How to prepare your product catalogue (PIM) for AI agents in 2026
An AI agent doesn't look at your store, it reads your data. A three-level checklist covering the data model, identifiers, variants, relations and feeds, plus Akeneo, Pimcore and Ergonode and a 30/60/90-day plan.
In the era of Agentic Commerce, the winner is not the prettiest store but the one whose products an AI agent can understand, compare and recommend. The foundation is the catalogue in your PIM system. Below is a concrete checklist and what is worth changing right now.
This is the third post in our series. The introduction, Agentic Commerce and AI Shopping Agents, explains what is changing in e-commerce. The post on Agentic Commerce modules for Magento 2 covers the ACP, UCP and MCP protocols. Here we go down to the product data itself.
What does it mean for a catalogue to be ready for AI agents?
A catalogue is ready for AI agents when every product can be read as a complete record: a stable identifier, filled-in attributes, explicitly described variants and relations, plus availability and price taken from the system that actually knows them. An agent infers nothing from page layout. If it isn’t in the data, it doesn’t exist.
How concrete those requirements are shows in OpenAI’s product feed specification for shopping in ChatGPT: there are nine required fields (among them the identifier, title, description, product URL, brand, seller, image, availability and price), and a row without a valid availability value is rejected (specification read on 25 September 2026).
Why does PIM decide product visibility in 2026?
Because AI agents, meaning ChatGPT, Claude, Gemini and the shopping agents stores build themselves, don’t browse a website the way a person does. They read structured data. A product with weak or contradictory data is invisible to them, or recommended for the wrong use. The ability to be understood correctly is increasingly called AI Visibility.
A PIM system therefore stops being just a content management tool. It becomes the single source of truth about the product, used by people, sales channels and machines alike. If the same product has one name in the store, another on a marketplace and a third in the feed, the agent receives three contradictory signals and picks a competitor’s product whose data agrees with itself.
How does data for a person differ from data for an agent?
A person reads context, images and tone. An agent reads fields and their values. A description that convinces a buyer can be useless to an agent if the key fact sits in a sentence instead of an attribute.
| What | Enough for a person | What an AI agent needs |
|---|---|---|
| Dimensions | “Compact, fits any cupboard” | width: 42 cm, depth: 30 cm, height: 55 cm |
| Variant | A photo with a colour swatch | A separate record for each colour and size, with a shared group identifier |
| Compatibility | “Fits most models” | A list of models as an “accessory for” relation |
| Availability | A green dot next to the button | A current value: in stock / out of stock / pre-order |
| Returns | A “Terms” page in the footer | The return window in days and the policy URL on the product |
| Identification | Name and photo | GTIN (EAN), manufacturer part number, brand |
What does an AI agent need from the catalogue?
An agent needs eight groups of data, but not all of them belong in the PIM. That distinction matters: stock levels and current prices change too often for the PIM to be their source of truth. PIM-to-store synchronisation typically runs once or twice a day, which is fine for descriptions and far too slow for stock.
| Data | Example | Source of truth |
|---|---|---|
| Structured attributes | material, power, dimensions, standard | PIM |
| Product relations | bundle, replacement, accessory, spare part | PIM |
| Variants and options | colour, size, capacity | PIM |
| Identifiers | GTIN (EAN), MPN, brand | PIM |
| Multilingual content | titles and descriptions per market | PIM |
| Units and B2B sales rules | unit of measure, minimum order quantity, pack size | PIM (pricing rules in the B2B platform) |
| Stock and prices | availability, price, contract price | ERP, WMS or the e-commerce platform |
| Policies | shipping, returns, warranty | E-commerce platform, linked from the product |
Only the feed or the store’s API joins those two halves into the one record the agent receives. OpenAI’s specification also sets limits worth turning into PIM rules: a title of up to 150 characters and a description of up to 5,000 characters, in plain text, without HTML.
Checklist: how to prepare your PIM for AI agents, in three levels
Start with the foundation, because the next two levels don’t work without it. The order matters: a rich description without an unambiguous identifier is still a product the agent cannot compare with a competitor’s offer.
Level 1: foundation (must-have)
- A clean data model: attributes, attribute groups and product families.
- No duplicates and no contradictory values for the same attribute.
- Enforced completeness of key attributes, separately for each channel and language.
- Unambiguous, stable identifiers. A GTIN has exactly 8, 12, 13 or 14 digits with a valid check digit, and leading zeros must be preserved, so store it as text, not as a number.
Level 2: agent-ready
- Factual, unique descriptions, with the same facts also held in attributes.
- Product relations with a named type (“replacement”, “accessory for”, “part of a bundle”), not just “similar products”.
- Versioning and change history, so you can reconstruct what reached a channel and when.
- Exports for agents: a JSON or CSV feed that meets the channel’s requirements, plus schema.org structured data on product pages (
gtin,mpn,isAccessoryOrSparePartFor,hasMerchantReturnPolicy). - Attribute mapping to standards: GS1 for identifiers and, in B2B, ETIM or ECLASS classifications, which an agent recognises without having to decode your internal naming.
