AI-Powered Recommendations

Revolutionary Breakthrough in AI Recommendations: RecGPT Foundation Model The recommendation systems landscape just experienced a seismic shift with the introduction of RecGPT, a groundbreaking foundation model that solves one of the industry's most persistent challenges: zero-shot generalization across domains without extensive retraining. >> The Core Innovation Traditional recommender systems hit a fundamental wall when encountering new contexts or cold-start scenarios, invariably demanding resource-intensive retraining cycles. RecGPT fundamentally departs from existing ID-based approaches by deriving item representations exclusively from textual features, enabling immediate embedding of any new item without model retraining. >> Technical Architecture Deep Dive Unified Item Tokenization: The system employs Finite Scalar Quantization (FSQ) to transform heterogeneous textual descriptions into standardized discrete tokens. This creates a domain-invariant representation that eliminates the semantic heterogeneity barriers plaguing existing systems. Each item embedding is partitioned into multiple sub-vectors, then quantized using a sophisticated pipeline involving sigmoid normalization and rounding functions with Straight-Through Estimator techniques for gradient flow. Hybrid Attention Mechanism: RecGPT implements a revolutionary dual attention system combining bidirectional attention for intra-item token processing with causal attention for sequential modeling. This enables comprehensive information sharing among tokens representing the same item while preserving temporal causality necessary for next-item prediction. Auxiliary Semantic Pathways: To counteract information loss during quantization, the framework incorporates original semantic representations alongside learned discrete token embeddings. These parallel embedding streams are combined through Layer Normalization to produce final input representations for transformer layers. Efficient Catalog-Aware Decoding: The system employs Trie-based prefix matching with beam search optimization, exploiting the sparse nature of valid item combinations within the vast token space. This transforms computationally intractable token-to-item mapping into practical real-time operations. Key Advantages: - Zero-shot transferability for immediate new item embedding - Cross-domain compatibility bridging diverse recommendation contexts   - Enhanced robustness in sparse-data environments - Real-time inference capabilities for production deployment The foundation model approach to recommendation systems has arrived, promising to transform how we think about personalization across platforms and domains. This breakthrough opens new possibilities for recommendation systems that truly understand content semantics rather than relying on brittle ID-based representations.

Surfer analyzed 289,105 URLs across ChatGPT, AI Overviews, Perplexity, and AI Mode to answer one question: Does AI recommend brands more strongly when they appear in more cited sources? Short answer: Yes. The correlation is moderately strong (0.41 Spearman). But here's where it gets interesting. ChatGPT shows the strongest correlation between brand presence and recommendation strength compared to all other AI platforms. Yet ChatGPT cites 84.7% more sources than AI Mode and 129.7% more than AI Overviews. That means getting mentioned in a high percentage of ChatGPT's cited sources is harder but more important. The second finding: Blog posts matter most. Brand mentions in blog posts correlate more strongly with recommendation strength than mentions in reviews, news articles, or any other content type. Even though blogs only make up 28.7% of cited sources, they drive the strongest recommendation impact. Third-party blogs (not your own) have the biggest effect. So what should you do? Stop trying to get mentioned everywhere. Start focusing on the blogs that AI cites repeatedly. Use tools (like Surfer's AI Tracker) to find which third-party blogs get cited most often for your target queries. Then figure out how to get mentioned in those specific posts. Reach out to authors. Pitch data they're missing. Send product samples for review. Offer expert quotes. According to Surfer's research, this is the game now. Getting your brand mentioned on frequently cited blog posts gives you the best shot at AI recommendations.

