AI shopping citations: what three retailers changed to earn them

🇹🇷 Türkçe: Bu yazının Türkçesini oku →

An e-commerce director I trust rang last month with the reasonable question: our organic traffic is fine, our paid is fine, our reviews are fine — why do the AI assistants never name us? She had already tried the marketing answer (more content), the SEO answer (fix the schema), and the ad-tech answer (buy into ChatGPT Shopping). None of it moved the needle. Sitting with the data, the answer turned out to be structural, and it had almost nothing to do with content.

The catalogues large retailers built for their own websites are not the catalogues an AI shopping surface reads. Two protocols shipped in the last twelve months, one from OpenAI and Stripe and one from Google and Shopify, and both push the entry ticket for citation down into a place most retail teams have never had to think about: the machine-readable product feed and the transactional endpoint an agent can call without a browser. Retailers that started rebuilding that layer in 2025 are showing up in the citation counts. Those that did not are watching brand and manufacturer sites take the click instead.

Three retailers are worth studying together — Walmart, Wayfair, Best Buy. Their responses are usefully different, and our clients ask us the same question that director asked.

Retailers are losing the AI shopping click to the brands they sell

Start with the number that reframes the whole problem. LLM Pulse analysed 391,073 citations across ChatGPT, Google AI Mode, Google AI Overviews, Gemini and Perplexity between April 12 and July 11, 2026, on 1,468 generic shopping queries in the “best X” / “where to buy Y” shape. Brand and manufacturer sites captured 64.7% of the citations. Reddit and YouTube together captured 10.4%. All of the major retailers and marketplaces put together captured 2.9% (LLM Pulse, “Where AI Actually Sends Shoppers”, 2026).

Inside that 2.9%, position matters more than volume. Amazon led with 1,114 citations at average position 9.67 — cited a lot but cited late. Walmart came second at 1,027 and position 5.72, materially earlier in the answer. Best Buy landed third at 562 citations, position 7.13. Target and Home Depot sat around 288 and 292 (same source). Being cited seventh is not the same as being cited first.

Two industry developments explain what these retailers have been doing about it. On September 29, 2025, OpenAI and Stripe released the Agentic Commerce Protocol, an open standard that lets an AI agent complete a purchase inside ChatGPT without handing the user off to a merchant site; OpenAI then shipped Instant Checkout in ChatGPT on February 16, 2026 (“Buy it in ChatGPT”, OpenAI). On January 11, 2026 at NRF, Sundar Pichai announced Google’s Universal Commerce Protocol, covering checkout, identity, order management, cart and product discovery, co-developed with Shopify, Etsy, Wayfair, Target and Walmart, and endorsed by 20-plus companies including Best Buy, The Home Depot, Macy’s, Zalando, Adyen, Stripe, Mastercard and Visa (NRF, January 2026).

Read those two together and the strategic picture is unambiguous. The retailers that are being cited are the retailers that made themselves callable by a software client, at the product and at the checkout, on protocols they do not fully own. That is a supply-chain decision as much as a marketing one, and it is the reason the three examples below are worth studying together.

Walmart is running a two-protocol bet on being everywhere at once

Walmart shows up as a co-developer of Google’s UCP alongside Shopify, Etsy, Wayfair and Target, and separately as one of the retailers Stripe and OpenAI cite among ChatGPT Shopping’s expansion partners. Very few retailers publicly committed to both stacks. That is not neutrality — it is the calculation that AI-mediated commerce will resemble payments more than search, so plugging into two open protocols is cheaper than betting on one.

The catalogue side of the move is quieter but arguably more consequential. In January 2026, Walmart shipped an update to its Supplier API and to the Item Setup and Maintenance feed schema, telling suppliers the “new feed files contain the latest product types, attributes, and values, to help suppliers improve content quality scores” (Walmart Supplier developer portal, January 2026). Plainly: Walmart is forcing tens of thousands of suppliers to describe their products more precisely, at the SKU level, in a machine-readable schema it controls. Those cleaner attributes are what an AI assistant sees when Walmart’s catalogue is queried through UCP or an agent surface.

• 1,027 citations from AI shopping engines, avg position 5.72 (LLM Pulse, Apr–Jul 2026).
• Co-developer of Google’s Universal Commerce Protocol, NRF, January 11, 2026.
• Launch-window retailer on Stripe/OpenAI’s Agentic Commerce Protocol.
• Supplier API + item feed schema refresh, January 2026, pushed down to every supplier.

The transferable lesson from Walmart is neither scale nor money: it is the sequencing. Fix the catalogue schema first, then plug into the transaction protocols. If the underlying data is thin, adding an agent-callable checkout does not improve what the agent picks up; it just makes the thin data available faster.

Wayfair rebuilt the pipes long before the protocols existed

Wayfair’s route is the most interesting one, because the work that has it inside UCP now is work it started in 2025 for a different reason. The company put its supplier integration through a full rebuild — new EDI file formats, a backend migration to a new platform, variation modelling changes that force each colour/size combination into a distinct feed relationship rather than a collapsed grouping (Wayfair Sell, “Adding Assortment”, 2025). None of that was pitched as an AI project. It was a supply-chain hygiene project.

The reason it now matters is that clean variation modelling is the single hardest thing to reconstruct after the fact. If an assistant asks for “a 60-inch walnut sideboard with cable management under $600”, the answer depends on whether the feed carries dimension, finish, feature and price as distinct, queryable attributes on each SKU — not as words buried in a description. Wayfair spent 2025 forcing that structure into its assortment; in 2026 it turned up as a UCP co-developer alongside Google and Shopify. The order of events matters. Retailers that treated product-data hygiene as a supply-chain problem, and finished it, walked into the protocol conversation as principals. The rest are still catching up.

