Product information management for ecommerce used to have one main job: get correct product data onto the online shop and the marketplaces, quickly. That job still exists. But for a growing share of shoppers, the first reader of your product data is a language model that compares your product with a dozen others before a person sees your page. This article covers what changed for buyers and for the teams producing product data, which risks are live right now, and how PIM software handles both in practice.
Key Takeaways
- AI assistants now shortlist products before shoppers reach a product page. Shoppers arriving this way convert better, so the product page increasingly works as a confirmation step.
- Assistants and agents match products against structured attributes. A spec that exists only in a PDF datasheet or inside a paragraph of copy is effectively invisible to them.
- Generative AI made product copy cheap. Verified facts are now the bottleneck, and AI-written text needs rules that stop it from inventing them.
- EU rules banning generic green claims apply since 27 September 2026, including to product names and to listings already online. Copy tools that default to "eco-friendly" wording create direct legal exposure.
- PIM software handles this with typed data models, per-channel completeness rules, validation, approval workflows, and feed mapping. AI generation works best as one controlled step inside that pipeline.
What PIM For Ecommerce Covers Now
A PIM system collects product data from its sources, enriches and checks it, and sends it to every sales channel. The sources are usually the ERP for SKUs and logistics data, supplier files, engineering systems, and a DAM for images and documents. The channels are your own shop, marketplaces, distributor portals, Google Merchant Center and, more recently, feeds for AI shopping surfaces.
The data has distinct layers, and each one breaks in its own way:
- Identifiers: GTIN, MPN, internal SKU and brand. Wrong or duplicate GTINs break marketplace listings and product matching across retailers.
- Typed attributes with units: weight in kg, dimensions in mm, power in W, noise in dB. These drive shop filters, feed fields, and AI matching.
- Variant structure: which values the parent holds and which ones differ per child, such as size or color.
- Content: titles, descriptions, SEO fields and alt text, per language and per channel.
- Assets and documents: images, manuals, certificates and safety data sheets.
- Relations: accessories, spare parts, successor models and compatible products.
- Compliance data: manufacturer contact details, safety warnings, energy labels, material composition and, soon, product passport identifiers.
Manufacturers create most of this data. Retailers and marketplaces mostly consume it, add prices and merchandising, and send error reports back. So the quality of a listing on a retailer's site usually traces back to the manufacturer's records. AI assistants that pull from many retailers at once make every inconsistency between those copies visible.
How AI Changed Buyer Behaviour And Expectations
Adobe bases its retail data on over a trillion visits to US retail sites, and its first-quarter AI traffic report shows how fast this moved. Traffic from AI sources grew 393% year over year in the first quarter of 2026. In March 2026, it converted 42% better than non-AI traffic. A year earlier, it converted 38% worse. Visitors arriving from AI sources stayed 48% longer and viewed 13% more pages. In the companion survey, 39% of US consumers said they had used AI for online shopping, and 66% said they believe AI tools give accurate results.
A likely explanation for the conversion jump is that the comparison step moved. A shopper asks an assistant for a cordless drill under 1.5 kg that comes with two batteries and a case. The assistant returns a short list, and the shopper clicks one. The product page then has one job: confirm. If it says 1.8 kg, or the case turns out to be an extra, the sale is gone.
Queries also got longer and more specific. Keyword search matched words. Assistants resolve constraints: a weight limit, a niche width, a noise ceiling, a certification. Shoppers ask about limits too, like "will this fit a 60 cm cabinet" or "does it work with my old charger". Each of these needs a typed attribute or an explicit compatibility record to match against. A dishwasher described as "whisper quiet" with no numeric dB value has nothing to offer when the question is "under 44 dB". Products described only with positive adjectives give the assistant nothing to answer with.
Consistency across surfaces matters more than it did. The assistant's answer, the marketplace listing, and your own product page now sit one click apart. When the price or a spec differs between them, shoppers read it as unreliable data. The assistant may simply pick a seller whose data contradicts itself less.
The second audience is the model itself. Adobe also benchmarked how much page content LLMs can actually read, on a score from 0 to 100%.
US retail product pages scored 66% on Adobe's AI visibility benchmark, the lowest of all page types in the report. Homepages scored 75%. Roughly a third of product page content is not readable by the models that send this traffic.
In the catalogs we work with, the usual causes are specs loaded through JavaScript tabs, tables embedded as images, technical data held only in PDF datasheets, and variant details that appear only after a click. Rendering is a shop problem. The PIM part of the fix sits upstream: when every spec exists as a typed attribute, the shop can output it as plain HTML and structured data, and feeds can send it directly without anyone scraping a page.
How AI Changed Product Data Work
Writing and translating product copy used to be the slow, manual part of a launch. Models now produce a first draft in seconds, in any language, for any channel. So the scarce resource moved. Text is cheap. Verified facts are the bottleneck.
