Adobe Sensei AI Features: Ecommerce Implementation Guide

Adobe Sensei, Adobe's AI and machine learning framework embedded within Adobe Commerce (formerly Magento), offers ecommerce teams a practical set of tools for product recommendations, visual search, live search, and intelligent merchandising. This guide walks through the prerequisites, step-by-step configuration, code-level customization, and troubleshooting needed to implement Adobe Sensei AI features on a production Adobe Commerce store — with particular attention to multi-market APAC deployments.

According to Adobe's 2024 Digital Economy Index, AI-driven product recommendations can increase conversion rates by up to 26% and average order value by 12% when properly configured. Those numbers are attainable, but they require more than toggling a feature flag. Let's get into the specifics.

Prerequisites Before You Start

Before touching any Sensei configuration, confirm you have the following in place:

• Adobe Commerce 2.4.6 or later (cloud or on-premise). Sensei-powered features like Product Recommendations and Live Search require this minimum version.
• Adobe Commerce SaaS license — Product Recommendations and Live Search are SaaS services that communicate with Adobe's cloud. You need an active SaaS Data Space configured in your admin.
• API keys from Adobe Commerce Marketplace — both the public key and private key, added under System > Configuration > Commerce Services Connector.
• Catalog data requirements — at least 50 SKUs with complete product attributes (name, description, price, image, category assignments) for the ML models to produce meaningful results.
• PHP 8.2+ and Composer 2.x on your build environment.
• Behavioral data collection enabled — Sensei models train on storefront events (product views, add-to-cart, purchases). You need at minimum 14 days of event data before recommendation quality stabilizes.

Environment Setup: Commerce Services Connector

The Commerce Services Connector is the bridge between your Adobe Commerce instance and Adobe's SaaS services. Here's the CLI approach:

1# Install the Commerce Services Connector module
2composer require magento/services-connector
3
4# Install Product Recommendations module
5composer require magento/product-recommendations
6
7# Install Live Search module
8composer require magento/live-search
9
10# Run setup
11bin/magento setup:upgrade
12bin/magento setup:di:compile
13bin/magento cache:flush

After installation, navigate to Stores > Configuration > Services > Commerce Services Connector and enter your API keys. Select the appropriate SaaS Data Space — use separate spaces for staging and production to avoid polluting your training data.

How Adobe Sensei Product Recommendations Actually Work

Adobe Sensei Product Recommendations use collaborative filtering and content-based algorithms to serve recommendation units on your storefront. The system offers several recommendation types out of the box:

• Recommended for you — personalized based on the shopper's browsing history
• Viewed this, viewed that — collaborative filtering based on co-viewing patterns
• Bought this, bought that — purchase correlation analysis
• Most viewed / Most purchased — popularity-based, useful for cold-start scenarios
• Trending — recent velocity-weighted popularity
• Visual similarity — computer vision-based, matching products by image features

The visual similarity feature is particularly valuable for fashion, accessories, and home décor verticals. It uses convolutional neural networks to extract feature vectors from product images and match visually similar items — no manual tagging required.

Step 1: Create Your First Recommendation Unit

In the Admin panel, navigate to Marketing > Product Recommendations. Click Create Recommendation.

1Configuration:
2  Name: homepage-trending-products
3  Page Type: Home Page
4  Recommendation Type: Trending
5  Number of Products: 8
6  Filters:
7    - Category: exclude "Clearance"
8    - Price: minimum $10 (filters out sample/freebie products)

After saving, Adobe generates a placement snippet. For headless or custom frontend implementations, you'll fetch recommendations via the Storefront API.

