The generation modules

Nine separate tasks, each with its own place in the sidebar, its own settings and its own kind of result.

How a job is created is the same for all of them and is explained on Generating content. This page is about what each task does and which settings belong to it. Those settings live in three places, and they are the same settings everywhere:

• Settings › Configuration — your shop-wide default (Owners and Managers). See Configuration.
• A template — a named preset you can pick per job.
• The job wizard — for this one run only.

Product description

Sidebar: Products › Descriptions. Produces: a long product description, as formatted text, one per product and language.

This is the workhorse. It reads what your shop knows about a product — name, manufacturer, number, categories, properties and any custom fields you selected as context — and writes a description in the language you asked for, ending with bullet points of the key facts. The AI is explicitly instructed not to invent anything that was not in the input, so the quality of the result depends heavily on how much real information the product carries. If descriptions come out thin, add more property groups and custom fields to the Product context of a template.

Its options

• Also generate Meta content — a checkbox in the wizard, on this task only: “Runs in the same job and uses the freshly generated content as input.” The meta texts are then written from the brand-new description instead of the old one, which is exactly what you want. It gets its own Template dropdown next to it. Greyed out if no model available to your shop can do meta content.
• Web research — “Search the web for additional product information before generating.” See Research options below.
• Keyword research — only if the keyword add-on is enabled for your shop. See below.
• Write result to — Product description (the field customers read) or a Custom field of yours. Pick the specific field in a template; it is the only place with the live list from your shop.
• With expert mode on: the raw System prompt and Generation prompt, Temperature and Max tokens.

Meta content

Sidebar: Products › Meta texts. Produces: three short fields per product and language.

• Meta title — aimed at roughly 50–60 characters.
• Meta description — aimed at roughly 150–160 characters.
• Meta keywords — five to eight, comma-separated.

These are the texts search engines show in their results, so they are the shortest texts you will generate and the ones most worth reviewing individually.

All three appear as separate blocks in the Review dialog, each next to what is in your shop today.

Its options

• Keyword research — if enabled for your shop. See below.
• With expert mode on: one prompt per output field (Meta title system prompt, Meta description system prompt, Keywords system prompt), plus Temperature and Max tokens.

Multi-field text

Sidebar: Products › Multi-field texts. Produces: up to eight product texts at once, per product and language — whichever fields you pick.

The two tasks above write fixed fields: a description, or the three meta texts. This one lets you decide what is written. You list the fields you want — the product name, the description, the meta texts, and any text custom field your shop has — give each one its own instruction, and the AI writes them all in a single pass. Because they are written together, they read as one voice: the name, the description and the meta title do not repeat each other word for word, and they do not contradict each other.

Use it when your shop renders bespoke fields (a USP box, a spec teaser, a layout text) that the fixed tasks cannot reach, or when you simply want everything about a product rewritten in one go.

Choosing the fields

The wizard’s What gets generated card holds the whole configuration, under Fields to write: “Each row is one product field. All of them are written in a single request, so the texts of one product fit together instead of repeating each other.”

Every row has two parts:

• Field — a dropdown in two groups. Product fields holds Name, Description, Meta title, Meta description and Meta keywords. Custom fields holds your shop’s own text fields, by their real names, read live from the shop when the page opens.
• Instruction for this field — what this one field should say, and how long. Picking a field fills an empty box with a sensible default instruction you can then rewrite; text you have already written is never overwritten.

The arrows move a row up or down, Remove deletes one and + Add field adds one. A job can write up to eight fields; the add button disappears once you reach that.

If we cannot reach your shop while the page loads, the custom fields are missing from the dropdown and the panel says why. The five standard product fields still work, so a job is still possible.

What happens afterwards

• Review shows one block per field, titled with the name you picked the field by, each next to what is in your shop today.
• The job page states every destination on the row: Goes into: Name, Description, Custom field “…” — and the confirmation page before sending repeats the list.
• A field the AI left empty is stored and shown, but never written into your shop.
• One item is charged once, whether it writes one field or eight: it is one request either way. Sending is charged per item as usual.

Its options

• Keyword research — if enabled for your shop. See below. The keywords are offered to every field, not just one.
• With expert mode on, in Settings › Configuration or a template: System prompt (the shared preamble in front of all the per-field instructions), Temperature and Max tokens per field — the budget of one request is that number multiplied by how many fields you picked.

Product images

Sidebar: Products › Product images. Produces: one or more new pictures per product, uploaded into your shop’s media library when you send them over.

Instead of a single prompt, this task uses an Image chain: “Each row produces one image for the product. A row can build on earlier rows, the product’s existing photos, or the reference images below.” Every row is run for every selected product, so a three-row chain over twenty products makes sixty pictures.

