Creating ETIM data with AI: from datasheet to BMEcat export

Your datasheet is finished, and yet your product is invisible in the wholesaler’s filters. Take an electrical planner who filters your wholesaler’s shop for “miniature circuit breaker, 16 A, B characteristic, 3-pole”. Your breaker meets all four criteria; it says so on page 1 of your datasheet. It is missing from the results list, while your competitor’s is there. The reason: in the ETIM record, the feature for rated current is empty.

The PDF has the values: rated current, number of poles, characteristic, dimensions. But they appear there as running text and a table, and the wholesaler needs them as class plus features plus values in the specified format. This typically happens when a manufacturer treats ETIM as an export format rather than as part of its product data. In between sits a product manager with Excel and a deadline. Today, AI can prepare most of the work of creating ETIM data, provided you stop a guessed value from slipping into your catalogue.

Why incomplete features cost you listings and shop filters

An ETIM class without filled-in features is like an address without a house number: the wholesaler may well add your item to its range. But its filters run on features, and without a value the item drops out of every search based on them.

Example scenario: a manufacturer supplies 4,000 items with the correct class, but only with the master data from the ERP: item number, description, packaging unit. The class features stay empty. The wholesaler lists the items, the filters don’t work, and the field sales team answers questions on the phone that a data record could answer.

You have the same gap in your own shop filter. And when a planner hands their research to ChatGPT or Perplexity, the value is missing there too, exactly where the assistant looks for a reliable figure.

ETIM, ECLASS and BMEcat: who governs what

In short: ETIM and ECLASS define what has to be said about a product. BMEcat and ETIM xChange define what the file you say it in looks like.

ETIM assigns every product to a class, and every class has defined features with a data type. As of October 2026, ETIM 10.0 is the current version; according to ETIM International, it was released in December 2024 and, according to figures from ETIM Deutschland, contains 5,640 classes and 17,377 features. ETIM 11.0 has been in beta since 15 October 2026, and the release is announced for 1 December 2026.

For your deliveries, the format question matters most. According to the announcement of Guideline 5.0 by ETIM International, the ETIM BMEcat Guideline 5.0 from June 2021 is the last version of this guideline; its successor is the JSON format ETIM xChange. Each national ETIM organisation decides for itself which format it prefers. In Germany, according to the BMEcat page of ETIM Deutschland, BMEcat remains the basis for communication. So if you supply German wholesalers today, you send BMEcat and keep an eye on xChange.

Creating ETIM data with AI, from datasheet to review

ECLASS itself shows that language models have arrived in classification. In Release 16.0, around 23,300 new class definitions were generated with AI and labelled in the attribute “source of definition” as “machine generated by GPT-4o mini”. For your feature data, this means the real work lies in reviewing the AI suggestions and in knowing where they came from.

The classic routes are manual work in the PIM or in Excel, a data service provider or a classification tool. For a few dozen items, the manual route is enough. A service provider delivers quickly, but the knowledge of your classes stays with them, and with the next release you commission the run all over again. From a few hundred items upwards, or when switching releases, I recommend a pipeline with four stages, of which only the last is human. As an example scenario, your three-pole miniature circuit breaker, 16 A, B characteristic, runs through all four.

Stage 1: extraction from the datasheet

It starts with extraction. A multimodal language model reads the datasheet as a PDF and returns a raw attribute set: “rated current 16 A”, “tripping characteristic B”, “number of poles 3”, each with a source reference such as “p. 1, technical data table”. You need the source reference later for the review. Values taken from dimension drawings are particularly error-prone, so the pipeline flags them separately for review.

Stage 2: assigning the ETIM class

The class is where the typical risk lies: a model formulates a plausible class that doesn’t exist in ETIM at all. That is why it searches the real ETIM class list, including synonyms, instead of formulating freely, and returns three candidates with reasons. For your breaker, two of them are close together: miniature circuit breaker or combined RCBO. The datasheet gives no rated residual current, so it is a miniature circuit breaker. If the margin is narrow and the datasheet offers no distinguishing feature, the case goes to manual review.

Stage 3: feature mapping by data type

Next comes feature mapping, and here the feature type decides what the model is allowed to do. For alphanumeric features such as the tripping characteristic, it only chooses from the ETIM value list. For numeric ones such as rated current, it converts to the unit that ETIM specifies for this feature.

