Enhancing Product Discovery in E-Commerce

Excited to share insights from Walmart 's groundbreaking semantic search system that revolutionizes e-commerce product discovery! The team at Walmart Global Technology(the team that I am a part of 😬) has developed a hybrid retrieval system that combines traditional inverted index search with neural embedding-based search to tackle the challenging problem of tail queries in e-commerce. Key Technical Highlights: • The system uses a two-tower BERT architecture where one tower processes queries and another processes product information, generating dense vector representations for semantic matching. • Product information is enriched by combining titles with key attributes like category, brand, color, and gender using special prefix tokens to help the model distinguish different attribute types. • The neural model leverages DistilBERT with 6 layers and projects the 768-dimensional embeddings down to 256 dimensions using a linear layer, achieving optimal performance while reducing storage and computation costs. • To improve model training, they implemented innovative negative sampling techniques combining product category matching and token overlap filtering to identify challenging negative examples. Production Implementation Details: • The system uses a managed ANN (Approximate Nearest Neighbor) service to enable fast retrieval, achieving 99% recall@20 with just 13ms latency. • Query embeddings are cached with preset TTL (Time-To-Live) to reduce latency and costs in production. • The model is exported to ONNX format and served in Java, with custom optimizations like fixed input shapes and GPU acceleration using NVIDIA T4 processors. Results: The system showed significant improvements in both offline metrics and live experiments, with: - +2.84% improvement in NDCG@10 for human evaluation - +0.54% lift in Add-to-Cart rates in live A/B testing This is a fantastic example of how modern NLP techniques can be successfully deployed at scale to solve real-world e-commerce challenges!

Rufus is an AI designed to revolutionize product discovery through natural language understanding, inference, and multimedia optimization. Here's how it works and how sellers can use it to boost their sales. Rufus changes the rules of product discovery by focusing on context, not just keywords. Instead of matching queries like "desk lamp" to products with the same exact words, Rufus identifies noun phrases and their relationships. For example: 1. A shopper asks: "What lamp is best for reading in bed?" 2. Rufus identifies key phrases like “reading lamp” and “bedside.” 3. It ranks products semantically, recommending items with phrases like “adjustable bedside reading lamp with eye-friendly light.” This ensures shoppers see relevant, high-quality products tailored to their needs. Key Features  1. Noun Phrase Optimization (NPO): Rufus focuses on detailed, descriptive phrases. Sellers should build product titles and descriptions differently: ▪️ Instead of: "Table Lamp" ▪️ Use: "Vintage Brass Table Lamp with Adjustable Arm for Home Office." 2. Visual Label Tagging (VLT): Rufus reads images as well as text. Adding overlays like “Energy Efficient | 6 Brightness Levels” directly on product images can increase discoverability. 3. Semantic Understanding: Rufus connects implied customer needs to product benefits. For example, it knows “easy-to-clean” is relevant for a query like “pet-friendly couch.” 4. Q&A Enhancement: Rufus thrives on clear answers to common customer questions. Example: Q: “Does it fit a queen-size mattress?” A: “Yes, our bed frame is designed for all queen-size mattresses up to 12 inches thick.” 5. Inference Optimization: Rufus maps product features to inferred benefits. A product labeled “durable non-stick pan” might also be shown for “easy-to-clean cookware.” Steps Sellers Need to Take 1. Optimize Product Titles with Rich Noun Phrases ▪️ Use descriptors like material, design, and purpose. Example: “Professional Chef Knife Set with German Steel Blades”. 2. Enhance Images with Text ▪️ Include labels like “Anti-Fog Coating | Shatterproof Design” directly on images. ▪️ Ensure images demonstrate key features clearly 3. Leverage FAQs ▪️ Anticipate shopper questions and weave them into your listings. Example: Q: “How do I clean this air fryer?” A: “Wipe with a damp cloth or place removable parts in the dishwasher.” 4. Use Semantic Context in Descriptions ▪️ Avoid keyword stuffing; write naturally. Example: “This ergonomic office chair supports your back during long hours at your desk, making it perfect for work-from-home setups.” 5. Update Content Regularly ▪️ Monitor trends in customer queries and adapt your listings accordingly. If shoppers search for “eco-friendly packaging,” ensure your products highlight those features. 6. Incorporate Click Training Data Insights ▪️ Analyze which features customers click on most and highlight them in your product content. Amazon’s Rufus thrives on detailed, customer-centric content.

