AI text generation · ChatGPT, Claude, Mistral & Gemini

Descriptions, meta titles, meta descriptions and keywords at the push of a button - in every language. With ChatGPT, Claude, Mistral or Gemini, all usable in parallel with provider selection per generation. For many products at once, ChatGPT, Claude and DeepL are available in bulk edit.

Four providers, all directly integrated

Each provider has its own section in the config matrix in the Texts & translations tab - with enable switch, API key, model selection, connection test and default tone. All providers can be active in parallel. DeepL is added for translations.

You set the model per provider in the config matrix. When several providers are active, you pick the provider per call in the AI window of the detail view - also for variants - and in category management, for example Claude for the demanding main description and another model for the meta data.

Set up API key

1. Create an API key with the respective provider: console.anthropic.com (Claude), platform.openai.com (ChatGPT), console.mistral.ai (Mistral), Google AI Studio (Gemini) or deepl.com/pro-api (DeepL)
2. In the config matrix, Texts & translations tab, enable the provider and enter the key
3. Use "Test connection" to check that the key is valid (separately per provider; for DeepL the character usage is shown as well)
4. Pick the model and the default tone
5. For ChatGPT, Claude and DeepL optionally enable the switch for bulk edit

What can be generated

The generator creates these fields, individually or together:

• Description - a written product description
• Meta title and meta description
• Keywords
• Product name - not preselected for safety so that existing names are not overwritten by accident; overlong names are cut to 255 characters

Lengths are controlled by the length targets in the config matrix (product name, SEO title, meta description, description in words or characters; 0 = no target). Generation works for a single product in the detail view or for many products via bulk edit.

Context chips · You control what the AI sees

Per generation you decide via chips which product data is sent to the AI provider:

• Product name and product number
• Existing description as a starting point
• Manufacturer, price, EAN
• Categories and properties
• Weight

In addition there is a field for your own keywords and notes, plus the master prompt from the config matrix. This way you control how much context is sent per call.

Five style templates

• Professional - factual, fact-oriented
• Casual & modern - more direct, younger tone
• Premium / luxury - refined expression
• Technical / factual - precise, with technical terms
• Emotional / promotional - storytelling approach

You set the default tone per provider in the config matrix; it can be adjusted when generating.

AI in bulk-edit

Only ChatGPT, Claude and DeepL: these three providers are available in bulk edit, each with its own switch in the config matrix (for example "AI text generation in bulk edit", "Claude in bulk edit", "DeepL translation in bulk edit"). Mistral and Gemini are not available there.

Flow: select products → bulk edit → AI & texts tab → enable a provider → choose context, fields, tone and target languages → start. Optionally every channel/language combination gets its own texts. The number of parallel requests for ChatGPT and Claude is set in the config matrix.

Preview before applying

In the detail view generated texts first appear in a preview. You check them and apply them with "Apply texts" or discard them. In bulk edit the AI writes directly into the selected fields - so only select the fields that should be newly created. A snapshot is created before the run and restores the previous state. For DeepL, "Overwrite existing texts" decides whether maintained translations are replaced or only empty fields are translated.

Costs and limits

Billing is directly with the respective provider (OpenAI, Anthropic, Mistral, Google or DeepL). Stone & Water charges no additional costs or margins.

Troubleshooting

• Check the API key under config matrix → Texts & translations → Test connection
• Check credit or billing status with the respective provider (Anthropic, OpenAI, Mistral, Google)
• For errors like 429 or timeouts, reduce the number of parallel AI requests in the config matrix
• Check server logs: var/log/prod-*.log

Copy from default language - a starting text for the AI

When a sales channel uses its own language, no own texts exist there at first. The storefront shows the inherited content, but in that language the AI has nothing to build on. The function "Copy from default language" in bulk edit (section AI & texts) closes that gap.

• Name, description, meta title, meta description and keywords are copied; fields and target languages are freely selectable.
• By default only empty fields are filled - existing texts stay unchanged. Overwriting can be enabled when needed.
• Whether a target language already has own texts is checked against the translation records, not against the inherited display values.
• Saving is bundled through the same path as the AI texts, including automatic retries on request limits.

An introduction to the topic is on AI product descriptions for Shopware.

What's next

That completes the tour through the docs. The manual has every feature in one compact reference page.

Frequently asked questions