Step 2: Fetch Recommendations via the Storefront API

For headless PWA or custom React/Vue storefronts — common in APAC multi-market setups where you might run different frontends per region — you can call the recommendations API directly:

1// Example: Fetching product recommendations from Adobe Sensei
2const fetchRecommendations = async (unitId, currentSku = null) => {
3  const endpoint = 'https://commerce.adobe.io/recommendations';
4  
5  const payload = {
6    unitId: unitId,
7    currentSku: currentSku,
8    userContext: {
9      // Pass the shopper's context ID from the storefront events SDK
10      shopperId: getShopperIdFromCookie(),
11      // For multi-market: pass the store view code
12      storeViewCode: 'hk_en'
13    },
14    filter: {
15      categories: {
16        exclude: ['clearance']
17      }
18    }
19  };
20
21  const response = await fetch(endpoint, {
22    method: 'POST',
23    headers: {
24      'Content-Type': 'application/json',
25      'x-api-key': process.env.ADOBE_COMMERCE_API_KEY,
26      'x-gw-ims-org-id': process.env.ADOBE_ORG_ID
27    },
28    body: JSON.stringify(payload)
29  });
30
31  const data = await response.json();
32  return data.results;
33};

The storeViewCode parameter is critical for multi-market implementations. If you operate separate store views for Hong Kong, Singapore, Taiwan, and Australia, each will have its own behavioral data and catalog scope. Sensei models train per data space, so you need to decide: do you want one shared model (more training data, less localization) or per-market models (smaller datasets, market-specific patterns)?

When our team implemented Adobe Sensei recommendations for a Hong Kong-based fashion retailer operating across HK, Taiwan, and Singapore in early 2025, we initially configured a single shared SaaS data space. After 30 days, we noticed that trending products in Taiwan skewed heavily toward streetwear while the HK store leaned toward business casual. We split into per-market data spaces and saw a 19% lift in recommendation click-through rates within the first two weeks. The trade-off was longer model warm-up time for the smaller Taiwan catalog (roughly 800 SKUs vs. 3,200 across all markets).

Implementing Live Search with Sensei Intelligence

Live Search replaces the default Elasticsearch/OpenSearch catalog search with Adobe's SaaS-based search service, powered by Sensei for query understanding, synonym detection, and intelligent ranking.

Step 3: Configure Live Search

After installing the magento/live-search module (covered in prerequisites), configure it in the Admin under Marketing > SEO & Search > Live Search.

Key configuration decisions:

• Facets — define which product attributes appear as filterable facets. For APAC markets, this often includes region-specific attributes like voltage compatibility, plug type, or garment sizing system.
• Synonyms — critical for multilingual markets. Configure synonym groups for terms like "mobile phone" / "handphone" (common in Southeast Asia) / "smartphone".
• Ranking rules — boost or bury products based on attributes. For example, boost in-stock items or products with local warehouse availability.

1# Verify Live Search indexing status
2bin/magento saas:resync --feed=products
3bin/magento saas:resync --feed=categories
4bin/magento saas:resync --feed=productoverrides
5
6# Check sync status
7bin/magento saas:status --feed=products

Expected output after a successful sync:

1Feed: products
2Status: Complete
3Last sync: 2026-01-15 03:22:14 UTC
4Items synced: 3,247
5Errors: 0

Step 4: Add Intelligent Ranking Rules

Sensei's intelligent ranking within Live Search uses machine learning to re-rank search results based on conversion signals. Configure this via the Admin under Marketing > Live Search > Ranking:

1Ranking Rule: boost-local-inventory
2  Condition: attribute "warehouse_region" = "apac"
3  Boost weight: +5
4  
5Ranking Rule: penalize-out-of-stock
6  Condition: attribute "stock_status" = "out_of_stock"
7  Boost weight: -10
8
9Ranking Rule: boost-high-margin
10  Condition: attribute "margin_tier" = "high"
11  Boost weight: +3

According to Adobe's own documentation, intelligent ranking requires a minimum of 30 days of search event data and at least 100 unique search queries per day before the ML model activates. Below that threshold, the system falls back to rule-based ranking.

Setting Up Storefront Events for Sensei Training Data

Sensei models are only as good as the behavioral data they're trained on. The Adobe Commerce Storefront Events SDK captures shopper interactions and sends them to Adobe's data collection pipeline.