AI search has changed how ecommerce visibility works. Rankings and keywords still matter, but they no longer determine whether you appear in AI answers. LLMs pull from many sources and validate information across channels, so brands now need a practical strategy to influence what these systems surface. Here is the actionable version of the playbook: 1. Strengthen the brand mention layer These mentions usually come from Reddit, Inc., media coverage, YouTube reviews, and user sentiment. To increase volume and quality: • Seed conversations on Reddit via customer outreach and product education. • Pitch journalists with data, not product claims. • Send reviewers standardized product kits with accurate specs. • Increase review velocity on Amazon, Walmart, and your own site. Goal: create widespread awareness that models can pick up reliably. 2. Win citations by becoming a source of truth Citations influence how the model describes your brand. To increase them: • Publish structured, factual resources. Examples include comparison charts, ingredient or materials breakdowns, and step by step usage guides. • Make your product pages machine readable with flawless schema markup, identical naming, and consistent specs. • Update all your public product data quarterly. Goal: give LLMs clean, verifiable information they can quote confidently. 3. Influence product recommendations This is the highest value layer. To increase recommendation frequency: • Get included in publisher listicles and buying guides through affiliate programs or product samples. • Make sure your product appears in retailer categories with high review counts and strong Q and A sections. • Encourage customers to mention specific attributes in their reviews so AI models learn your positioning. • Publish expert testing results wherever possible. Goal: fit the queries that drive buying decisions, not just general awareness. 4. Build consensus across independent sources Models reward brands whose reputation looks the same everywhere. To create this: • Audit every major source in your category twice a year. Look at Amazon sentiment, YouTube reviews, TikTok results, Reddit threads, niche forums, and publisher guides. • Identify mismatched attributes, outdated specs, or conflicting feature claims. Fix them one by one. • Monitor competitor positioning across these same channels to understand why they appear more often in AI answers. Goal: eliminate contradictions so LLMs treat your brand as the consistent default. 5. Fix consistency across every product feed LLMs cross check your data across Amazon, Shopify, Walmart, Google Merchant, and any structured source. To avoid exclusion: • Standardize SKU names and attribute formats. • Keep pricing aligned within a narrow range. • Use the same dimensions, specs, and materials everywhere. • Remove outdated product copies from old marketplaces. Goal: reduce confusion so AI systems trust your data. (Find Step 6 and 7 in the comment section)

Incredible Shoptalk conversation this week with Jason Del Rey from The Aisle, Ashye Marcus from Stripe and Kristina Elkhazin from Klarna on agentic commerce and what brands need to do to get their products recommended in this fast growing discovery channel. Carla DeSantis from PwC said it well: Brand sentiment and awareness used to be the center of the show and the vehicle through which products were brought to market. Now, we’re seeing that flip with the role of brand sentiment being disrupted and product-level data and attributes becoming the most important (yet previously neglected) component of getting your product recommended in AI. Key takeaways: 1/ The 3 problems brands & retailers face in discovery: - AI can’t see your products at all (unreadable / non-existent catalog feed) - AI can see your products but doesn’t recommend you (low-quality, low trust attribute data) - AI can see you and recommends you, but in the wrong contexts & conversations (wrong attribute data) 2/ What is causing this: - Existing catalog/merchant schemas are basic and generic. - Different product types (e.g., face wash vs socks vs peanut butter) require very different attribute data. - LLM-driven discovery needs deeper category and intent-specific product data than what current feeds provide. 3/ The solution: - Know what category contexts you’re trying to win in (what products are you being compared to and considered against) - Infuse that category specific data into your product schema (introduce the information needed for products to show up in the conversations you care about) - Get that data out consistently, everywhere (put it in all of the channels you sell across the internet) 4/ Lastly, this doesn’t just apply to brands. Retailers need to make sure that when a product is recommended, their store is the place the consumer is being sent to buy it. This requires the same “golden catalog” data. If your feed is full of garbage, AI will ignore you as well. 

