• Shopware Gold Partner & Shopify Partner

Chapter 2

Your Shopware store through the eyes of an AI

Table of contents

Open one of your own product pages in the browser. You see images, price, description, reviews, the add-to-cart button. Everything is there. Now imagine the same page as plain text – without layout, without color, without the arrangement you chose so carefully. How much of it actually reaches a machine? For many AI systems the answer is: less than most store owners assume.

Why does a lot of content never reach AI crawlers?

Many modern store frontends – including Shopware storefronts with a custom or headless setup – load content via JavaScript instead of delivering it directly in the HTML. For a human visitor, that makes no difference. For most AI crawlers, that is exactly the wall.

Analyses by Vercel and MERJ show: the large independent AI crawlers, for example from OpenAI, Anthropic and Perplexity, do not execute JavaScript. Google is set up differently – Gemini uses the Googlebot infrastructure and can render. Where price, availability or description are only created by a script, an OpenAI or Perplexity crawler sees a largely empty page, no matter how cleanly the robots.txt is configured.

What does the customer see, what does the crawler see?

The same product page, two perceptions:

  • Product: The customer sees "running jacket" – so does the crawler.

  • Description: The customer sees all of it – the crawler sees that it exists.

  • Price: The customer sees 129 euros – for the crawler it is missing.

  • Availability: The customer sees "in stock" – for the crawler it is missing.

  • Variants: The customer sees four colors – for the crawler they are missing.

  • Rating: The customer sees 4.8 stars – for the crawler it is missing.

For the customer, the page is complete. For the machine, exactly the information it needs for a recommendation is missing.

Why is not every bot the same bot?

This is where most misunderstandings arise in practice. An AI provider usually runs several crawlers with different tasks. At OpenAI, GPTBot collects data for training the models. OAI-SearchBot, on the other hand, makes public web content accessible for ChatGPT search.

The practical difference: if you block GPTBot for data protection reasons, you can still specifically allow OAI-SearchBot. Protecting training data and appearing in search answers are two separate decisions. The same separation exists at Perplexity, where PerplexityBot serves search and Perplexity-User covers user-initiated requests.

A robots.txt that blocks all AI bots across the board or allows all of them makes the wrong choice for part of your own goals in both cases.

Where does the problem usually lie?

In our consulting practice, the most common cause is not the robots.txt at all, but the content delivery network or the firewall – through default settings that nobody in the company ever consciously chose.

Since July 2025, Cloudflare has blocked AI training crawlers by default on newly created domains. Since September 15, 2026, Cloudflare distinguishes its rules more strongly by the purpose of a crawler: search, agent and training are treated separately. For new domains, new websites of existing customers and all customers on the free plan who have not adjusted their settings themselves, training and agent crawlers are blocked by default on pages with ads, while search remains allowed.

The decisive point is in the detail: multi-purpose crawlers that combine several functions are treated according to the most restrictive rule enabled. Cloudflare explicitly names Googlebot, Applebot and BingBot as examples. So if you block training crawlers, under certain settings you may also block classic search engine crawlers – putting not only AI visibility but your Google ranking at risk.

This is not a theoretical scenario. Many store owners enabled the original "Block AI Bots" switch from 2025 with a single click, without anyone being able to foresee the consequences for 2026. If you have never checked it, you should.

What structured data does Shopware bring by itself?

Shopware has long delivered structured product data in the standard storefront. With version 6.7.9.0, the technical implementation was fundamentally modernized: instead of the previous microdata scattered across HTML attributes, the storefront now outputs structured data as JSON-LD in the page head.

A detail that is regularly overlooked in practice: the function is optional at first and is enabled via the feature flag JSON_LD_DATA. By default, the flag is switched off – and the old microdata output keeps running. If you have installed 6.7.9.0 and assume you are automatically delivering JSON-LD, you are mistaken. The old microdata is deprecated and will be removed with version 6.8.0.0.

In terms of content, the new product schema goes much further than the earlier microdata. It includes, among other things:

  • all product images instead of only the cover image

  • AggregateRating including the required review count, plus up to ten recent individual reviews

  • OfferShippingDetails with delivery time for products with a simple price

  • typed dimensions and weights as well as gtin13 and mpn, if maintained

  • new: WebSite schema with search function, Organization schema with store logo, ItemList schema on category and search result pages, and VideoObject entries for product videos

If you work on an older version or use your own storefront templates, do not assume the same information is automatically present.

Tech box for developers and agencies: Each JSON-LD schema lives in its own Twig template under storefront/layout/structured-data/. Each template provides two blocks: an outer one with the data logic and an inner _script block for the output. Extensions use the standard template mechanics – merge data into the respective data object, then call {{ parent() }}. If you added your own itemprop attributes in the past, you must migrate them before switching to 6.8.0.0, because the microdata will be removed then.