A model asked to describe a product with half its attributes empty fills the gaps with plausible values. That's how a drill ends up with a torque figure nobody measured, translated into every language you sell in and sent to every channel you serve.
Let AI write the sentences. Never let it supply the facts.
AI tasks that work well inside a PIM share one property: the input is data the system already holds, and someone has verified. Generating titles and descriptions from attributes fits. So does translation against a glossary of approved terms, SEO fields per category, and alt text drafted from image content and product data. Classification suggestions for imported supplier products work too, as do attribute values extracted from supplier PDFs, as long as they land as suggestions with a source reference until a person confirms them. Fields with legal weight need a tighter path: safety warnings, environmental and durability claims, compatibility statements, and anything that ends up in a product passport.
Distribution changed as well. AI platforms now take product data directly. Google launched the Universal Commerce Protocol in January as an open standard for agent-driven commerce. According to Google's Merchant Center documentation on UCP, it lets participating merchants in the US, Canada, and Australia offer checkout on eligible listings in AI Mode in Search and in Gemini. Only listings with the native_commerce attribute show that Buy button. Merchant Center stays the hub for preparing product data, and Google's advice for getting ready is to keep everything from product feeds to brand assets complete and current.
OpenAI moved the other way on checkout. In March, it scaled back Instant Checkout in ChatGPT, let merchants keep their own checkout, and kept merchant product feeds as the input for product discovery.
So where the purchase happens is still open. The need for a complete and current structured feed holds on every platform. Each surface has its own feed specification and refresh rules, and the number of surfaces keeps growing. Teams that maintain these feeds by hand from spreadsheets end up spending their time on format conversion.
Risks That Matter Right Now
Generic Green Claims In Generated Copy
The EU Directive on Empowering Consumers for the Green Transition applies from 27 September 2026. It bans generic environmental claims such as "eco-friendly" or "green" unless the trader can demonstrate recognised excellent environmental performance, for example through the EU Ecolabel. It also bans sustainability labels that aren't based on a certification scheme or set by public authorities, and climate claims based on offsetting.
A Latham & Watkins analysis of the new rules points out two things that matter for product data teams. The rules apply to claims on products already on the market, with no transition period. And Commission guidance treats product names, brand names, imagery and colors as potential claims. Member States must allow maximum fines of at least 4% of annual turnover in the countries concerned. Transposition is uneven: in May the Commission opened infringement procedures against 20 Member States, so the details vary by market.
Under the new EU rules, a green leaf icon next to a product or the word "eco" in a product name can count as an environmental claim. This applies to listings already online, with no grace period for old content.
In our experience, language models asked for appealing copy for household or outdoor products reach for this vocabulary unless a rule stops them. The exposure also sits in old content. A catalog built over years contains descriptions written by many people and agencies. A PIM can search every text field in every locale for a list of blocked terms and route each hit to review. Doing the same across a shop, several marketplaces, and a stack of distributor spreadsheets is slow and error-prone.
Errors That Spread To Every Channel At Once
A wrong attribute value used to mean one wrong page. Now it goes to the shop, every marketplace, and every AI surface that ingests the feed. Some of those surfaces refresh on their own schedule, so the wrong value can stay visible after you fix it at the source. Generated text adds a multiplier: a hallucinated value in one description gets translated into every language.
Stale Price And Availability
Agents compare price and stock across sellers in one view. A feed that refreshes once a day while prices change several times a day produces mismatches the shopper notices at checkout. Google Merchant Center already disapproves items when the feed price doesn't match the landing page. Price and stock need a faster path than descriptions, usually delta updates through an API.
Product Passports And Listing Duties
Compliance data is turning into product data. The EU battery passport becomes mandatory in February 2027. Under the ESPR, the Commission's working plan puts iron and steel first for digital product passports, followed by textiles. The General Product Safety Regulation already requires online listings to show manufacturer contact details, a product identifier, an image or type designation, and safety warnings. The European Accessibility Act, in force for many online shops since June 2025, adds requirements such as text alternatives for images. Each of these is a set of attributes that needs a clear owner and a change history.
Product Data Sent To External AI Providers
Unreleased products, purchase prices, and supplier terms often sit in the same system as the descriptions you want generated. Before connecting an AI provider, check where it processes data and whether it uses inputs for training. Also check whether you can switch providers without rebuilding prompts and workflows. Self-hosted PIM installations and provider-agnostic AI connectors reduce that lock-in.
How PIM Software Handles Product Data In Practice
A Typed Data Model For Each Product Type
The PIM defines an attribute set per product type through classifications or product families. Drills get torque, voltage, chuck size, and battery capacity. Dishwashers get noise level, place settings, niche dimensions, and energy class. Every attribute has a data type and a unit. The text "approx. 1,5kg" is useless for a filter. The value 1.5 with the unit kg works in a shop filter and in an agent's constraint check alike, and the PIM converts it on export when a channel wants grams.
