In AI-Era Retail, the Catalog Became Content — and a Stale Price Became a Brand Risk

That's the sharpest difference between an operation ready for generative search and one that merely has a good-looking website. It's also the area where the work is most consistently underestimated, because it looks like configuration when it's actually data architecture.

Schema: What Google Recommends Specifically for Commerce

Google recommends Product and Offer for merchant listings and shopping experiences, plus ProductGroup for variants and Organization / LocalBusiness for the brand and physical-store layer.

One detail that tends to slip past teams: Google states that structured data improves the accuracy of how price, discount and shipping are understood, and that Merchant Center may use that information in feed verification. In other words, schema isn't just a request for a rich result — it's an input to verification.

The rule underpinning all of it: structured data must stay consistent with the visible text. Markup that declares a price different from what the page displays isn't optimization, it's inconsistency — and inconsistency is precisely what erodes an engine's trust in you as a source.

On Open Graph, worth repeating: it complements sharing metadata and supports unfurls and secondary scrapers, but it does not replace schema or Merchant Center.

Formal Feeds: The Two Paths That Matter Today

Google Merchant Center

On Google, Merchant Center feeds remain the path with the deepest commercial reach, and they're mandatory for some surfaces, such as the Shopping tab. This isn't an alternative to technical SEO — it's a distinct layer with its own rules, distributing your catalog to Search, Shopping, Images, Lens and Maps.

There's no cost for organic listings; paid media is optional.

OpenAI ACP

In OpenAI's Agentic Commerce Protocol, the feed is the starting point of the integration: it gives ChatGPT the catalog data it needs to index products, understand core attributes and display correct price and availability. ACP also distinguishes eligibility for discovery from eligibility for checkout through its own flags.

There's an architectural point worth getting right, because it's frequently explained badly: under ACP, checkout is rendered in the ChatGPT interface, but the real order state, payment processing and the decision to accept or reject the order remain in the merchant's systems — which also assess risk and fraud in their own stack.

That isn't a technical footnote. It's the guarantee that the retailer isn't outsourcing the commercial decision or the anti-fraud policy to a conversational layer.

Adjacent Surfaces: Why This Is Bigger Than It Looks

Google documents that product data can appear not only in Search, but also in Google Images, the Shopping tab, Lens, Maps and Business Profile.

For omnichannel operations, that creates a direct bridge between GEO, local SEO and inventory visibility. An ecommerce business with a physical store or click-and-collect gains far more coverage when it connects catalog, store, Business Profile and local inventory — because the same entity and product data feeds several surfaces at once.

The trajectory is expansion: generative engines will keep advancing into recommendation by image, video, voice and context state. That raises the stakes on product media, explainer video, visual attributes, and consistency between feed, page and multimodal representation.

The Problem That Destroys the Most Value: Price and Stock Out of Sync

This is the part that deserves attention at board level, not just from the technical team.

Google explicitly documents that site data and Merchant Center can become inconsistent when the site changes before the feed, and recommends automatic updates and greater control over feed timing to reduce the conflict.

In GEO, that becomes a trust problem of a different magnitude. A generative engine quoting an outdated price may well drive traffic, but it destroys conversion and brand perception. The customer arrives convinced of a price that doesn't exist. The frustration doesn't get attributed to the engine — it gets attributed to you.

The architectural consequence is direct: any commercial RAG architecture needs a short TTL for volatile attributes, prioritization of fresh documents, and mechanisms to invalidate embeddings and context caches whenever price, availability or promotion changes.

Four Repositories, Four Rules

From a data-flow perspective, the ideal is to separate at least four logical repositories, each with its own rule.

The most important rule in that table is the last row, and it's absolute: identity context never enters the public corpus. Mixing public content and personal data in the same index, with no isolation, is the most serious architectural failure in this space — and it's the subject of a whole separate article in this series.

Answer-Oriented Content: What Changes on PDPs

With the catalog structured, the editorial layer gets a clearer objective. Citable modules built by intent outperform generic copy:

• →Comparisons — this model against that one, with explicit criteria.
• →Compatibility — what works with what.
• →Measurements and specs — numbers, not adjectives.
• →Maintenance and durability — how to care for it, how long it lasts.
• →Policies — exchanges, returns, delivery windows, warranty, in plain language.
• →Reviews — with appropriate markup.

The test is always the same: is the information objective, verifiable and attributable? If the answer is "it's nicely written, but there's nothing to quote," it isn't working for you in generative search.