Product descriptions for SKU catalog (SEO optimization)

An AI agent based on a low-code scenario takes raw product data and turns it into a ready SEO description for a SKU card. The editor receives a draft that only needs to be read and adjusted, not written from scratch. According to Reactively reports, the cost of one description drops from £64 to £6.40, while at Fnac Darty draft accuracy holds at 95%.

Specific process steps:

1. The agent retrieves the product card from CMS or file storage (CSV, XLSX, JSON export from PIM).
2. Normalizes attributes: category, brand, material, size, compatibility, key characteristics.
3. Pulls target keywords and SEO rules for the corresponding category.
4. Generates three blocks: a short title/lead, an expanded description, and bullet points with characteristics.
5. Checks the draft against a checklist: length, keyword presence, prohibited phrasing, tone of voice.
6. Saves the result back to CMS as a card draft and marks it with the status "under review".
7. A copywriter or category manager reads, edits, and publishes.

What automation does NOT do:

• Does not replace the category manager and editor. The final decision on publishing and product positioning remains with a human.
• Does not verify the factual accuracy of characteristics. If there is an error in the source data (wrong size, incompatible SKU), it will appear in the description as well.
• Does not replace a full SEO strategy. The agent writes within the defined rules and keywords, but does not form a query cluster and does not decide which categories to promote.

Automation addresses three e-commerce catalog pain points at once: slow creative output when launching new collections, repetitive routine tasks for copywriters, and manual data entry/copying between PIM and CMS. As a result, the marketing team spends time on editing and strategy, not on rewriting characteristics in their own words.

At the core is a low-code scenario that connects the product data source, LLM, and CMS. The architecture is built on a workflow engine or similar orchestrator and does not require a dedicated backend for the task.

The technical flow looks like this:

1. Trigger. A new SKU appears in the CMS/PIM, or a file with a batch of products is uploaded to file storage. The scenario catches the event or runs on a schedule.
2. Context collection. The agent gathers product attributes, image links, category, brand, target audience, and keywords from a separate SEO rules table.
3. Prompt preparation. The data is templated into a structured prompt with strict constraints: tone of voice, length, prohibited words, required sections (lead, description, bullets).
4. LLM call. The agent queries the model and receives a draft in the specified JSON format.
5. Validation. The scenario checks length, keyword presence, and the absence of hallucinations in attributes (every number and attribute is verified against the source data).
6. Result writing. The draft is placed back into the product card in the CMS with the status "under review" and the AI-draft tag.
7. Editor notification. The copywriter receives the task in the familiar CMS interface or in a separate queue.

Implementation steps:

1. Week 1. Catalog audit, selection of the pilot category (200-500 SKU), collection of examples of "good" descriptions for few-shot.
2. Week 1-2. Setting up access to CMS and file storage, describing the product data schema.
3. Week 2. Building the scenario in the low-code orchestrator, first run on 20-50 SKU, prompt calibration.
4. Week 2-3. Validation rules, connecting the SEO dictionary, editor review of the first batch.
5. Week 3-4. Launch across the full pilot category, collecting metrics on edit time and the share of accepted drafts, decision on scaling.

Scenario components:

To scale to the full catalog, it makes sense to split the scenario into two queues: a fast stream for new SKU (event-driven) and a batch run for the bulk regeneration of old descriptions, to avoid hitting API limits and disrupting the editors' review process.

Automation relies on product data already existing in a structured form with programmatic access to it. Without this, the agent will write polished but reality-detached texts.

What you need in terms of data and access:

• A product catalog in a CMS or PIM with attributes (name, category, brand, key specifications, images).
• API access or export to file storage (CSV/XLSX/JSON) for reading product cards and writing drafts.
• SEO dictionary: a table of keywords and rules by category (minimum for the pilot category).
• 20-50 reference descriptions that the team considers 'the right way' — for few-shot and style validation.
• A tone of voice guide and a list of prohibited phrases.

What you need from the team:

• A category manager or marketer — the process owner, makes decisions on the description structure.
• A copywriter/editor — reviews drafts and provides feedback in the first iterations.
• A technical executor — sets up the low-code scenario and integrations with the CMS.
• Readiness to work with drafts: the team must accept that an AI description is a draft that gets edited, not published automatically.

Implementation timeline: 2-4 weeks for a pilot in one category (weekend complexity with ready data, up to 4 weeks if the catalog needs to be put in order). Scaling to the full catalog — another 2-4 weeks depending on the number of categories and the quality of the source attributes.