Level 3: competitive advantage
- B2B context: who can buy which range, in which unit and at which contract price (data from the B2B platform, linked to the product from the PIM).
- Quality signals where available: ratings, return reasons, the questions customers ask most.
- AI-assisted content enrichment inside the PIM, always with a human approving it before publication.
- Readiness for direct agent queries: an API with a narrow, read-only permission scope rather than full access to the system.
Which mistakes most often cost products their visibility in AI agents?
The most common mistake is filling missing data with stand-in values. OpenAI’s specification says it plainly: don’t use strings such as null, unknown or n/a, don’t treat an empty value as zero, and don’t invent a manufacturer part number to replace a missing GTIN. An empty field is honest. A made-up one misleads the agent.
The other mistakes we see most often:
- Key data held only in the store. Attributes typed by hand into Magento rather than the PIM drift away from marketplaces and feeds at the first change.
- Different values for the same attribute across channels. “Black”, “black”, “Blk” and “BLK” are four different colours to an agent.
- Descriptions written only for SEO. Keyword-heavy text in which the technical parameter appears once, in the third paragraph, reads worse to an agent than a short attribute table.
- No relations and no context of use. Without an “accessory for” relation, the agent cannot build a complete basket.
- Stale stock and prices. Usually because they were put in the PIM and are waiting for the nightly sync.
- Unstable identifiers. OpenAI’s specification forbids reusing an identifier for a different product and putting the price into an offer ID. An identifier has to survive changes in price, stock, title and images.
How do you prepare Akeneo, Pimcore and Ergonode for agents?
All three systems have what levels 1 and 2 need. They differ in how much you have to add yourself.
| System | What helps most | What is worth refining |
|---|---|---|
| Akeneo | Completeness per channel and locale, product families and variants, data quality scoring, product associations | Association types named the way the channel understands them (replacement, accessory) instead of one generic “related”; automation rules in the paid editions |
| Pimcore | A fully custom data model, field validation, approval workflows, a GraphQL API through Datahub | Flexibility cuts both ways: the model has to be designed deliberately, because the system won’t impose a structure for you |
| Ergonode | Completeness on product templates, workflow statuses, built-in AI for translations and descriptions | Approving AI-generated content before publication; consistent value lists for select attributes |
In all three, the thing that matters most is the same: data quality rules. An attribute required for a channel, a list of allowed values and format validation (GTIN, units) catch an error in the PIM before it reaches an agent.
How do you connect PIM with Magento or OroCommerce so the agent gets the full picture?
The PIM is the source of truth about the product, the e-commerce platform consumes that data, and the ERP supplies stock and price. The agent only sees the result of that assembly, so a weak link anywhere in the chain spoils the whole record.
Three integrations are critical:
- Product data from the PIM to the platform (Magento 2, OroCommerce): attributes, variants, relations and content. The store corrects nothing by hand here.
- Stock and prices from the ERP to the platform, as often as possible. In B2B that includes contract price lists and customer account rules.
- The agent layer: a product feed, structured data on product pages and an API for agentic protocols. Which ACP, UCP and MCP modules already exist for Magento is covered in our post on Magento 2 and Agentic Commerce.
How to design the data flow between those systems is described on our system integrations page.
What should you do in the next 30, 60 and 90 days?
Measure in the first month, clean up in the second, expose the data to agents and test in the third. The order matters more than the pace: a feed built on a messy data model has to be rebuilt later.
| When | What to do | How you know it’s done |
|---|---|---|
| Days 1–30 | Catalogue audit: completeness of key attributes, duplicates, contradictory values, GTIN coverage | You know the share of products with complete data for each channel |
| Days 31–60 | Clean up the data model and put quality rules in place in the PIM | A new product cannot reach a channel without its required attributes |
| Days 61–90 | Feed and structured data for agents, tests with real AI agents | The agent answers questions about your products correctly and builds a basket from them |
The agent test is simple and free: ask ChatGPT, Claude or Gemini a question your customer would ask (“which replacement fits model X?”) and check whether the answer rests on your data or on a competitor’s.
Summary
Preparing your catalogue for AI agents is not a future IT project. It is work that already decides whether the agent shows your product or a competitor’s. Companies with a well-ordered PIM will gain a clear advantage in the coming months, because the rest of the road, the feeds and the protocols, cannot be built on chaotic data.
At Macopedia we have delivered more than 30 PIM implementations (Akeneo, Pimcore and Ergonode). We help with the whole journey: from a catalogue readiness audit, through the data model and quality rules in the PIM, to integrations with Magento 2 and OroCommerce and the agent layer. If you want to check how ready your catalogue is, let’s talk.
Co-founder and CEO of Macopedia, Tomasz is an expert in e-commerce, PIM, WMS, and ERP solutions. For over 15 years, he has been shaping the e-commerce market, driving digital transformation for numerous companies and implementing innovative technological solutions. His articles provide strategic and practical insights for businesses aiming to optimize processes and achieve growth in the digital world.
More from this author →