Ecommerce product data for AI agents
Product data quality has always mattered for feeds and merchandising. When the consumer is an AI agent the criteria change: completeness becomes a filter, freshness becomes a trust signal, and internal inconsistency becomes disqualifying rather than merely untidy.
Product data for AI agents is the machine-readable representation of a catalogue — identity, offer, attributes and policy — evaluated on completeness, freshness, internal consistency and resolvable identity rather than on presentation.
Four axes, four different failures
Each axis fails in its own way and each has its own fix. Treating "data quality" as one thing is why remediation stalls.
| Axis | Question | Failure | Fix |
|---|---|---|---|
| Completeness | Are the fields present? | Product excluded from constrained queries | Fill the product template, not individual pages |
| Freshness | Do they reflect current state? | Recommends sold-out or mispriced items; source gets discounted | Drive availability and price from live inventory |
| Consistency | Do markup and page agree? | Contradiction read as unreliability | One source of truth for page, markup and feed |
| Identity | Can this be reconciled elsewhere? | Cannot be matched to the same product on other sources | Stable sku plus gtin or mpn |
Freshness is a trust signal, not a detail
A human who lands on an out-of-stock product is mildly annoyed. An agent that recommends an out-of-stock product has produced a wrong answer, and the correction propagates: sources that produce wrong answers get weighted down.
This is the argument for wiring availability to real inventory rather than emitting a constant InStock in the template. A missing availability field costs you constrained queries; a wrong one costs you standing.
Variants are where catalogues quietly fail
Size, colour and material commonly change both price and availability, and a single parent-level record collapses all of that into one claim that is wrong for most variants.
For agents handling variant-constrained requests — a specific size, a specific finish — a parent-only record means you cannot satisfy the constraint even when you have the item in stock.
- Per-variant offers
- Emit an Offer per purchasable variant with its own price and availability, or use AggregateOffer with an accurate low and high price.
- Explicit attributes
- size, color, material as fields rather than as substrings of the product title.
- Variant-level identifiers
- A distinct sku per variant so the agent can reference exactly what it recommends.
This is an operations problem, not a project
Every replatform, theme update, app install and bulk import can silently drop fields. The stores that stay readable are the ones that check continuously rather than the ones that did a good migration once.
Practically: own it in the product template, generate feeds and on-page markup from the same source, and re-scan on a schedule rather than after a complaint.
Frequently asked questions
- What product data do AI agents need from an ecommerce store?
- A complete record per purchasable item: name, description, images, brand, a stable identifier, and an offer carrying price, currency, availability, shipping details and return terms — served as JSON-LD in the initial HTML.
- How often does product data need updating for AI agents?
- Price and availability should reflect live state, because a recommendation of an out-of-stock item is a wrong answer rather than a stale page. Descriptive fields change with the catalogue.
- Do product variants need their own structured data?
- Yes, where price or availability differ. A parent-only record cannot satisfy variant-constrained requests, so you lose queries for a specific size or finish even when the item is in stock.
- What happens if my markup disagrees with my page?
- The disagreement is read as unreliability rather than averaged. Generate the page, the structured data and any feed from one source of truth so they cannot drift apart.
Score your product data across all four axes
Completeness per field, consistency against the page, and what an agent would build from your catalogue.