Commerce search isn't just about cramming keywords into a title anymore. Shoppers are treating Amazon like a knowledgeable store clerk. If your product pages aren't ready for a real conversation, you are going to lose sales to the brands that are. This massive shift is driven by AI shopping assistants like Amazon's Rufus. People are moving away from typing basic, two-word searches. Instead, they are asking complex, highly specific questions. 🤔 For example, they aren't just searching for a stroller. They want to know if a specific model will easily fold up and fit in the trunk of a compact car. If your listing only relies on traditional keyword lists, Rufus won't have the context to recommend your brand. It is time to rethink how we build our product pages. The focus must shift to answering the deep, situational questions your buyers are already thinking about. Here is how you can adapt your listings to win in this new era of AI discovery: ➡️ Solve for constraints: Tell shoppers exactly what your product can and cannot do so the AI knows when to confidently suggest it to the right person. ➡️ Define your ideal buyer: Call out who your product is perfect for and who should look elsewhere, which naturally reduces your return rates. ➡️ Highlight everyday scenarios: Turn technical specs into real-world outcomes, like explaining that a suitcase actually fits in an overhead bin. ➡️ Build a knowledge base: Treat your product page like an FAQ section that addresses common concerns pulled directly from your customer reviews. These updates do much more than just feed the algorithm. They give your human customers the absolute confidence they need to finally check out. 🙂 Take an hour this week to look at your top-selling products and see if they genuinely answer the questions your customers are asking. You have everything you need to adapt and own this new landscape. https://lnkd.in/gS9iYnBJ #AmazonRufus #EcommerceStrategy #RetailMedia #AISearch #MarketingTips

There's been a lot of buzz these past two weeks since ChatGPT launched both Atlas and announced its partnership with Walmart — amongst other AI-related announcements — and with each we've heard one question over and over from brand and retail leaders: what does this mean for us? It's clear that AI is having a profound impact on product discovery, and that discovery is the new marketing battlefield, something to be studied and mastered. A recent study by Yale and Columbia, "What Is Your AI Agent Buying?", shows a profound causal relationship between what AI-driven summaries and shopping agents recommend and improvements in average ratings and review volume (study here: https://my.ugc.bz/9ZtrnK). It turns out that these are huge trust signals for AI, just as they are for humans. And so our take at Bazaarvoice is simple: You have two consumers of content — AI and people. You need to feed AI if you want to have your products be discovered and chosen to be featured in its outputs. What do you feed it with? The content that it trusts: Ratings, reviews, and other user-generated content. So it's clear: Ratings and reviews are essential for your products to be discovered and surfaced by AI, and then trusted and chosen. And that UGC must be Accessible, Authentic, and Abundant — what we call the Triple-A framework. Here's what that looks like in practice: → Accessible: Clean, centralized, tagged, and AI-ready content that LLMs can easily ingest and surface accurately. → Authentic: Verified, trustworthy content. AI trusts what people say about your products more than what you say about them. → Abundant: High volume, fresh UGC, specifically reviews. Recency and depth are the new currency of AI discoverability. My recommendation to you: Start stress-testing where your brand stands today on these three points as soon as possible because the brands that master Triple-A will be the ones AI and shoppers will find first. The LLMs like ChatGPT and Gemini are evolving quickly, and data hygiene is critical. Here at Bazaarvoice, our innovation roadmap is geared specifically towards improving Triple-A, with a specific focus on GEO and ASEO, backed by our immense network reach, content scale, and authenticity efforts. We'd love to hear from you.