How Walmart Uses AI to Recommend Kitchen Tools When You Buy Milk Ever wonder why Amazon suggests a milk frother when you add milk to your cart? The challenge is harder than it looks—how do you bridge the gap between routine grocery purchases and discovering useful general merchandise? 👉 The Challenge Traditional recommendation systems excel at suggesting similar items (milk → cheese) but struggle with cross-category discovery (milk → milk frother). The problem becomes even trickier when customers have 20+ items in their grocery cart—which products should drive recommendations, and how do you rank them intelligently? 👉 Walmart's Dual Approach Researchers at Walmart Global Tech developed a two-stage system that combines the best of both worlds: Stage 1: Smart Candidate Generation - Historical co-purchase analysis identifies proven item pairs - Large Language Models (LLMs) generate contextual suggestions that go beyond purchase history - Example: For eggs, the LLM suggests egg poachers, timers, and specialized pans—items that enhance the egg experience but might not appear in traditional data Stage 2: Real-Time Cart Context Ranking - A transformer-based neural network analyzes the entire cart contents - Uses cross-attention to understand how potential recommendations relate to all cart items simultaneously - Considers customer persona, platform, and sequential shopping behavior 👉 The Results The system delivered impressive improvements: - 36% increase in add-to-cart rates for LLM-generated recommendations - 27% lift in ranking quality (NDCG@4) when using full cart context - 4.7x more unique recommendation coverage compared to traditional methods 👉 Why This Matters This research tackles a fundamental e-commerce challenge: helping customers discover products they didn't know they needed. By combining AI reasoning with behavioral data, the system creates those "aha moments" where a grocery run becomes an opportunity for useful discovery. The approach shows how modern AI can enhance rather than replace traditional recommendation methods, creating more engaging shopping experiences while driving meaningful business results. Paper by Akshay Kekuda, Murali Mohana Krishna Dandu, Rimita Lahiri, Shiqin Cai, Sinduja Subramaniam, Evren Korpeoglu, and Kannan Achan from Walmart Global Tech.

Graph-Based Memory and Recommendation Systems in AI Agents ⭕️ Memory systems in AI agents achieve new levels of sophistication when integrated with recommendation mechanisms, particularly within graph-based architectures. This integration creates a dynamic system that not only stores and retrieves information but actively suggests relevant knowledge and actions based on complex patterns of relationships and interactions. Graph-based memory systems naturally evolve into recommendation engines through their ability to traverse and weight relationships between entities. When an agent encounters a new situation, it can leverage its graph structure to recommend responses, solutions, or information based on similar past experiences. This capability transforms passive memory storage into an active system that anticipates needs and suggests relevant actions. Consider how this works in practice: When a user interacts with an agent, each interaction creates nodes and edges in the memory graph. These connections capture not just the content of the interaction, but also its context, outcome, and related factors. The agent can then use this rich network of relationships to make informed recommendations in future situations. For instance, if a user asks a technical question, the agent can traverse its graph to find similar questions, successful solutions, and related concepts that proved helpful in analogous situations. The power of graph-based memory for recommendations lies in its ability to identify patterns across multiple dimensions simultaneously. Unlike traditional recommendation systems that might focus on simple similarity metrics, graph-based systems can consider: Temporal patterns: How solutions evolve over time Contextual relationships: How different pieces of information relate in various contexts User behavior patterns: How different users interact with similar information Outcome patterns: Which approaches tend to lead to successful results When these patterns combine in the graph structure, they create a sophisticated basis for generating recommendations. The graph structure makes it possible to discover these indirect relationships and use them to enhance recommendations. Graph-based memory systems excel at incorporating feedback to improve future recommendations. When users respond positively or negatively to suggestions, this feedback modifies the weights of relevant connections in the graph. Over time, this creates a self-improving system where successful recommendation patterns strengthen while less helpful patterns naturally weaken. The adaptation process works across multiple scales simultaneously. At the individual level, the system learns user-specific preferences and patterns. At the aggregate level, it identifies broader patterns that apply across user groups or problem domains.

