GenAI for E-commerce Ops: Catalog Enrichment, Search, and Merchandising on GCP

E-commerce ops is the category our team has spent the most time working in. Cataloging, search relevance, merchandising, returns handling, customer messaging. The economics are unforgiving in a way that’s clarifying: a 2% lift in product-listing conversion shows up as real money next quarter. There’s no philosophical debate about whether AI is useful. There’s only a question of whether this AI feature, in this part of the customer journey, lifts the right metric.

Google Cloud has a specifically strong product stack for this category. AI Commerce Search (formerly Vertex AI Search for commerce; incorporates Recommendations AI), Gemini’s multimodal capabilities for product imagery, and BigQuery for the analytics substrate. This post is the reference set of patterns we deploy.

The use case landscape

Four areas where GenAI consistently delivers value in e-commerce ops:

Each has a clean architecture pattern. We’ll walk through them.

Catalog enrichment

This is the highest-volume use case for most e-commerce platforms. The shape of the problem:

• Some products have thin or missing data (no description, missing attributes, low-quality title).
• Some products have inconsistent data across SKUs that should be similar.
• New products arrive constantly and need enrichment before they can rank well in search or be merchandised.
• Marketplaces with seller-provided content need editorial quality controls.

Gemini’s multimodal capabilities (image and text together) are well-suited here. Given a product image and the existing partial data, Gemini can generate complete listings that match the platform’s style and quality bar.

ENRICHMENT_PROMPT = """
You are completing the listing for a product. Use the image and the existing data
to generate a complete listing in our standard format.

Style requirements:
- Title: 60-80 chars, no ALL CAPS, no excessive punctuation
- Description: 2-3 paragraphs, conversational but professional
- Attributes: select from the controlled vocabulary provided
- Search keywords: 5-10 terms a shopper would actually use

Existing data:
{existing}

Controlled attribute vocabulary for this category:
{vocab}

Output JSON matching the schema. Be specific about visible product features.
Don't invent claims you can't see in the image or existing data.
"""

from google import genai
from google.genai import types

client = genai.Client(vertexai=True, project=PROJECT_ID, location=LOCATION)

async def enrich_product(product_id: str, image_uri: str, existing_data: dict):
    response = await client.aio.models.generate_content(
        model="gemini-2.5-flash",
        contents=[
            types.Part.from_uri(file_uri=image_uri, mime_type="image/jpeg"),
            ENRICHMENT_PROMPT.format(
                existing=json.dumps(existing_data),
                vocab=json.dumps(get_category_vocab(existing_data["category"])),
            ),
        ],
        config=types.GenerateContentConfig(
            temperature=0.3,
            response_mime_type="application/json",
            response_json_schema=ENRICHMENT_SCHEMA,
        ),
    )
    enriched = response.parsed

    # Validation: don't auto-publish, route to merchant review
    quality_score = score_enrichment(enriched, existing_data)
    if quality_score > 0.85:
        return {"status": "ready_to_publish", "data": enriched}
    return {"status": "needs_review", "data": enriched, "score": quality_score}

The architecture:

Product event (new SKU or thin data)
   │
   ▼ Pub/Sub
   │
   ▼
Cloud Run worker
   │
   ├──► Gemini multimodal (image + existing data → enriched listing)
   │
   ├──► Validation: schema, controlled vocabulary, length checks
   │
   ├──► Translation: Gemini for additional locales if multi-region catalog
   │
   └──► Output:
        High confidence ──► auto-publish to catalog
        Medium confidence ──► merchant review queue
        Low confidence ──► reject, log for diagnosis

Watch out for: hallucinated attributes (“Material: Cotton” when the product is polyester), promotional language that violates platform policies, and locale-specific style violations. The validation layer catches these.

Search relevance

E-commerce search is the canonical case for retrieval-augmented behavior. Customers don’t search the way the catalog is labeled. They search “running shoes for flat feet under $100,” not “athletic footwear, neutral support, MSRP $50-100.”

AI Commerce Search (the rebranded Vertex AI Search for commerce, which now incorporates the Recommendations AI personalization capabilities) is the GCP product purpose-built for this. It combines:

• Semantic understanding via Gemini-derived embeddings
• Lexical retrieval for exact terms (brand names, model numbers, SKUs)
• Merchant-controlled ranking (promote SKUs, demote others, boost on margin, etc.)
• Personalization signals (the former Recommendations AI surface, now folded in)
• A/B testing built into the platform

For most e-commerce builds, the right pattern is to use AI Commerce Search as the search backend and layer custom logic on top for very specific behaviors. Don’t roll your own from scratch unless you have requirements it genuinely can’t meet.

Where Gemini adds value on top of AI Commerce Search:

• Query rewriting. “running shoes for flat feet” gets expanded to include relevant attribute filters (support type: stability) plus the natural language phrase.
• Conversational search. A chat-style interface where the customer refines results across turns. “Show me more colorful ones.” “Under $80.” “Something with better arch support.”
• Result explanation. “Why is this the top result?” for transparency. Useful for merchant tooling and for high-consideration purchases.

Merchandising assistance

Merchants spend a lot of time on tasks that are creative-judgment-heavy but assistance-friendly: building collections, writing campaign copy, choosing hero products for a category page, drafting promotional emails, A/B test hypotheses for category pages.