🎨 Google 's visual #AIMode update wasn't random. It’s part of a deliberate product and e-commerce strategy playing out across their entire AI Search Ecosystem. Here's how the dots connect between Google’s recent updates, product discovery and shopping pre #holidayseason. ⚡ The Two Critical Changes in AI Mode 1. Testing AI-Generated Product Summaries (September 22nd) Here's how it works: users click a product in AI Mode, a right-hand panel opens, and Google dynamically generates an AI summary in real-time. That summary is the first impression—not your beautifully designed product page. • If your source content is messy or vague? You get poor summaries. Worse yet, the AI might just feature your competitor's product instead. ⚠️ • The key: your content needs to be clean, clearly structured, with well-written features, pros/cons, specs, and user reviews. 2. Introducing Visual Search Fan-Out in AI Mode (September 30th) now analyzes every image detail—objects, background, textures, metadata—runs multiple parallel queries behind the scenes, then serves up visual grids users can refine conversationally. With 50 billion Shopping Graph listings (2 billion refreshed hourly), poor image metadata means you simply don't exist.👻 The interesting part? • Agentic experiences are live in #SearchLabs for some areas now, helping users buy and find exactly what they want. • But #Google still wants final purchases on your site, making your product page the make-or-break conversion point. 👀The E-commerce Triage Pattern I'm Watching: BrightEdge research found that in September #AIOverviews grew 43.6% in E-commerce. I don't think this was random—it looks like a clear product hierarchy being deployed ahead of the holidays. The pattern: Google's using #AIO for quick product info while keeping traditional results for research and comparison shopping. Layer in this new visual #AIMode experience, and shoppers now have options tailored to exactly what they need based on their query type. 🎯 Why This Matters Right Now: #Black Friday is closing in. Holiday shoppers are already searching conversationally: "Show me cozy gifts under $50" or uploading photos asking, "find me this in blue." • If your product pages aren't AI-friendly and your images lack metadata, you're not showing up in these AI-curated results during the biggest shopping season of the year. • That 43.6% September surge? Feels like Google stress-testing their entire AI Search ecosystem and fine-tuning for holidays, specially around shopping and Agentic AI. My Take: Your product pages are now AI training data. Your images are search assets. The optimization window for holiday is closing fast.⏰ E-commerce brands, how are you preparing your product catalogue for #Google AI holiday discovery and purchase?🛍️ https://lnkd.in/gH6Ehuvk #Ecommerce #Commerce #AI #SEO #BrightEdge #Shopping #Retail #AgenticAI #VisualSEO

I’ve seen 8 reasons e-commerce teams prioritize an AI shopping assistant. All 8 address friction tied to revenue. The same patterns show up across brands and categories. I’m sharing the eight reasons as a practical reference for e-commerce leaders shaping their AI roadmap. The Top 3 reasons teams start 1. PDP abandons kill conversion 90%+ of shoppers leave product pages without adding to cart. Because concerns are unaddressed. Sizing. Fit. Specs. Compatibility. An AI shopping assistant anticipates and resolves these concerns 1:1 Result: 2× conversion for assisted shoppers. This is the most common starting point because ROI is immediate, while data to match product and intent data builds in the background. 2. Traffic keeps getting more expensive Teams report 10–30% YoY increases in acquisition costs, with flat conversion. An AI shopping assistant focuses on converting high-intent traffic already on-site. Result: 10-25% revenue contribution means AI is making your store more productive. 3. Pre-revenue questions overwhelm support Transactional questions drive up support volume. Shipping timelines. Delivery status. Return policies. “Where is my order?” alone can represent 25–35% of support interactions. An AI shopping assistant answers these questions instantly and proactively. Result: 4 to 10x more engagement with 50% less support workload The Next 5 reasons that expand the business case 4. Shoppers bounce in discovery When discovery feels hard, shoppers leave. Search and filters generates too many choices. AI guides selection with recommendations all along the way. Result: 60%+ click-through on AI-recommended products. 5. Returns are eroding margins NRF estimates 20% of online purchases are returned. AI improves decisions before checkout. Result: Eliminate returns that are due to misset expectations. 6. Cross-sells don’t lift AOV Static recommendations convert at 1–2%, even on high-traffic pages. AI suggests more relevant add-ons that build trust and AOV. Result: Higher AOV without discounts. 7. Slow, generic interactions hurt loyalty Slow or generic answers break trust. AI delivers fast, contextual, on-brand responses. Result: CSAT in the 85–90% range and stronger repeat behavior. 8. Teams need scale through insights, not headcount Revenue goals grow faster than teams can. AI scales an organization with insights on the shopping journey and what their customers really want. Result: Grow expertise, not headcount. Takeway Teams don’t prioritize AI shopping assistants to automate. They prioritize removing shopping friction. Start with one of the Top 3, and build from there! -- Like this? Save, and repost. Follow Mike de la Cruz for more.

