Smart Catalog in Retail: 80% Less Registration Time with Gemini Enterprise

A smart catalog in retail powered by Gemini Enterprise is an automated system that uses multimodal AI to analyze product images and raw data, extract structured attributes, and generate SEO-optimized descriptions. For enterprises, this technology eliminates cataloging backlogs by reducing SKU registration time from 42 to 7 minutes while boosting product detail page conversion by 18 percent.

Product registration in a marketplace is the invisible work that decides the platform's success: poorly written descriptions, wrong attributes, non-standard photos = products that don't sell.

This case shows how a well-designed agent turned a bottleneck into a competitive advantage.

The starting point

• ~80k active SKUs, 1,200 new per week.
• 8 catalogers with an average time of 42 min per SKU.
• Typical backlog: 4-6 weeks — products reached the site 1 month after the supplier sent them.
• Attribute inconsistency: 23% of SKUs with broken site filters.
• Generic descriptions, copied from suppliers, with no SEO optimization.
• Translation to ES (Argentine market) took +1 week and contained errors.

What the agent does

1. Receives product photos (5-15 images) + supplier spreadsheet with raw specs.
2. Gemini 2.5 Pro multimodal analyzes the images: color, apparent material, shape, condition, angles, presence of defects.
3. Extracts and normalizes attributes according to the PIM schema (category, subcategory, brand, color, size, material, dimensions, weight).
4. Generates a structured spec sheet.
5. Writes a rich description SEO-optimized (250-400 words) with the brand's voice.
6. Suggests bullets and title for the PDP.
7. Translates to neutral ES (LATAM market).
8. Identifies similar SKUs already registered (deduplication).
9. Returns for human review: cataloger validates in 5-7 min and publishes.

Architecture

• Gemini Enterprise Plus + Vertex AI Agent Builder.
• Gemini 2.5 Pro multimodal: vision + text.
• Vertex AI Search: existing product base (deduplication) + brand glossary.
• Tools:
  • Reading a folder in Drive (supplier).
  • Product lookup by name/brand/UPC (PIM REST).
  • Draft SKU creation in the PIM.
  • Optimized image upload to CDN.
  • SKU code generation.
• Review panel in PWA for the cataloger to validate and publish.
• Quality pipeline in Cloud Run: schema tests, image weight, presence of required fields.

Results in 60 days

The 18% conversion gain is the number that definitively bought the project: descriptions written with copywriting + long-tail SEO techniques convert better than supplier copies. The agent democratized quality copy.

What we learned

1. Brand glossary is non-negotiable

Without a glossary, the agent writes correct but "neutral" descriptions. We built a document with brand voice, words to use/avoid, and award-winning description examples. Quality improved within 3 cycles.

2. Multimodal vision covered more than expected

For clothing, it identifies: sleeve type, neckline, cut, length, pattern (solid/printed/striped), apparent material type. For home: dominant color, style, function, suggested space. Reduced manual fields by ~70%.

3. Embedding-based deduplication avoided duplicates

Marketplace catalogs have 5-15% hidden duplicates. The agent compares visual + textual embedding with the existing catalog — the cataloger receives an alert before creating a duplicate.

4. SEO became an internal capability

Previously, SEO was outsourced to an agency. Now every SKU comes with a title tag, meta description, slug, and bullets designed for Google Shopping.

5. PT→neutral ES translation worked

Regional terms were glossary-mapped to neutral variants. The agent's automatic translation required 4% revision (vs ~30% in the initial pilot).

Mistakes we avoided

• We did not publish without human review. The "publish if confidence > 0.9" rule was tried and failed on descriptions with subtle factual errors. Manual review remains cheap at 7 min — it's worth the insurance.
• We did not use a single model for everything. Visual categorization runs Gemini Flash (faster); rich descriptions run Gemini Pro (higher quality). Cost/performance mix.
• We did not train a custom model in the MVP. Gemini 2.5 Pro was sufficient. Fine-tuning was put on the roadmap only if metrics stagnate.

Governance

• Supplier content has revised terms of use (internal use + AI processing).
• Images go through visual DLP: no identifiable human faces, no customer information.
• Audit log: each SKU has a history of which prompt version and model generated each field.
• Human reviewer remains legally responsible for published content.

+18% conversion bought the project. But the real gain was taking 5 catalogers out of "order taker" mode and returning their careers — today they do curation, merchandising, and catalog quality work that AI doesn't do well.

Replicability

The pattern applies to any medium/large marketplace or retailer with a PIM (VTEX, Salesforce Commerce, Shopify Plus, custom headless). Typical implementation time: 45-60 days.

Frequently Asked Questions sobre Smart Catalog in Retail: 80% Less Registration Time with Gemini Enterprise

What is the product registration time after the solution implementation? The product registration time has been reduced from 42 minutes to 7 minutes.

What is the impact on the conversion of product detail pages (PDP)? There was an 18% increase in the conversion of product detail pages (PDP).

What language model is used in the solution? The solution uses Gemini Enterprise Plus and Gemini 2.5 Pro multimodal.

How does the solution help in translating content to other languages? The solution translates the content to Spanish (ES) instantly.

Also read

• Dumb Search: 30% of Sales Lost in Retail
• Gemini Enterprise Connectors
• Is Your B2B Buying Like It's 1990?