Standard classifications such as ETIM, ECLASS, GS1 GPC or UNSPSC give manufacturers attribute sets their distributors already expect. Mapping a catalog to the relevant standard once replaces repeated work on every distributor template.
Inheritance And Variants
Shared values live on the parent product and flow down to the variants. Values that differ, such as size or color, live on the child. Change the material on the parent and every variant updates. Assistants depend on this structure for questions like "available in size 44 in black", which only work when size and color are separate attributes on each variant.
Completeness Rules Per Channel And Locale
Each channel has its own required fields. Google expects identifiers and a category-specific set of attributes. A distributor portal might require ETIM features and a CAD file. An AI shopping feed needs fulfillment options alongside price and inventory. The PIM computes completeness per channel and per language and uses it as a publishing gate: a product goes to a channel only once the fields that channel requires are filled. Dashboards show which product families lag, so teams can close gaps in order of revenue impact.
Validation Rules That Catch Errors Before Publication
Completeness checks whether a field is filled. Validation checks whether the value makes sense. Typical rules:
- GTIN length and check digit, plus duplicate detection across the whole catalog
- plausible ranges per attribute, so a handheld vacuum entered at 20 kg gets flagged
- conditional requirements: if "battery included" is yes, battery type and capacity become mandatory
- consistent units, with dimensions stored in one unit and converted per channel on export
- blocked terms across all text fields and locales, such as generic environmental claims, with hits routed to compliance review
- numbers in generated descriptions checked against the stored attribute values, where the PIM supports scripted rules
Workflows, Roles And Approval
A status model turns data work into a controlled sequence. A product moves from draft to technically complete, then to enriched, approved, and published. Field-level permissions decide who changes what. Product management owns technical values, marketing owns copy, compliance owns claims and safety fields, and the ecommerce team owns channel mappings. AI-generated content carries a flag until a person approves it.
Workflows connect the steps. When a product reaches "technically complete", the workflow generates descriptions in all active languages and assigns review tasks to the right editors. In AtroPIM, for example, the AI integration module builds prompts from attributes already stored in the system and lets teams switch AI engines (ChatGPT, Gemini, or Jasper). Combined with the workflow module, it triggers generation on status changes. Editors see the output and approve or edit it before it is saved and published.
Channel Mapping And Feed Delivery
One master value feeds many outputs. Title templates differ per channel; for example, brand, product type, key attribute, and variant for a marketplace, and a shorter form for the shop. Attribute mapping translates internal fields into each feed specification. Unit conversion, language fallbacks, value lists, and character limits run on export. Full exports go out on a schedule. Price and stock go out as delta updates through the API. With this setup, a new AI surface is a new mapping on top of data that already exists.
Supplier Data Intake
Retailers and distributors get most of their data from suppliers, and manufacturers do the same for bought-in components. The PIM imports supplier files in BMEcat, Excel, CSV, or via API, and a column mapping set up once gets reused on every update. AI extraction can fill gaps from supplier PDFs. Each extracted value lands as a suggestion that names its source document until someone confirms it.
Change History As Evidence
Every change gets logged with user, time, old value, and new value. When a regulator or a competitor questions a claim, the history shows when the claim entered the catalog, which evidence it relied on, and who approved it. Product passports need the same traceability.
In projects we implemented for manufacturers selling consumer goods in the EU, the September deadline for green claims exposed the same pattern. Descriptions written over years by different agencies, plus newer AI-generated text, contained generic environmental phrases that nobody had checked against evidence. The fix combined a blocked-terms rule across all locales with a dedicated claim-evidence attribute linked to the certificate in the asset library. A description may now mention a certification only when that attribute is filled, and the workflow routes every flagged text to compliance before it can go live again.
Choosing PIM Software For AI-Driven Ecommerce
A dedicated PIM starts to pay off when you sell through more than one channel, in more than one language, with product data from more than one source system, or with regulated attributes. The shop backend stores product data for one channel. A PIM manages it for all of them.
Feed specifications keep changing, so the data model should be configurable without development work. API coverage needs to include import and export, with delta updates for price and stock. Per-channel and per-locale values with completeness scoring are the core of channel readiness. Workflows and field-level permissions make AI generation safe to run at catalog scale.
On the AI side, a choice of providers matters more than one built-in model, since models and their pricing change quickly. Hosting matters when product data and prompts must stay inside your own infrastructure. Open-source systems such as AtroPIM, published under GPLv3 and built on the AtroCore data platform, can run on-premises and let your team inspect the code that sends data to an AI provider.