One Inventory Truth for Agentic Commerce: Feed, API and Page Must Say the Same Thing

What is inventory truth in agentic commerce?

Inventory truth in agentic commerce means that a product's price and stock carry the same value on the page, in the schema, in the feed and in the API at the same time, and that the value is current at the moment the agent asks. Classic inventory management was back-office work: demand forecasting, reorder points, safety stock. In 2026 stock data moved to the shop window, because the agent recommending your product checks stock itself at the moment of purchase. The UCP Catalog capability Google announced on 19 March 2026 lets an agent pull live stock, price and variant data directly from your system right before it recommends; the OpenAI feed can refresh every 15 minutes. A brand that misstates stock no longer loses only one order; it slips in the agent's reliability score.

We first published this article in September 2024 under the heading of AI demand forecasting and inventory optimisation. Forecasting is still valuable; but in two years the question changed. "How much stock should I hold" now matters as much as "who reads my stock, when, through which channel, and is the number they see correct". Our thesis at Webtures: stale stock data is more dangerous than no stock. If the product is gone the agent skips you; if it looks available and is not, the agent flags you.

How does an agent read stock?

When an agent arrives with "blue, size M, delivered in two days" it can read stock from three places. First, the product feed: the availability field in Merchant Center and the OpenAI feed (in_stock, out_of_stock, preorder, backorder). Second, the schema on the page: availability and inventoryLevel inside Offer. Third, a live query: with UCP Catalog the agent asks you directly for variant-level stock and price at that moment; the cart capability adds several products in one flow. When the three sources disagree the agent does not pick the one it trusts most; it takes the product out of the comparison.

The second half of the mechanism is feedback. Most AI shopping surfaces track merchant data reliability: if a product you marked in stock turns out to be unavailable at checkout, the transaction fails, the failure is written to the merchant's reliability score, and the agent shows you less in the next round of recommendations. In our internal decks we summarise why in-chat checkout was pulled back in 2025 and 2026 in two words: stale data. In scraped static copies, stock, delivery and shipping information had gone out of date by the moment of purchase; at the moment trust matters most the agent was presenting a false reality.

Industry measurements show the size of the problem: average retail inventory accuracy sits around 83%, meaning one SKU in six carries wrong stock data at any given moment; roughly 60% of e-commerce catalogues contain missing GTINs, inconsistent attribute names or stale availability states. A person tolerates that; an agent does not.

Why must four layers say the same number?

The same product is represented on four layers, and each layer has its own update path. Inconsistency usually comes from layers being fed from different sources at different frequencies.

The goal is for all four layers to be generated from a single product record. The first question in our team's Agentic Commerce Assessment Set is exactly this: "Is product data held as a master record in one source, or scattered across channels?" The sixth question measures frequency: "How often and how accurately are stock and price synchronised across channels?" The answer is a synchronisation log; "it is usually up to date" cannot score above 2.

How do you keep stock data current for an agent?

1. Name one source. The stock table in the ERP, PIM or commerce platform is the master; page, schema, feed and API read from it. No hand-updated copy survives.
2. Build a two-layer update. A daily full snapshot plus API-based upserts for intraday changes. For fast-moving products the target lag is under 15 minutes; the OpenAI feed accepts that cadence.
3. Give stock at variant level. "Shirt in stock" means nothing to an agent; "blue, size 42: 3 units" does. Group with item_group_id in the feed and give every variant its own availability.
4. Be honest below the threshold. At the last one or two units, mark limited stock in availability and show the count on the page; the agent weighs the risk of non-delivery as much as price.
5. Code pre-order and lead time separately. Instead of showing an out-of-stock product as available, use preorder / backorder with availability_date; the agent can tell the user the wait.
6. Keep schema in sync with the page. An "InStock" value embedded in a static template is the most common source of inconsistency. Generate Offer.availability from the stock table; carry the count in inventoryLevel where available.
7. Open a live query. On platforms like Shopify, UCP Catalog support arrives with little effort; on custom stacks publish a product API (price, stock, variant) and make sure the ACP and UCP feeds are fed from it.
8. Measure inconsistency. Report the price and stock mismatch rate between page, schema, feed and API weekly. In our measurement framework this is the "data integrity" row, verified with consistency monitoring.

What is AI still good for in stock forecasting?

The core of the old article still holds: generating demand forecasts from sales history, season, campaigns and lead times, setting reorder points and safety stock dynamically, spotting dead stock early. What is new is the signal entering the forecasting engine. Agent-driven demand does not behave like human demand: an assistant can recommend a product in thousands of queries one evening and switch to a competitor the next day. Our assessment set has a question for it: "Has fulfilment capacity been tested against sudden agent-driven demand spikes?"

In practice three additions are enough: tag agent-driven orders separately (part of dark AI traffic cannot be measured, but the order source can be separated), feed AI visibility data (which product is recommended in which prompts) into the forecasting model as a signal, and hold higher stock thresholds for products agents recommend. AI keeps forecasting; the difference is that it also plans what the agent will see.

What are the most common stock mistakes?

• "InStock" embedded in the template. The schema always says available; when the product sells out the agent recommends it anyway.
• A daily file behind a live shop window. A product that sells out at 9am stays "available" in the feed until the next morning.
• Marketplace and site stock living apart. The same product is sold out on the marketplace and available on the site, or the reverse; the agent reads both channels.
• Total stock instead of variant stock. "12 units" is true, but zero in the size the user wants.
• A stock label loaded by JavaScript. No stock in the raw HTML; the agent treats it as unknown.
• Removing the out-of-stock page. A product that returns 404 becomes a broken link in the agent's memory; keep the page, mark out_of_stock and show a substitute.

How should cross-border brands set up stock?

On platforms such as Ticimax, ikas and İdeasoft the stock table is usually central; the break happens on the feed and schema layers. The first audit is three questions: does the feed travel by file or by API, is the schema generated from the stock table, is the marketplace integration two-way? In multi-market selling inventory truth is defined per market: a product in the German warehouse may be "backorder" for the United States; shipping and availability in the feeds must be given per country. In our own export matrix, B2B markets such as Germany run agent procurement through APIs; a supplier without a live stock API is ignored on the agent exchange.

Inventory truth is part of one data discipline together with product descriptions agents can read and category architecture for agent discovery. We placed UCP Catalog and the ACP feed on the protocol side in the protocol map. You can see whether your four layers say the same number with the Agentic Commerce Readiness scan, and set up consistency monitoring with our Agent Experience team.