Visitors who visit 3+ products are 2x more likely to buy. Yet most home pages fail at helping visitors find relevant products. 'Relevant' being the key word. In this post, taking this site's example, I'll be sharing 10 ways you can enable relevant product discovery: 1. Have an open search bar. Relevant if you have 100+ products and highlight product names on social posts. 2. Navigation with new and popular. Helps visitors find best selling products. Also include 'Sale' when live (that page usually sees a high conversion rate). 3. Mention the CTA in the banner. This clearly tells shoppers what to expect when they click on the banner. 4. Highlight top selling categories. This should ideally be below the banner. You can highlight best selling 'sub' categories here too. 5. When a sale is live, give it visual prominence. Mention the sale % and till when is the sale live. 6. Help shoppers explore broad sale categories (e.g. girls instead of girl tops). Show these categories under the sale banner. 7. Use lifestyle imagery on the homepage. This creates aspiration. Do this in the lifestyle featured collection section. 8. Highlight in season products. These are your best selling items during last 1-4 weeks. You can also show all 'bestsellers' and have them divided by categories. 9. Show curated themes. You want to give visitors different ways to shop so they end of finding something of their interest. 10. Have a bestsellers banner / product slider. If a banner, use lifestyle images. If a product slider, show tabs of categories with a CTA to go their category page. When visitors find relevant products, something they really like.. Your site conversions dramatically increase. Use this as a checklist for your homepage. Found this helpful? Let me know in the comments. 

🌌 Amazon’s COSMO: Next-Gen Search! 🛒🚀 Amazon just launched COSMO, a game-changing AI-powered search system designed to revolutionize how we discover products. Unlike traditional algorithms that rely on keywords, COSMO focuses on customer intent and context, creating a shopping experience tailored to user needs. Here's what I takeway: 🎯 What is COSMO? COSMO uses commonsense knowledge graphs to understand why, how, and for whom a product is used. Think smarter searches—when someone looks for “shoes for pregnant women,” COSMO suggests slip-resistant options instead of generic results. Another example? A search for “gear for winter hiking” might prioritize insulated boots, thermal jackets, and crampons, understanding the user’s need for products suited to cold and icy conditions instead of generic outdoor gear. 🥾❄️ 🤖 How it Works: COSMO leverages large language models to process search intent and behavior. Integrates visual and textual data for a comprehensive product understanding. Already impacting 10% of Amazon searches with reported 8% engagement rate increases. 💡 What This Means for Brands: 1️⃣ Content Overhaul: Move beyond keywords. Focus on how your product is used, who uses it, and why it matters. Context is the new king! 👑 2️⃣ Visuals Matter: COSMO reads images! Optimize photos to show products in action and highlight benefits. 📸 3️⃣ Reviews are Gold: Reviews help train COSMO. Encourage customer feedback to boost your listing's relevance. ⭐ 4️⃣ No More Shortcuts: Keyword stuffing won’t cut it. Brands that embrace context-driven optimization will thrive. 📈 Why This Matters Now: The days of static product searches are ending. COSMO represents a shift towards personalized, AI-driven shopping experiences. Brands that prepare today will dominate tomorrow. 💬 Your Take: How do you see COSMO changing the e-commerce landscape? Are your listings ready for this transformation? Let’s discuss! 👇