Building the chain

Each row has:

• A prompt — what this picture should show. Below it sit clickable chips labelled Insert variable: that drop a placeholder into the text so the same prompt adapts to every product: {{ product_name }}, {{ manufacturer }}, {{ product_number }}, {{ category }}, {{ properties }}. So “A studio photo of the {{ product_name }} on a white background” becomes a different instruction for every product in the job.
• Pictures this step may build on (optional) — tick boxes naming what the AI may look at while making this image: Current cover photo, All existing photos, Image from step 1 (and 2, 3… for earlier rows), Uploaded picture 1 (and so on, for your reference images). This is what makes it a chain: row 2 can take the picture row 1 just produced and change one thing about it.
• Image shape — Square (1:1), Landscape (4:3), Wide (16:9) or Portrait (3:4), per row.

The arrows at the top of a row move it up or down — order matters, because a row can only build on rows above it. Remove deletes a row, + Add image adds one.

The rest of the panel

• Reference images — “Optional. Upload images you can then tick as an input on any row above.” Use Choose files, or drag and drop, or paste (click the upload area first, then paste). JPEG, PNG and WebP up to 12 MB each. These are shared by the whole job, which is why they are numbered.
• Target media folder — which folder in your shop’s media library the finished pictures are filed into. Default (no folder) lets your shop decide.

There is no test run for picture tasks. Start with one or two products, look at the result, then create the real job.

Category description

Sidebar: Categories › Descriptions. Produces: a description text for a category landing page, one per category and language.

It writes from what the category is, where it sits in your tree and a sample of the products inside it — including products in subcategories, which matters because a top-level category has almost no products of its own.

Its options

• The category picker in the wizard: Include subcategories, Only categories that contain products, Skip categories that already have a description. Folders, links and inactive categories are never listed.
• Write result to — Category description or a Custom field.
• Keyword research — if enabled for your shop.
• With expert mode on: System prompt, Generation prompt, Temperature, Max tokens.

Category images

Sidebar: Categories › Category images. Produces: a new cover picture for a category.

This is the one image task whose settings are ordinary configuration fields rather than a bespoke panel — which means you can edit all of them, watermark included, right inside the job wizard under How the AI should write (optional).

Its options

• Prompt mode — how the instruction for the picture is put together:
  • Manual prompt — you write it, once, for all categories.
  • AI-authored prompt — a text model writes the image instruction first, per category. Better when your categories are very different from one another.
  • Recreate existing cover — your text becomes an edit instruction applied to the category’s current cover picture. Good for “same picture, but on a light background”.
• Image prompt — the text itself. Placeholders available: {{ category_name }}, {{ category_path }}, {{ category_description }}.
• Image size — 1:1 (square), 4:3 (landscape), 16:9 (wide) or Portrait.
• Look the category up on the web first — only used in AI-authored prompt mode. It searches the web for the category and lets the model use what it finds. “This costs one extra research credit per category the first time it is looked up.” Only search this website narrows the search to one address, for example a manufacturer’s site.
• Mark images as AI-generated and its four settings — editable here, in a template, or in the wizard.
• With expert mode on: Author system prompt, Author temperature, Author max tokens and the Target media folder id.

FAQ suggestions

Sidebar: Categories › FAQ suggestions. Produces: question-and-answer suggestions for a category — and, if you want them, for single products inside it.

This task produces suggestions, not entries. None of it is in your shop until you have accepted it and sent it. A finished run leaves one FAQ suggestions panel per category on the job page, where you accept, reject or edit each suggestion on its own — and only the push writes the accepted ones into your shop as FAQ entries.

For the category, or for one product

Every suggestion belongs either to the whole category or to one single product. That is the distinction this task turns on, because an entry on the category is inherited by everything below it: its products, its subcategories and the products in those.

This is why a second pass checks every category-level suggestion against every product beneath the category before it is offered to you. “Machine washable at 30 degrees” can be true for 40 of 42 textiles — and on the two leather items it would be a promise to your customers that is simply wrong. That check has three outcomes:

• It holds everywhere — it stays a category entry.
• It holds except on a few products — it stays a category entry and those products are excluded from it (see below).
• It holds only for a handful of products — it becomes one product entry per product, marked in the review panel as Written for the category, kept for this product only. If it holds almost nowhere it is dropped, and only the item’s summary still mentions it.

Exclusions for single products

An exclusion is the answer to “true almost everywhere here”: the entry hangs on the category but does not appear on the excluded products. The review panel lists them under Products this entry is hidden on; unticking one shows the entry on it after all. How many exclusions a single entry may carry is capped, with expert mode on, by Max share of excluded products and Max excluded products.

In your shop the same thing appears in the plugin’s FAQ bulk editing screen: pick the category on the left and the Products in this category card lists each product’s entries — its own, the inherited ones marked Inherited, and the excluded ones as inherited, hidden here. You can also set an exclusion there yourself (Hide here) or lift it again.

