini can pull it through UCP. Perplexity Shopping picks it up. The merchant stops fighting for visibility one ad at a time and starts being chosen automatically when the request matches.

This is also the foundation that makes other growth investments compound. AI citation optimization works better when products mentioned in editorial content actually exist in clean form in the catalogue.

For a catalogue under 1,000 SKUs with clean source data, the full 27-step implementation runs 4 to 6 weeks. The bottleneck is never the spreadsheet work. It is reconciling source systems (ERP, PIM, Shopify) that store the same field under three different names, and getting product teams to write structured return policies when they have been shipping PDF attachments for years.

For a catalogue above 5,000 SKUs, plan 2 to 3 months. The 27 steps are the same, but automation becomes mandatory. You need scripts that enforce completeness, not manual checks. This is typically where a partner is non-negotiable. Carrefour reached 26 million French ChatGPT users after investing in exactly this layer before the integration, according to Carrefour's own announcement. The feed work was what unlocked the distribution, not the other way around.

Most catalogues fail on the same three blocks. Trust signals are the most neglected, with return policies hidden in PDFs instead of being structured. Real-time stock is the second, with sync delays of 30 minutes or more breaking agent trust. And protocol integration is the third, with merchants relying on Merchant Center alone and being completely absent from ChatGPT and Claude.

A fourth failure mode is rising fast: variant-level attribute drift. A parent product has a clean entry but the variants inherit incomplete or stale data. When an agent resolves "blue medium", it hits the variant, not the parent, and rejects the offer. Parent-level cleanliness does not absolve variant-level gaps.

A fifth is silent protocol rejection. A feed validates in Merchant Center but silently fails UCP completeness checks. Merchants only discover the gap weeks later when competitors start being selected and they are not.

Fixing these five categories alone moves a catalogue from invisible to selectable in most cases.

This is the work we build at L'Atelier Growth. Catalogue audit, feed restructuring, and protocol integration across UCP, Apps SDK, and MCP. We turn product data into a system agents can read, trust, and recommend.

This is not consulting. These are systems we design, build, and operate. Contact L'Atelier Growth to audit what you have today and ship what is still missing for the agent layer.