• Full EDI file format overhaul and platform migration, 2025.
• Variation modelling reworked so each SKU carries queryable attributes (colour, size, dimension, feature) at the item level.
• Co-developer of Google’s Universal Commerce Protocol, January 11, 2026.

The transferable lesson: the AI-shopping citation you want in 2027 is being decided by product data work you either do or defer this quarter. There is no shortcut that begins with content marketing.

Best Buy is treating standardisation itself as the moat

Best Buy is not a UCP co-developer. It is on the endorser list — one of the twenty-plus retailers, payment networks and processors that signed on when Google announced UCP at NRF (Sundar Pichai’s NRF 2026 remarks). That is a deliberately smaller bet, and it is worth reading carefully.

Best Buy has been running structured product data — schema.org markup, GTIN discipline, category-specific attributes — for years, because in consumer electronics a mismatched attribute (wrong screen size, wrong port count) generates a return. Its 562 citations at position 7.13 in the LLM Pulse study suggest the existing catalogue is legible enough that becoming a protocol co-developer would not have paid back. Endorsing UCP costs less, signals intent, and keeps optionality open for whichever standard wins. If ACP holds on the OpenAI side and UCP wins on the Google side, Best Buy’s catalogue can surface on both without owning either.

• 562 citations from AI shopping engines, avg position 7.13 (LLM Pulse).
• Endorser (not co-developer) of Google’s Universal Commerce Protocol, January 11, 2026.
• Product data hygiene treated as a returns-reduction discipline, not a marketing programme.

The transferable lesson: if your catalogue is already in reasonable shape, resist the temptation to co-develop a protocol. Endorse the standards, keep your options open, and put the budget into the vertical-specific attributes that stop returns.

What the three responses have in common is that none of them is a content play

Every AI-search agency this quarter will offer to write you a fresh set of buying guides, add FAQ schema, and improve your content depth score. None of that is wrong; the LLM Pulse split (64.7% to brand and manufacturer sites) says long-form content on your own domain matters. But the retailers earning citations at the top of the AI answer are not there because of their content programme. They are there because their catalogue is structured cleanly at the SKU level, and they have made it callable by software over an open protocol. The entry cost used to be an SEO budget line; it is now a product-data engineering budget line, with an integration budget line next to it.

What we would ship first inside a WooCommerce store

For a real store we handle — mid-market, single-country, catalogue in the low thousands of SKUs — the sequence we run is boring, unglamorous, and cheaper than the SEO retainer it replaces. I will be specific about it because generic “improve your schema” advice has done enough damage already.

Product-data audit at the SKU level. One row per variation, not per parent product. Columns for GTIN, MPN, brand, price, availability, condition, dimensions, primary material, colour, size, and category-specific attributes (compatible model, screen size, port count). WooCommerce’s default product-variation model handles this if the attributes are set up as global attributes; it does not if they are stored as free-text in descriptions. Rebuilding that structure on a 3,000-SKU catalogue is a two- to four-week job.

schema.org/Product markup at parity with the feed. The JSON-LD on the product page should carry the same attributes as the Google Merchant Center feed, generated from the same source of truth. Divergence between the two is the single most common issue we find on WooCommerce stores that already invested in schema plugins; the plugin was written before Merchant Center’s attribute list expanded, and nobody keeps them in sync. Fix it once, and either write a validation step into CI or pay for a service that does.

A clean Google Merchant Center feed. GTIN coverage above 95%, zero disapprovals, image URLs stable, prices and availability accurate within the hour. Merchant Center is the data pipe most AI shopping surfaces read on the Google side; the ChatGPT side is starting to read it too (an early-2026 Search Engine Land study found roughly 83% of ChatGPT product carousel entries came from Merchant Center data). Clean feed, and you are visible on both surfaces before you touch any protocol.

A watching brief on ACP and UCP, not an integration yet. Both protocols are early and moving quickly. Instant Checkout inside ChatGPT is limited to specific platforms and geographies; UCP’s live surface is Google AI Mode and Gemini. Unless you run on Shopify or are a Stripe-first merchant with headline SKUs, there is no shipping-today integration that repays the engineering budget. There will be inside twelve months. Prepare the data now so the retailer next door does not have a six-month lead when the path opens.

Effort estimate for a typical mid-market WooCommerce store: six to ten weeks of engineering for the catalogue rebuild and schema parity work, ongoing feed maintenance thereafter, and a one- to two-week scoping engagement whenever an agent-callable surface becomes commercially viable for the vertical. A fraction of a content-led “AI SEO” retainer, and the part that shows up in citation counts twelve to eighteen months out.

The part that is not settled yet, and why we are honest about it

Two things nobody in this space credibly knows. First: whether ACP and UCP converge, coexist, or knock each other out. Both are open standards on GitHub, both are MCP-compatible on paper, both have serious retailer backers. History is not kind to duplicate open standards, but the twelve- to twenty-four-month interim asymmetry could hurt whoever bet on the losing side. Second: whether Amazon — 1,114 LLM Pulse citations at position 9.67, almost no presence inside ChatGPT or Perplexity — builds its own protocol, joins one, or keeps Rufus as a walled surface. If Rufus becomes the default for a meaningful minority of shoppers, the calculation shifts again.

What is not in doubt is the direction. Retailers earning citations from AI shopping in 2026 made their catalogues legible to software before the AI wave, and are plugging into open protocols on top of that foundation. If you are evaluating an agency or a platform right now, the question I would put in front of the technical team is narrower than the trade press suggests: not “what is your AI strategy”, but “show me one SKU end-to-end, from source of truth to Merchant Center feed to schema.org markup to variation attribute, and tell me where the divergences are”. If the answer is short and boring, you are looking at people who can do the actual work. If it is long and involves the word “content”, keep looking. That is the conversation we would want to be in on your behalf.