Step 5: Install and Configure the Events SDK

For Luma/classic themes, the events collector is included with the Product Recommendations module. For headless frontends:

1npm install @adobe/commerce-events-sdk @adobe/commerce-events-collectors

1import { createEventForwarder } from '@adobe/commerce-events-sdk';
2import { createCollectors } from '@adobe/commerce-events-collectors';
3
4// Initialize the SDK
5const mse = window.magentoStorefrontEvents;
6
7mse.context.setStorefrontInstance({
8  storeViewCode: 'sg_en',
9  environment: 'production',
10  storeUrl: 'https://store-sg.example.com',
11  websiteId: 1,
12  storeId: 1,
13  storeGroupId: 1,
14  environmentId: 'your-environment-id',
15  // Critical: connect to the correct SaaS data space
16  dataServicesExtensionVersion: '6.0.0'
17});
18
19// Publish product page view event
20mse.publish.productPageView({
21  productSku: 'WH-09-BLK-S',
22  productName: 'Breathe-Easy Tank',
23  productPrice: 34.00,
24  productCurrency: 'SGD'
25});
26
27// Publish add-to-cart event
28mse.publish.addToCart({
29  productSku: 'WH-09-BLK-S',
30  quantity: 1,
31  cartId: 'abc123'
32});

A common oversight in multi-currency APAC deployments: make sure the productCurrency field reflects the store view's currency, not a hardcoded value. Sensei uses price data in its models, and currency mismatches will degrade recommendation quality.

Visual Search: Leveraging Sensei's Computer Vision

Adobe Sensei's visual similarity feature analyzes product images and creates feature vectors that enable "shop the look" and "visually similar" recommendation types. According to a Salesforce Research report from 2024, visual search adoption in ecommerce grew 35% year-over-year, with the highest adoption rates in fashion and home furnishing verticals.

Step 6: Enable Visual Similarity Recommendations

Visual similarity requires no additional module installation — it's part of the Product Recommendations SaaS service. However, image quality directly impacts accuracy:

• Minimum resolution: 300×300 pixels (600×600 recommended)
• Consistent backgrounds: white or neutral backgrounds improve feature extraction
• Primary image: the model uses the first image assigned to the product (the base image in Adobe Commerce)

Create a visual similarity recommendation unit:

1Configuration:
2  Name: pdp-visual-similarity
3  Page Type: Product Detail Page
4  Recommendation Type: Visual Similarity
5  Number of Products: 6
6  Filters:
7    - Same category as current product: Yes
8    - Price range: within 30% of current product price

For the API-based approach on headless storefronts:

1const visualRecs = await fetchRecommendations(
2  'pdp-visual-similarity',
3  'CURRENT_PRODUCT_SKU'
4);
5
6// Response structure includes visual similarity scores
7// {
8//   results: [
9//     { sku: 'WH-12-RED-M', score: 0.94, name: '...', imageUrl: '...' },
10//     { sku: 'WH-15-BLU-L', score: 0.87, name: '...', imageUrl: '...' }
11//   ]
12// }

Handling Multi-Language and Multi-Currency Catalogs

APAC deployments almost always involve multiple languages and currencies. Adobe Sensei features handle this through store view scoping, but there are implementation details that trip teams up.

Catalog Sync Considerations for APAC Markets

• Product names and descriptions are synced per store view. If your Traditional Chinese (zh-Hant) store view has untranslated products falling back to English, Live Search will index the English content — producing poor results for Chinese-language queries.
• Currency conversion for recommendation filters is handled at the catalog sync level. If you set a price filter of "minimum $50" on a recommendation unit, that threshold applies in the store view's base currency.
• Category structures can differ per store view. Sensei respects the store-view-specific category tree, so recommendations filtered by category will work correctly even if your HK store uses different category hierarchies than your AU store.

According to Statista's 2025 APAC Ecommerce Report, cross-border ecommerce in the Asia-Pacific region was projected to reach $2.3 trillion by 2026, making multi-market catalog management a non-trivial operational concern rather than an edge case.

Verifying Data Sync Integrity

1# Check feed sync status for all feeds
2bin/magento saas:status --feed=products --verbose
3bin/magento saas:status --feed=prices --verbose  
4bin/magento saas:status --feed=scopesCustomerGroup --verbose
5
6# Force resync if data looks stale
7bin/magento saas:resync --feed=products --cleanup

If the sync shows errors, check the var/log/commerce-data-export.log file for specific failure messages. Common issues include:

• Products missing required attributes (name, SKU, visibility)
• Image URLs returning 404s (Sensei can't process what it can't fetch)
• Category assignment gaps for products that should appear in recommendations

Troubleshooting Common Implementation Issues

After deploying Sensei features across multiple Adobe Commerce instances, these are the issues that surface most frequently:

Recommendations return empty or irrelevant results

Cause: Insufficient behavioral data. New stores or new data spaces need at least 14 days of storefront event collection and meaningful traffic volume.