Range features, such as a temperature range, get a lower and an upper limit; yes/no features get true or false. If the datasheet doesn’t provide a value, for example because it only appears in the test report, the field stays empty and is flagged as a gap.

A guessed value is worse than an empty field, because nobody looks for it any more. A gap ends up on the review list. A hallucinated rated current ends up in every retailer catalogue that draws on your data.

Stage 4: review by a person

The fourth stage is review by a person. A product manager sees the class, features, confidence and the source reference in the datasheet side by side, and confirms, corrects or rejects. The trap comes after the first clean runs: anyone who has seen 50 correct suggestions clicks Confirm on the 51st without looking.

To keep the review fast and still genuine, the interface guides the eye. Gaps come first: for the missing value from stage three, the product manager checks with engineering or deliberately leaves the field empty. Then come the uncertain suggestions, with confident ones only checked by sampling, and the source reference is always shown right next to them.

Every correction goes into the prompt for the next round as an example. Whether this raises the share of records that need no correction is something you measure in the pilot. In the end, the bottleneck of this pipeline is review time per item, not the model.

One data record for wholesale, your Shopware shop and AI search

A typical mistake after classification is to park the ETIM data in an export table that exists only for the wholesaler. Your shop then runs on different product data from your retailer, and at the next release change you maintain everything twice. Class and features belong on the product itself, in a PIM or at least in a leading product data source that all channels read from. Whether a dedicated system pays off depends on catalogue breadth, channels and languages; with 4,000 classified items, the answer is usually yes.

From this source, the wholesaler gets its BMEcat according to ETIM Guideline 5.0 or, where it already accepts it, ETIM xChange. Your Shopware shop gets the same features as properties. Once set up, an import mapping translates each ETIM feature into a property group, including filter visibility, and from then on the faceted filter fills itself without manual work per item.

In the example, the “miniature circuit breakers” category previously only has filters for manufacturer and price. Afterwards, the customer filters by rated current, tripping characteristic, number of poles and rated breaking capacity. At the wholesaler, the same happens on a large scale: in electrical wholesale shops with six-figure item counts, your complete features can end up in the filter, provided the retailer adopts them from your delivery.

According to Google, AI features don’t require any special markup, but Google does recommend that structured data matches the visible text. A visible feature table on the product page, fed from the same ETIM data as your product markup, meets that requirement. If the planner from the opening asks an assistant for “miniature circuit breaker 16 A, 3-pole, B characteristic”, the answer is on your page with nothing left open. Whether an assistant quotes it is outside your control. It certainly won’t quote an empty value.

Pilot with 50 to 100 items: how to get started

Don’t start with the whole catalogue. Take 50 to 100 items from two or three ETIM classes with good datasheets, and find a wholesaler that actually checks the data. That way you hit real edge cases, and a product manager can still look at every record. In the example, miniature circuit breakers and RCBOs would be good pilot classes, because class assignment wobbles precisely for this pair.

In the pilot, you measure:

• the hit rate of class assignment against the manual decision,
• the fill rate of the features before and after the run,
• the review time per item.

Before you start, measure how long manual maintenance of ten items takes, and use that to set the hit rate and review time at which you scale. If the review time in the pilot isn’t clearly below the manual figure, the pipeline won’t scale. Anyone who sets the threshold only afterwards will always find a reason to carry on.

Then have the files validated: ETIM Deutschland checks BMEcat XML for its members, and ETIM International offers a validation tool for xChange. Only once the pilot runs stably is it worth connecting the ERP and Shopware, so that prices, stock levels and ETIM features come from the same source. If you integrate earlier, you build the interface twice, because the data model is still changing during the pilot.

ETIM data is product data, not a mandatory deliverable

The pressure comes from the release cycle. Every new ETIM version brings changed classes and features that have to be mapped, with ETIM 11.0 just as with every release after it. If you maintain your data manually, you redo this work with every version. A pipeline repeats the run and only puts the changes in front of you for review.

The AI takes on the translation work; the decision on each record stays with your product management. Whether you build the pipeline yourself, use a tool or bring in a partner depends above all on one question: who in your company puts in the review time. The first step is the same in every case: this week, pick the two ETIM classes with the best datasheets. That is your pilot.