• Shopware Gold Partner & Shopify Partner

Chapter 5

Structure, Schema.org and machine-readable data

Table of contents

Chapter 2 showed what Shopware delivers automatically in the standard template. This chapter answers the follow-up question: what do you have to add yourself, and in which order? The answer up front: machine readability does not come from adding as many technical formats as possible. It comes from serving the same correct information consistently everywhere.

Why are unique product identifiers the foundation?

Before any schema type is of use, the product identity has to be right – and identity is more than a GTIN. For branded goods, the identifiers assigned by the manufacturer are the most valuable: GTIN and MPN, the manufacturer part number. Both need to be added, not newly generated. The internal article number, the SKU, is something different – it identifies the product in your system, not in the market.

  • GTIN: unique product identity in the market. Assigned by the manufacturer, for private labels via GS1.

  • MPN: the manufacturer's part number for the product. Assigned by the manufacturer.

  • SKU: internal article number from your own system – not a market identifier.

For genuine private labels, where the store itself acts as the manufacturer, different rules apply: correct GTIN assignment requires a membership with GS1 and an assigned company prefix. Inventing a GTIN or buying one from a trading platform can violate the policies of Google Merchant Center and comparable feeds and, in the worst case, lead to the feed being suspended. For products that have not been assigned such unique product identifiers – for example certain one-offs or custom-made items – Google Merchant Center provides `identifier_exists = false` instead of faking an identifier.

Where is schema actually still missing?

Shopware's automatic JSON-LD covers product, category and search pages. What it does not cover: buying guides, instructions, editorial content. You add these yourself – but according to a simple principle: schema is selected based on existing content; content is not produced just so a schema can be used.

For buying guides, Article or BlogPosting schema is a good fit, for step-by-step instructions HowTo schema. For question-and-answer sections, FAQPage schema can still make sense if the marked-up questions and answers are actually visible on the page. You should not expect an independent GEO effect from it, though. Google discontinued HowTo rich results back in 2023; FAQ rich results have also no longer appeared in Google Search since May 7, 2026. Whether and how individual AI systems use this markup for their source selection is not reliably documented.

How do you implement schema cleanly?

Two rules that are violated most often in practice. First: implement schema at template level, not page by page. A store does not have a thousand individual pages, but a handful of templates that generate a thousand pages.

Second, and this is the real core of this chapter: schema must correspond exactly to the visible content. If a product is sold out and the page shows that, but the schema still reports availability, a discrepancy arises – with errors and, in the worst case, the loss of rich result eligibility.

The safest safeguard: such discrepancies rarely arise from bad intent, but almost always because visible data is updated while the schema comes from a separate source. Feed the schema from the same data source that renders the visible page.

Why one data foundation for several output channels?

Product data should be maintained cleanly and completely once internally – several platform-specific outputs can then be generated from it. Google Merchant Center, Shopware's agentic product feed and other AI platforms have different formats and requirements. The wrong conclusion would be to maintain several independent product data sets for them. The right conclusion is a reliable data foundation in the store from which the respective feeds are generated automatically.

What distinguishes feeds from sitemaps?

A point that regularly gets mixed up in practice: a product feed transports structured commerce data to a platform that expects and processes this data. A sitemap helps a crawler systematically discover a store's relevant URLs and content. Both can make sense, but one does not replace the other.

Which conflict around variants is overlooked?

With variants, there is a conflict between two legitimate requirements. From an SEO perspective, variant URLs are consolidated onto the main product page via canonical so that Google does not index hundreds of nearly identical pages and spread ranking signals unnecessarily.

For AI crawlers, this creates a different problem: anyone following the canonical practically no longer sees the individual variant options. Whether the running jacket exists in red, size XXL and with exactly this product image is in the store – but it is no longer reliably available to the crawler as related information.

The solution is not to give up the canonical logic. It is to provide this detailed information additionally through a separate, machine-readable route.

That is exactly why we at arboro developed the AI Sitemaps for Shopware. The extension provides four separate sitemaps for products, images, FAQ content and category topics. Variant attributes such as color, size and other properties are output in a structured way directly with the product – regardless of how the variant URLs are handled via canonicals.

In addition, the extension normalizes product information that in many stores exists only as free text or inconsistently maintained fields, such as dimensions, material or volume. Instead of an AI crawler having to interpret tens of thousands of product and variant URLs one by one, it receives the relevant information in a compact, machine-readable form.

The AI Sitemaps deliberately run separately from the classic SEO path: they can be opened to AI crawlers and excluded from indexing by classic search engines. The arboro AI Sitemaps are available for free in the Shopware Store.

How much does an llms.txt achieve?

For about two years, the recommendation has been circulating to create an llms.txt file in addition to robots.txt – a table of contents in text form specifically for AI systems. The evidence so far does not justify high priority. In its official guidance on optimizing for generative AI search from May 2026, Google explicitly states that such additional AI text files are not necessary – although maintaining them does no harm either. An analysis by Limy of more than 500 million AI bot requests over 90 days found only 408 requests for the file. And Ahrefs found across around 137,000 websites that 97 percent of existing llms.txt files were never requested at all.

Our assessment is therefore simple: you can do it. You do not have to right now. If you have missing GTINs, deliver poor product data, block AI crawlers or generate important content only via JavaScript, you have more important construction sites. Prioritizing an llms.txt over these basics optimizes the wrong place.

For Shopware stores, the effort is low anyway these days: since version 6.7.12.0, llms.txt and agents.md can be enabled directly in the sales channel settings, without a plugin and without custom development. If you are up to date, you can therefore take the files along. You should not expect a measurable effect from them at the moment – and they do not replace any of the basics above.

Our assessment: the right order

For product data, this order applies:

  1. Clean product identity

  2. Central data foundation in the store

  3. Schema and feeds

  4. Discovery structures such as sitemaps

  5. Experimental formats

For product data, the order starts with clean product identity. Next comes a central data foundation in the store. From it, structured page data and the feeds for platforms that are supposed to process your products are generated. Discovery structures such as sitemaps then help with discoverability. Experimental formats come last.

If you reverse this order and start with the experimental format, you end up with a technically modern setup on an unclean data foundation. That is the most expensive way of still not being recommended in the end.

Frequently asked questions

Not as a priority right now. Google says such files are not necessary, and an analysis of more than 500 million AI bot requests found that the large AI search crawlers almost never request the file. Missing GTINs or blocked crawlers are more important construction sites.

No. Correct GTIN assignment requires a GS1 membership. Invented or purchased GTINs can violate the policies of Google Merchant Center. For products without an identifier, identifier_exists = false is provided.

It can make sense if the questions and answers are visible on the page. You should not expect an independent GEO effect: Google has not shown FAQ rich results since May 2026, and how AI systems use the markup is not reliably documented.