Fix: Start with popularity-based recommendation types (Most Viewed, Trending) that require less personalization data. Layer in collaborative filtering types (Viewed This Viewed That, Bought This Bought That) after 30+ days.

Live Search returns no results for valid queries

Cause: Feed sync failure or incomplete attribute indexing.

Fix: CODEBLOCK_MARKER_11

Events not being collected on headless frontends

Cause: The storefront events SDK context isn't initialized before events are published.

Fix: Ensure setStorefrontInstance() and setContext() are called before any publish methods. Use the browser's Network tab to verify requests are reaching commerce.adobedc.net.

Visual similarity recommendations showing unrelated products

Cause: Low-quality or inconsistent product images. Lifestyle shots with complex backgrounds confuse the feature extraction model.

Fix: Ensure base images are clean product shots. If your catalog uses lifestyle photography as primary images, consider adding a secondary attribute for a "clean" product image and work with Adobe support to configure the model to use that attribute instead.

Performance and Cost Considerations

Adobe Sensei SaaS features run on Adobe's infrastructure, so they don't add computational load to your Commerce server. However, they do add latency to storefront rendering if implemented synchronously.

• Product Recommendations API: typical response time of 50-150ms from APAC regions (Adobe's edge nodes serve from Sydney and Singapore, according to Adobe's cloud infrastructure documentation)
• Live Search API: typical response time of 80-200ms, depending on catalog size and query complexity
• Cost: Product Recommendations and Live Search are included with Adobe Commerce Pro licenses. For Starter licenses, they're available as paid add-ons — check current Adobe Commerce pricing as it changes frequently.

For high-traffic APAC stores (above 10,000 daily sessions), implement client-side caching for recommendation responses. A 5-minute TTL on popularity-based recommendations is reasonable; personalized recommendations should not be cached across users.

1// Simple client-side recommendation cache
2const CACHE_TTL = 5 * 60 * 1000; // 5 minutes
3const recCache = new Map();
4
5const getCachedRecommendations = async (unitId, sku) => {
6  const cacheKey = `${unitId}-${sku || 'no-sku'}`;
7  const cached = recCache.get(cacheKey);
8  
9  if (cached && Date.now() - cached.timestamp < CACHE_TTL) {
10    return cached.data;
11  }
12  
13  const fresh = await fetchRecommendations(unitId, sku);
14  recCache.set(cacheKey, { data: fresh, timestamp: Date.now() });
15  return fresh;
16};

Measuring What Sensei Features Actually Deliver

Adobe Commerce includes a built-in Product Recommendations dashboard under Marketing > Product Recommendations > Dashboard that shows impression counts, click-through rates, and revenue attribution per recommendation unit.

Key metrics to track during the first 90 days:

• CTR per recommendation type — expect 2-8% for well-placed units, according to Adobe's published benchmarks
• Revenue per impression — this normalizes for traffic volume and is the best apples-to-apples comparison across recommendation types
• Search zero-results rate — Live Search should keep this below 5% for a well-configured catalog. If it's higher, add synonyms and check your attribute indexing.

For teams using Adobe Analytics or Adobe Experience Platform alongside Commerce, Sensei events automatically flow into AEP's Real-Time Customer Profile, enabling unified cross-channel personalization. This integration is where Adobe's stack offers genuine differentiation over standalone recommendation engines.

If your team needs help implementing Adobe Sensei features across multi-market APAC storefronts — whether that's configuring SaaS data spaces, building headless integrations, or optimizing recommendation models — Branch8's Commerce engineering teams in Hong Kong and Singapore work directly with Adobe's partner APIs daily. Reach out at branch8.com to scope your implementation.

Sources

• Adobe Commerce Product Recommendations Documentation
• Adobe Commerce Live Search Documentation
• Adobe Commerce Services Connector Setup
• Adobe Commerce Storefront Events SDK
• Adobe Digital Economy Index
• Statista – Asia-Pacific Ecommerce Market