Amazon's AI just became your biggest competitor …and your most important customer. Amazon rolled out a feature last week that should have every seller and advertiser paying attention. "Help Me Decide" uses AI to analyze a shopper's recently viewed items and hand-picks 3 products for them: a "best pick," a "budget option," and an "upgrade pick." It provides AI-generated reasons why each product was selected, pulling from reviews, features, and the customer's shopping history. As an Amazon PPC agency working with sellers across many categories, we believe this feature introduces both opportunity and a new priority. ✅ Opportunity: Shoppers are being guided faster to a decision. If your listing is chosen by the AI, that could mean fewer distractions, fewer competing clicks, and essentially a more direct path to purchase. 📌 Priority: The AI picks products based on browsing behavior + listing signals (features, reviews, alignment with shopper’s prior purchases). That means you need to ensure your listing and your PPC strategy tick all the boxes: compelling unique selling proposition, strong review profile, clear feature/benefit language, and pricing that works in the context of “budget vs upgrade”. From a PPC lens, here are three immediate action steps we recommend: ➣ Review your listings and ensure the key differentiators are front-and-centre (because the AI will use them). ➣ Adjust your ad creative and targeting to capture interest before the indecision threshold (i.e., fewer “I’m just browsing” shoppers). ➣ Monitor your metrics and attribution: as this feature rolls out, watch for shifts in how often your ads lead to a listing view vs. how often a listing gets surfaced by AI-driven recommendation. In short: if Amazon is automating more of the decision process, the quality of your listing and how well your PPC strategy supports that listing become more critical than ever. If you’re a seller wondering how to make sure you’re not passed over by “Help Me Decide”, send me a DM.

Sharing key learnings and insights from our Real-Time (In-Session) Personalization journey at CARS24 — a capability that has transformed how we personalize the car buying experience at scale. Leveraging advanced sequence-based neural networks and real-time Kafka streaming infrastructure, we've developed a dynamic machine learning pipeline that processes more than a million user interactions daily. Our deep learning models rapidly adapt to user behaviour, delivering personalized car recommendations with sub-200ms latency. Highlights: ✅ Advanced sequence-based neural network architecture  ✅ Real-time streaming and processing of user behaviour signals with Kafka  ✅ Rapid feature engineering and inference using optimized real-time databases  ✅ High scalability for continuous model retraining and deployment Performance Impact: 📈 Across all discovery widget we achieved a highest Impression-to-View (I2V) rate and on the 'Best Matches' recommendation rail on our car detail page and buyer home page. 📈 Delivered a strong Impression-to-Booking Initiation (I2BI) conversion rate across different discovery widgets, underscoring high user relevance and engagement. Business Outcomes: 🚀 Significant uplift in user engagement  🚀 Marked reduction in user drop-offs  🚀 Enhanced personalization and superior user experience The attached flow chart outlines the architecture behind this AI-powered personalization pipeline — from real-time clickstream ingestion to ML inference and personalized recommendations. #RealTimePersonalization #AI #MachineLearning #DeepLearning #Kafka #DataScience #RecommendationEngine #TechInnovation #AI  #Personalization #pubsub #CARS24 #transformers #llm #genai

I've been asked the same question 6 times this week: "How should we add AI tools like ChatGPT to our HDYHAU attribution surveys?" So this morning, we analyzed 80,000+ free-form HDYHAU responses to find out exactly how consumers describe discovering brands through AI. The biggest insight? consumers don't see AI as just another search tool—they see it as a 𝗿𝗲𝗰𝗼𝗺𝗺𝗲𝗻𝗱𝗲𝗿. They're writing in: • "ChatGPT recommended" • "AI suggested" • "Asked Perplexity for recommendations" AI isn't replacing Google—it's creating an entirely new discovery channel that acts more like a trusted friend than a search engine. And this trusted friend is scaling fast, LLM mentions are up 5x since January across the entire Fairing data set. 𝗢𝘂𝗿 𝘄𝗼𝗿𝗱𝗶𝗻𝗴 𝗿𝗲𝗰𝗼𝗺𝗺𝗲𝗻𝗱𝗮𝘁𝗶𝗼𝗻: "AI Recommendation (e.g. ChatGPT)". ChatGPT is dominating in our sample, so we'll use that as the example in the response. (Add this option when AI hits >1% of your responses) Full analysis by our data scientist Enia Xhakaj, PhD in the comments 👇