How does Shopware open a direct route to AI platforms?

With version 6.7.10.0, Shopware introduced its own "Agentic Commerce" sales channel type, initially as an experimental feature. It lets you export products as a JSONL feed, a preconfigured template covers the mandatory fields, and orders and traffic from AI platforms can be tracked through the existing affiliate infrastructure.

There is now also a dedicated extension by Shopware that goes considerably further: in addition to feeds for OpenAI and Google, it supports the Universal Commerce Protocol.

Two things that are often confused:

  • Product feed: makes sure your products reach the AI – as a structured data stream.

  • Universal Commerce Protocol: goes one step further. AI agents can interact with the store – query the catalog, fill the cart, trigger checkout, view orders.

What this means strategically is covered in Chapter 9.

How seriously should you take this today?

Beta really does mean beta here. Shopware explicitly labels the extension as a beta version; scope and behavior may still change. At the same time, adoption is still low.

That is not an argument against the direction – it is clearly the right one. It is an argument against uncritical use in production. If you try the extension today, do it on a staging environment first and test cart, checkout, orders and the functions released for AI agents thoroughly.

There is also a geographical restriction: registering as a merchant in the ChatGPT marketplace is according to the Shopware documentation currently limited to the USA and not yet available in Europe. For Shopware merchants with US business, the channel is therefore already relevant today. For the purely European market, the technical groundwork can be prepared while the rollout is still pending.

Why this asymmetry is interesting: if you are active in the US, you are gaining experience today that European competitors will only gain later. And if you only sell in Europe, you have lead time to set up product data and feed structure cleanly. Both are better than being caught off guard.

Two practical points during setup: attributes such as color or size are not system fields in Shopware but have to be created as properties and mapped manually in the feed. And a URL to the return policy is mandatory – Shopware does not provide a uniform default page for it, because the legal requirements differ by region.

What does a case from Shopware's own code teach us?

How quickly technical accessibility can tip over is shown by a case from Shopware's own development. The links in the Agentic Commerce feeds carry a tracking parameter to attribute the sales channel. However, the storefront's default robots.txt blocked all URLs with parameters with a blanket rule – so Googlebot could not crawl exactly these landing pages. The result: Google Merchant Center rejected the affected products. This was only fixed with version 6.7.13.0.

A single line in the robots.txt threw an entire product range out of a sales channel – while the store worked flawlessly for every human visitor.

The point is not to blame Shopware. Such interactions occur wherever new functions meet configurations that have grown over time. The point is: this check belongs on the checklist – and not just once, but after every major update.

What should you not expect from structured data?

Structured data is mandatory. It is not a magic bullet. An evaluation by Ahrefs of 1,885 pages that newly introduced JSON-LD schema between August 2025 and March 2026 found no meaningful citation advantage from schema alone compared with around 4,000 control pages – neither in Google AI Overviews and AI Mode nor in ChatGPT. However, the pages studied were already being cited regularly.

For context: schema makes content more machine-readable. It does not make bad content better. If you wrap weak product information neatly in JSON-LD, you end up with one thing above all: neatly structured weak product information. Technology creates the precondition – it does not replace the substantive answer to the question a customer actually has.

What you can check in ten minutes today

  1. Fetch your most important product page with the OAI-SearchBot user agent. Do you get a normal page with content back – or an error message, an empty page, a redirect?

  2. Check your robots.txt for explicit entries for OAI-SearchBot, PerplexityBot and comparable search agents – not just for the better-known training crawlers.

  3. Take a look at the source code: is the price directly in the HTML, or does it only appear after loading in the browser?

  4. If you use Cloudflare or a comparable CDN: check the bot management settings. What was enabled with a click a year ago may have a different effect today than intended back then.

  5. If you run Shopware 6.7.9.0 or newer: check whether the feature flag for JSON-LD is actually enabled.

Frequently asked questions

According to analyses by Vercel and MERJ, the independent crawlers from OpenAI, Anthropic and Perplexity do not. Google, by contrast, renders through the Googlebot infrastructure. Content that is only created by a script is therefore often missing for ChatGPT, Claude and Perplexity.

That is a separate decision from visibility in ChatGPT search. GPTBot collects training data, OAI-SearchBot makes content accessible for search. If you block GPTBot, you can still specifically allow OAI-SearchBot.

From version 6.7.9.0 yes, but only after enabling the feature flag JSON_LD_DATA. By default, the old microdata output keeps running; it will be removed with version 6.8.0.0.