Your Product Feed Is the Homepage

OpenAI is expanding product discovery in ChatGPT through richer merchant data, direct integrations, and Shopify Catalog. Retailers including Target, Sephora, Nordstrom, Lowe's, Best Buy, The Home Depot, and Wayfair have integrated with its Agentic Commerce Protocol for discovery.

An AI shopping conversation begins with intent that is more specific than a typical category-page visit. The shopper can combine budget, use case, fit, timing, materials, compatibility, and policy preferences in one request. The agent then translates those constraints into a small set of products.

OpenAI says product results are selected for relevance. Its commerce guidance says merchant ranking can consider availability, price, quality, primary-seller status, and checkout capability. Its shopping help guidance also describes the use of product and merchant metadata, reviews, price, and availability.

The product page still matters. Shoppers need a place to validate the recommendation, inspect details, understand the brand, and complete any step that the agent cannot handle. Yet the product page may arrive after the most consequential filter.

A homepage once framed the category and guided the buyer into the catalog. An agent can now enter through the data layer and skip that framing. Brands need to make their strongest product truths legible before the click. This is the next operating layer behind the broader shift in shopping influence.

Many feed programs grew up as compliance work. The goal was to avoid disapprovals and keep price and availability correct enough to stay listed. Agentic discovery raises the standard.

Those fields are the floor. The specification also supports product attributes, variant relationships, sale pricing, return details, and review aggregates. A minimal row can establish existence. A rich row can establish fit.

A rich row also gives brand strategy somewhere concrete to live. If a company wins because its replacement parts are available for ten years, the service period should exist as governed data. If a material is the source of comfort or durability, the catalog should name it precisely. If a product is easier to repair, the parts, instructions, and warranty terms should make that advantage retrievable.

Consider a skincare product for someone asking for a fragrance-free mineral sunscreen that works under makeup and ships before a trip. “Premium daily protection” offers little help. The useful evidence is concrete: active ingredients, SPF rating, fragrance status, finish, skin-type guidance, package size, testing claims, current inventory, delivery range, and return policy.

This is persuasion expressed as verifiable attributes. It asks marketing to carry the evidence behind a claim instead of relying on promotional adjectives.

Every unanswered question adds friction to the recommendation. A system can ask the shopper for more detail, yet it cannot invent a missing warranty term or safely infer a product dimension from brand copy.

Human shoppers already treat operational facts as part of the offer. Baymard's product-page research reports that 64% of users look for shipping costs on a product page and 60% look for return information. Those studies concern human behavior, yet they reveal which facts carry decision weight.

An agent compresses that evaluation. It may compare several products, reconcile constraints, and provide a reason for each recommendation. When one merchant exposes delivery, returns, and variant availability clearly while another does not, the second product creates more uncertainty.

OpenAI acknowledges that shopping research can make mistakes about details such as price and availability. Direct, current merchant data helps reduce the error surface. It does not remove the need for a merchant-owned source of truth.

Google's product guidance points in the same direction. It recommends providing Product structured data on product pages and a Merchant Center feed because the two can expand eligibility and help Google understand and verify product information. Shipping, returns, ratings, availability, and variants all have structured roles.

Consistency matters because outside systems may combine more than one source. A channel can read the merchant feed for price, the page markup for a policy, and the landing page for supporting context. When those surfaces disagree, the brand creates a verification problem exactly where a recommendation system needs confidence.

Product facts often cross several systems before an agent sees them. A product information system stores attributes. The commerce platform controls price, inventory, and variants. The product page renders copy and structured data. Merchant feeds transform fields for external channels. Agentic storefronts or marketplaces apply their own mappings.

Marketing may own the description, merchandising may own the taxonomy, ecommerce may own the product page, operations may own inventory, and legal may own policy language. Each team can be locally correct while the external record becomes inconsistent.

A brand with five conflicting versions of a product has one unresolved data problem appearing in five places. Product-feed optimization needs to begin upstream with definitions, ownership, and evidence.

In an OpenAI case study, Wayfair reported correcting 2.5 million tags across more than one million visible and frequently purchased products. Jessica D'Arcy, Wayfair's associate director of catalog merchandising, connected better data quality with customer trust, better buying decisions, and fewer returns caused by misrepresented products.

A color, material, dimension, or compatibility field looks small inside a spreadsheet. Across discovery and recommendation systems, it becomes a promise the brand needs to keep.

The strongest program begins with buyer questions and ends with governed data. Six steps make that real.

1. 01

   Collect high-intent questions

   Use site search, support logs, reviews, sales conversations, marketplace questions, and AI referral landing pages. Capture the constraints buyers actually express.

2. 02

   Turn claims into evidence

   List the promises that matter for each category. Define the test, policy, record, or operational condition that makes each promise true.

3. 03

   Map evidence to fields

   Decide where each fact lives, which format it uses, how it reaches every channel, and who owns its meaning.

4. 04

   Treat variants as products

   Give selected variants stable IDs and their own price, availability, URL, images, and qualifying attributes.

5. 05

   Synchronize volatile facts

   Set freshness targets for inventory, price, promotions, delivery, and eligibility. Measure the lag from source change to external availability.

6. 06

   Test real decisions

   Run representative shopping prompts, inspect the candidates and explanations, then classify every failed match by cause.

The marketing sentence is the visible surface. The evidence model is the reusable asset. “Dishwasher-safe” may require a test standard and care instructions. “Works with Model X” needs a maintained compatibility record. “Arrives tomorrow” depends on inventory, cutoff time, fulfillment location, and carrier logic.

Variant discipline deserves special attention. OpenAI's specification calls for separate rows for selected variants, stable item IDs, a shared group ID, and each variant's own price, availability, URL, and images. A shoe in size 12, a laptop with 32 GB of memory, and a sofa in a particular fabric are distinct decisions even when merchandising treats each set as one product family.

The same distinction applies to freshness. Product titles and materials may move through governed review, while inventory and price can change many times in a day. Promotions require start and end dates. Shipping promises can vary by location and cutoff time. Set an update target by field type, then monitor the lag between a source change and the value an outside channel receives.

Shopify's product-discovery guidance treats catalog data as authoritative and says inventory and pricing are continuously updated. It also recommends previewing raw catalog search while warning that external channels can re-rank results. A feed can pass validation and still lose the decision, so teams need both data QA and real-query testing.

Traditional ecommerce reporting starts when a session reaches the site. Agentic discovery creates meaningful activity before that point. A product can be considered, compared, rejected, or recommended without producing a pageview.

Adobe reported that AI-referred retail visitors converted 42% higher than non-AI traffic in March 2026. The benchmark is encouraging, though it does not establish a universal result. It describes the visible visitors who clicked through, not every upstream recommendation decision.

Connect those operational measures to AI-referred sessions, assisted conversions, product-page conversion, returns, support contacts, and margin. Shopify's agentic storefront reporting already exposes channel-level sales, orders, sessions, and conversion for supported surfaces.

Keep a reason code for each failed recommendation test. The product may be ineligible, missing an attribute, stale, mapped into the wrong category, supported by weak evidence, or simply a worse fit than another option. Those causes belong to different owners. A clean diagnosis keeps the team from treating every loss as a copywriting or ranking problem.

The crucial shift is to measure whether a product was selectable before measuring whether a visitor clicked. Traffic is a late signal when the shortlisting happened elsewhere. This is another expression of the AI search measurement gap.