Improvements of entries you already have

Before writing, we read the FAQ entries your shop already has for this category and its products. Anything already answered is not suggested a second time; a thinly answered entry can come back as an improvement instead. An improvement creates nothing: it overwrites the question and the answer of the entry it names. The review panel shows both versions side by side, as Current in shop and Suggested.

If you deleted that entry in your shop in the meantime, pushing the item fails with a sentence that says so. The card then offers exactly one button: The old entry is gone — send as a new entry.

An improvement stays exactly where the entry it rewrites already is. Its card names every place your shop shows that entry — which can be several categories, including ones outside this run — and offers no Shown for switch: the push replaces the text and touches none of the assignments. If you do want to move it, turn it into a new entry first with the button above. Because an improvement is checked against exactly the products your shop shows it on, one that reaches a category this run did not read is dropped instead of guessed at; the prompt log says which.

Reviewing the suggestions

On the job page, each item has an FAQ suggestions panel underneath the item table (the table links to it as FAQ candidates). Per suggestion it offers:

• Accept, Reject and Reset — plus Accept all / Reject all for the whole item. “Only accepted suggestions are sent to your shop.”
• Question and Answer, editable. Preview shows the last saved version, formatted.
• Shown for — The whole category or One single product; for a single product you then pick which one. Improvements do not have it (see above).
• The entry’s exclusions, its keywords, and under Why the answer says this the product data the answer came from.
• A Possible duplicate of “…” in … note when the same question was suggested for a second category of the same run. Nothing is dropped for you — which of the two categories should keep the question is a call only you can make.

Anyone who may see a job but not change it sees the same cards without the forms; that includes an approval request in the approvals inbox.

One language per job

This task writes one language per job. Its language picker is therefore radio buttons rather than checkboxes, with the line “This task writes one language per job.”, and the draft page names the language once more: “FAQ suggestions are generated in <language>; other languages can be added later.” So you get one item per category, not one per category and language.

So that a new entry is never blank in the storefront, we write the text into your shop’s system language as well: the plugin reads the system language as its base, and an entry without that version would render as an empty line. To add another language, run a second job over the same categories.

Research options

• Keyword research — this category’s target keywords go into the request, so the questions are phrased the way your customers phrase them.
• Use researched shopper questions — looks up which questions people actually search for on this subject and prefers those, as long as your product data can answer them. “Only used when keyword research is switched on as well.”
• Web research — stays disabled until a Trusted research domain is entered: “Only used with a trusted domain (e.g. your manufacturer’s site): FAQ answers must not come from arbitrary web pages.” An FAQ answer is a promise to your customers; it must not come from whatever page a search happened to return.

What a run costs

You are charged per category, not per suggestion: one item costs the task’s price whether it produces two suggestions or ten. Pushing costs the usual one credit per item. An item where you rejected everything, or where a suggestion is still undecided, is not sent at all: the confirmation page marks it Not sent and it costs nothing. The keyword- and web-research surcharges are only charged when a lookup actually ran; an answer from the cache is free. The numbers are on Credits.

After the push

Accepted suggestions become ordinary FAQ entries of your shop: active, inherited downwards, and carrying the Show on category pages / Show on product pages options the run was configured with. From then on you edit, deactivate or delete them in Shopware itself — we never touch an entry we have sent, and pushing again never creates a second copy.

The position inside your help center is the order of the suggestions: the first suggestion of an item becomes position 1, the second position 2. If you want them in a different order, reorder them in your shop’s FAQ bulk editing screen. Entries you already had are never renumbered. Optionally every entry carries the badge category picked in Help center category, which becomes a filter above the questions in the storefront.

What actually reached your shop is listed on the task page under Sent FAQs — with the date, the question, the target and the FAQ id from your shop. A second run over the same category reads those entries back and does not suggest them again.

Alt text

Sidebar: Media › Alt texts. Produces: one short descriptive text per picture and language, written for people using a screen reader — and read by search engines too.

This task actually looks at the picture. It therefore needs an AI model that can read images, and the wizard tells you so: “This task looks at your product pictures, so only AI models that can read images are offered below.” If no such model is enabled for your shop, the wizard says No suitable AI model available.

Media source

The most important setting on this task, a row of radio buttons that decides which pictures are worked on:

• Product images (the default) — the pictures attached to products. You then choose the products with the normal product picker below.
• Category images — the cover pictures of your categories.
• Shopping Experiences (CMS) — pictures used on your Shopping Experiences pages.
• Media folder — everything in one folder. A second dropdown appears, Choose a media folder….
• All media — the entire media library. Powerful, and easily the most expensive option in the whole app. Check the estimate on the summary page carefully.