A merchandising copilot pattern:

Merchant request: "Build a Mother's Day collection for the home goods category"
   │
   ▼
Gemini Agent (Vertex AI Agent Builder, now part of the Gemini Enterprise Agent Platform — the Cloud Next 2026 umbrella covering Agent Builder, ADK, Agent Engine, Agent Studio, Agent Garden, and Agentspace)
   │
   ├──► Tool: query_products(filters)
   │   Returns: candidate products matching criteria
   │
   ├──► Tool: get_product_performance(skus, period)
   │   Returns: conversion, AOV, return rate per SKU
   │
   ├──► Tool: get_inventory(skus)
   │   Returns: stock levels, ensuring no out-of-stock items
   │
   ├──► Tool: query_recent_campaigns(theme)
   │   Returns: past performance of similar themed collections
   │
   └──► Output: proposed collection (20 SKUs) + rationale + campaign copy draft
        Routes to merchant for review

The agent doesn’t auto-publish. It proposes. The merchant edits and publishes. This is the agent-assist pattern from our customer support post, applied to a different workflow.

Operational AI

Several smaller patterns that compound:

Returns triage. Customer returns include free-text reason. Classify it (defective, wrong size, didn’t match description, changed mind), pull product and order context, route to the right team. gemini-2.5-flash does this in a single call.

Dispute summarization. A customer has 8 messages in a support thread about a delayed order. Generate a 3-sentence summary for the support agent who picks up the case. Faster context-loading for the human.

Review moderation. Filter out spam, fake reviews, policy-violating content. Gemini’s safety filters plus a domain-specific classifier handle most cases; human review on the edge.

Out-of-stock detection from images. For physical retail or photo-based catalog, detect when shelf or product images don’t match expectations. Trigger replenishment workflows.

Data architecture

The backbone for any of these is having product, order, and behavioral data colocated and queryable. Our default:

┌──────────────────────────────────────────────────────────────────────────┐
│  Source systems                                                            │
│  Commerce platform (Shopify, BigCommerce, custom), PIM, OMS, CDP           │
└────────┬─────────────────────────────────────────────────────────────────┘
         │
         ▼ Datastream or Cloud Data Fusion or custom ELT
         │
┌────────▼─────────────────────────────────────────────────────────────────┐
│  BigQuery                                                                  │
│  - Product catalog (curated)                                               │
│  - Order history                                                           │
│  - Behavioral events (clickstream, search queries)                         │
│  - Inventory snapshots                                                     │
│  - Review and support ticket text                                          │
└────────┬─────────────────────────────────────────────────────────────────┘
         │
         ▼ Embeddings pipeline (Dataflow or Cloud Run jobs)
         │  (or skip — Vector Search 2.0 supports auto-embeddings, where you
         │   store text and the service generates vectors server-side)
         │
┌────────▼─────────────────────────────────────────────────────────────────┐
│  Vertex AI Vector Search 2.0 / AlloyDB pgvector                            │
│  - Product embeddings for similarity                                       │
│  - Behavioral embeddings for personalization                               │
│  - Vector Search 2.0 collapses index/endpoint/feature-store into a single  │
│    Collection, adds built-in hybrid search and self-tuning ANN params      │
└──────────────────────────────────────────────────────────────────────────┘

AI Commerce Search sits on top, integrating with the structured catalog and the embeddings.

BigQuery as the substrate matters because it’s where the analytical questions get answered. “Which enriched listings outperformed the originals by conversion rate?” “What’s the relationship between search query reformulation and purchase?” These are SQL queries, not RAG queries. The eval harness for the GenAI features lives in BigQuery too. See our evaluation guide.

Vertical specifics

A few things particular to e-commerce:

Seasonality. Eval sets that worked in March don’t reflect Q4 traffic patterns. Refresh the eval set seasonally, especially for search.

Promotional periods. During Black Friday, every metric is anomalous. Don’t deploy model changes during peak season; don’t trust drift detection signals during peak season.

Inventory dependence. A search result for an out-of-stock item is worse than no result. Inventory state has to be a first-class input to ranking, recommendation, and merchandising features.

Regulatory and platform compliance. Marketplaces have policies about claims, accessibility, and prohibited content. The enrichment and merchandising AI has to respect these. Build the constraints into prompts and validators.

Where our experience fits

Our team’s background in e-commerce gives us a deep starting point for these engagements. We’ve worked on product cataloging, search, and merchandising problems for years, and the GenAI capabilities now available are the next chapter in problems we already understand. That domain experience includes shipping work on Prism, the e-commerce analytics product from Calibrated Intelligence (a separate firm), where members of our team built and scaled retail data and ML systems. The credibility carries across.

How Accelyze helps

Accelyze designs and builds e-commerce GenAI on Google Cloud across catalog enrichment, search relevance, merchandising assistance, and operational AI. Engagements typically start with a use-case ranking exercise (which features will move which metrics), an eval design specific to commerce metrics, and a phased rollout that proves ROI before scaling. If you’re considering GenAI in your e-commerce operations and want a team with domain depth as well as GCP depth, get in touch.