Improving recommendations isn’t just about better models — it’s about creating more meaningful user experiences. In our latest work at CVS Health 🏥, Madhumita Jadhav and I enhanced the "You May Also Like" (YMAL) system by combining the contextual power of LLMs with the efficiency of Word2Vec. The result: smarter, faster, and more relevant suggestions. 🔗 Read the full details on the CVS Health Tech Blog - https://lnkd.in/djiwVvbx Inside the post: ✨ A look at our hybrid architecture ⚙️ Practical lessons from real-world deployment 💡 Why LLMs and traditional embeddings work better together Grateful to collaborate with such a sharp and supportive team 🙌 Madhumita Jadhav, Sarah Boukhris-Escandon, PhD, PE, Jyothsna Santosh, Sowmya Vasuki Jallepalli, Ryan Berns and more! Always curious to hear how others are applying LLMs in production — feel free to share your thoughts! #AI #LLM #RecommenderSystems #MachineLearning #CVSHealthTech #TechInnovation #NLP

74% of "how-to" queries now show AI Overviews instead of organic results. But that's not even the most shocking shift in search. A groundbreaking study by First Page Sage reveals something even more disruptive: ChatGPT, Gemini, Claude, and Perplexity are becoming the new gatekeepers to your customers. Welcome to the era of GEO (Generative Engine Optimization). Here's what's happening: When someone asks an AI chatbot "What's the best CRM software?" or "Who are the top marketing agencies?" - they're getting recommendations not from Google, but from AI. And the algorithm is completely different. FirstPageSage analyzed 11,128 commercial queries across the top 4 AI chatbots. The results? Mind-blowing. Here's what ACTUALLY determines AI recommendations: For ChatGPT (61.3% market share): • Authoritative list mentions: 41% • Awards & accreditations: 18% • Online reviews: 16% • Customer examples: 14% • Social sentiment: 11% For Google Gemini (13.3% market share): • Authoritative list mentions: 49% • Google website authority: 23% • Awards & accreditations: 15% • Online reviews: 13% For local businesses, it's even more fascinating: • Local business reviews: 38% • Authoritative list mentions: 29% • Online reviews: 19% This changes EVERYTHING about digital marketing. While you've been obsessing over keyword density and backlinks, your competitors might be securing spots on "top 10" lists that rank highly on Google. Why? Because when ChatGPT gets asked about your industry, it's scanning those lists to make recommendations. What this means for your business: 1. Create and rank "best of" lists that include your company (This influences 41-64% of AI recommendations!) 2. Showcase awards, certifications, and affiliations prominently (Influences 15-19% of recommendations) 3. Generate positive reviews on trusted platforms (Influences 13-31% of recommendations) 4. Publish customer case studies and usage data (Influences 13-14% of recommendations) 5. Monitor and improve social sentiment (Influences 11% of ChatGPT recommendations) The companies that master GEO now will dominate the next generation of search. Those who don't might disappear from the customer journey entirely. Our AI Agent can help you analyze your GEO potential in minutes: • Scan for list inclusion opportunities • Identify award and certification gaps • Analyze review sentiment across platforms • Generate a complete GEO strategy The future of search isn't coming. It's already here. Are you ready? Source: FirstPageSage's Generative Engine Optimization (GEO) Study (link in comments)

💡 "Find me a gift for my sister who loves cooking, likes sustainable brands, and has a small kitchen." Traditional retail search: *Returns 500 random kitchen products* AI-powered search: *Curates space-saving, eco-friendly tools with reviews and next-day delivery* The difference? Customer loyalty. 🎯 in our latest blog on Databricks.com, John Gilman and I explore how AI is transforming retail from search to sale—and why this isn't just about better recommendations. It's about creating experiences so personalized that customers feel *understood*. The data is compelling: • 15-25% boost in conversion rates • 10-15% increase in average order volumes • 25% reduction in service backlogs during peak periods But here's what caught our attention: The retailers winning aren't just using AI for product recommendations. They're using unified customer data to power retail media networks, turning customer insights into high-margin revenue streams while maintaining privacy compliance. So ask yourself, if your competitor can instantly understand "sustainable cooking gifts for small kitchens" while you're still matching keywords, how long before your customers never come back? The holiday season is coming. The retailers who've built real-time personalization capabilities will capture loyalty. The rest will compete on price. Which camp are you in? Read the full post: https://lnkd.in/gG_SksWf #RetailAI #CustomerExperience #DataStrategy #RetailTech #Personalization