Its other options

• Skip images that already have an alt-text — strongly recommended. “Skips a given image and language when that language already has a non-empty alt-text.” It is per language, so an image with a German alt text still gets an English one. Skipped items are never charged.
• With expert mode on: Cover image only (one alt-text per product) — only meaningful for the product source. Unticked, you get one alt text per language for every picture of every selected product.
• The prompts (System prompt, User prompt) are ordinary settings on the Alt text card in Settings › Configuration — you do not need expert mode for those. Use them to set the length and tone of your alt texts.

Alt texts are written onto the picture itself in your shop, not onto the product.

Image generator

Sidebar: Media › Image generator. Produces: free-standing pictures from your own descriptions — banners, mood shots, background images, anything not tied to a particular product.

This is the only task with no catalog subject: there is no product picker, no category tree and no language selection. Your prompts are the job.

Its options

• Image prompts — “Each prompt produces its own image(s). Use the count to request several variations of one prompt.” Write into the box marked Describe the image to generate… and press + Add prompt for another. Remove drops a row.
• With expert mode on, each row also has How many versions of this image (1–10). Ask for several versions when you want a choice; you are charged per picture produced, so three versions cost three times as much as one.
• Reference images — “Optional. Uploaded images are given to the model as inputs for every prompt above.” Use Choose files, drag and drop, or paste from the clipboard (click the upload area first, then paste). JPEG, PNG or WebP, up to 12 MB each. This is how you keep a house style, a product shape or a colour scheme consistent.
• Image shape — Square (1:1), Landscape (4:3), Wide (16:9), Portrait (3:4), for the whole job.
• Target media folder — where the pictures are filed in your shop’s media library, or Default (no folder).

Generated pictures go into your media library when you send them over. They are not attached to any product.

Marking images as AI-generated

All three picture tasks share one block of settings. It is off by default, and it is worth turning on.

What it does

The checkbox is Mark images as AI-generated: “Stamps a small ‘AI-generated’ label onto every image this task produces, and embeds a machine-readable AI marker in the file. Recommended for shop imagery: from August 2026 the EU AI Act expects AI-generated content to be recognisable as such.”

So there are two marks, not one: a small visible label in a corner of the picture, and an invisible marker inside the file that machines can read. Both are applied as the picture is produced, before it ever reaches your shop.

The four settings underneath

• Watermark text — what the label says. The default is KI-generiert; change it to “AI-generated” or whatever suits your shop’s language. Leave it empty and the default is used.
• Position — which bottom corner: Bottom left, Bottom centre or Bottom right (the default).
• Style — Dark badge (white text on a dark pill; stays legible on almost any picture, and the default), Light badge, or Text only (with shadow), which is the most discreet.
• Size — Small (default), Medium or Large. It scales with the picture, so a label looks the same on every format.
• With expert mode on: Opacity (%), between 10 and 100.

The marking is applied to newly generated pictures only. Pictures already in your shop are untouched, and turning the setting on later does not go back and mark what you generated before.

Research options

Two settings let the AI look things up before it writes. Both cost extra credits per item and both are therefore off by default. See Credits for the surcharges.

Web research

Available on Product description (as Web research), on Category images (as Look the category up on the web first) and on FAQ suggestions (there only once a trusted domain has been entered). It searches the web for the item and hands what it finds to the AI as extra input. Use it when your own product data is thin — a bare name and a number will produce a bland description, while a manufacturer’s page full of specifications will not.

Trusted research domain (product descriptions) and Only search this website (category images) limit the search to one address. Pointing it at the manufacturer’s own site is the usual reason: you get facts rather than whatever a price-comparison site says.

Which country’s results are used comes from Country for web research on the Global card in Configuration.

When a generated item used web research, its Review dialog shows the source at the bottom, labelled Web research source:, so you can check where a claim came from.

Keyword research

Available on Product description, Meta content, Category description and FAQ suggestions: “Use keyword-research suggestions to steer the generated content toward real search demand.” The AI is asked to weave the suggested phrases in naturally, adapting spelling and capitalisation to the language rather than pasting them in raw.

This requires the keyword-research capability to be switched on for your shop. If it is not, the whole block — checkbox and all its tuning fields — is hidden everywhere, and no external lookups are ever made.

FAQ suggestions adds a third option on top: Use researched shopper questions looks up which questions people actually search for on a subject. It is only used together with keyword research and carries no surcharge of its own — it travels in the same lookup.

With expert mode on you also get Keyword country, Keyword sources, Max keywords, Min search volume, Max competition and Keyword cache TTL (days). The cache setting is the one worth knowing about: results are reused for that many days before a fresh — and billable — lookup happens.

Where to go next

• Generating content — the job wizard step by step.
• Configuration — setting these options shop-wide.
• Templates — saving different answers for different jobs.
• Sending content to your shop — where each task’s output ends up.