AI commerce: discovery outlives checkout

A product feed is a structured file of your catalog (identifiers, descriptions, price, availability, media) that AI shopping surfaces read directly instead of inferring from your pages. Over the last year it turned out to be the most durable piece of AI commerce infrastructure, while the checkout layer built on top of it turned out to be the least.

That gap is worth understanding even if you never sell a physical product, because it is a clean demonstration of which AI visibility investments survive vendor churn.

#The churn happened at checkout

OpenAI and Stripe published the Agentic Commerce Protocol in late 2025, and ChatGPT shipped Instant Checkout on top of it: buy without leaving the conversation. It did not last. OpenAI retired the feature in March 2026 and sent buyers back to merchants' own checkout flows, refocusing its own effort on product discovery (CNBC). Trade coverage at the time noted OpenAI never disclosed usage numbers for it (TechBuzz).

Google went the other way on a different clock. On January 11, 2026 it announced the Universal Commerce Protocol, co-developed with Shopify, Etsy, Wayfair, Target and Walmart, alongside a Business Agent that retailers can activate in Merchant Center and a Direct Offers pilot in AI Mode (Google). Google's developer documentation describes UCP as an open standard for direct, instant purchases across AI surfaces including AI Mode in Search and the Gemini app, with the retailer remaining merchant of record (Google for Developers).

Two vendors, two protocols, opposite trajectories inside twelve months. A team that spent a quarter building an Instant Checkout integration spent it on something that no longer exists. That is the normal risk profile of a transaction layer during a land grab, and it is not a reason to panic. It is a reason to be deliberate about which layer you fund first.

#The discovery layer did not churn

Underneath both protocols, the inputs have stayed the same the entire time: a structured feed, and a product page a crawler can read.

OpenAI's commerce documentation describes the merchant feed as a secure, regularly refreshed file carrying identifiers, descriptions, pricing, inventory, media and fulfillment options, and orders the work so that feed sharing comes first, ahead of checkout handling and payment (OpenAI Developers). The Product Feed Specification accepts TSV, CSV, XML or JSON, so most merchants can adapt an export they already run (Agentic Commerce Protocol).

Google's UCP documentation points merchants at an asset they have maintained for years, telling them to use their existing Merchant Center shopping feeds to capture high-intent customers during discovery (Google for Developers).

The feed outlived the checkout feature it shipped alongside. It will outlive the next one too, because every agent that wants to talk about your product needs to know what the product is, what it costs, and whether it is in stock. No protocol removes that requirement.

#Freshness is a contract, not a ranking tactic

The ACP feed spec says the system accepts updates every 15 minutes, and that frequent updates improve match quality and reduce out-of-stock or price-mismatch scenarios (Agentic Commerce Protocol).

Read that as a correctness requirement rather than an optimization. If an assistant quotes a price you no longer charge, the user finds out at your own checkout page. The damage is a broken promise made in your name by a system you do not control, and the assistant has every reason to stop trusting the source that caused it. Stale feed data is a different failure class from a thin product description. One makes you less visible; the other makes you wrong in public.

The same logic already applies to editorial content, which is why freshness signals like dateModified carry weight in AI answers. Commerce just makes the cost of being stale immediate and measurable.

#The page still does the persuading

Feeds settle the facts. They do not settle the choice.

OpenAI's help center states that product results in ChatGPT are organic and unsponsored, selected independently rather than influenced by partnerships, and that merchants are ranked on factors including availability, price, quality, and whether they are the maker or primary seller of the item (OpenAI Help Center).

Availability and price come straight from the feed. The other two do not. "Quality" and "maker or primary seller" are judgments assembled from the open web: your product page, your specification table, your documentation, third-party reviews, and whether the entity behind the listing is consistently identifiable as the manufacturer rather than one reseller among many.

That is ordinary generative engine optimization work applied to a commercial page. Answer the obvious question in the first paragraph instead of burying it under lifestyle copy, which is the answer-first structure that earns citations elsewhere. Put specifications in a real table, because AI engines prefer tables when the content is comparative. Name the entity the same way everywhere, so entity clarity does not force the model to guess who makes the thing.

#Where the feed and the page disagree, you lose twice

There is a quieter failure mode than a missing feed: a catalog whose feed and whose HTML tell different stories.

Feeds are usually owned by an ecommerce or data team and generated from the order system. Product pages are usually owned by marketing and edited in a CMS. Nothing forces them to agree on the model number, the material, the dimensions, the bundle contents, or which variant is the default. When they disagree, an assistant reconciling both sources has to pick one, and a source that contradicts itself is a weaker source than one that does not.

Treat it as a crawl-time check rather than a quarterly audit. Pull the feed, fetch the corresponding page, and compare the fields that a buyer would ask about. Any drift is a bug with an owner.

#What to fund, in order

1. Get the feed correct and current. Right identifiers, right price, right stock, updated on the cadence the surface accepts.
2. Make the product page fetchable and answer-shaped. If the specs only appear after client-side rendering, they are not available to most AI crawlers.
3. Reconcile feed against page automatically, and treat mismatches as defects.
4. Establish who the maker is, unambiguously and identically across your site, your feed and your third-party listings.
5. Only then evaluate a checkout protocol, on the assumption that the specific one you pick may be replaced.

The ordering matters more than the individual steps. Discovery compounds: work you do on feed accuracy and page structure keeps paying out through every protocol revision, every new assistant, and every reshuffle of who owns the transaction. Checkout integrations do not compound. They are bets on a particular vendor's particular quarter.

Instant Checkout ran from its launch in 2025 to its retirement in March 2026. Merchant Center feeds have been running for years and just got handed a new job. That is the whole lesson.