AI in E-commerce: Personalization, Search, and Inventory

Adobe Analytics recorded 4,700% year-over-year growth in AI-driven visits to US retail sites in 2025, and those visitors convert 42% better than traditional search traffic. Retailers with AI agent integrations saw roughly seven times better sales growth during Cyber Week 2025 than those without.

The catch is what happens on arrival. Analysis of agent-driven traffic found conversion still lagging badly against other referral channels, for a specific reason: merchant infrastructure was never built for machine buyers. That's the thread connecting all three pillars of ai in ecommerce below, because personalization, search, and inventory turn out to run on the same underlying foundation.

Personalization is the most commercially proven of the three. Companies leading on it report revenue increases of up to 40%, AI-driven recommendations typically account for 25% to 35% of revenue where they're well implemented, and McKinsey found shoppers are 4.5 times more likely to purchase after clicking an AI recommendation.

What separates a system delivering those numbers from one that doesn't is rarely the algorithm. Modern recommendation approaches, collaborative filtering, content-based matching, and hybrid models using learned embeddings, are well understood and widely available. The differentiator is the quality of the signal going in: clean behavioural data, accurate product attributes, and enough interaction history to learn from. A sophisticated model on thin or inconsistent product data reliably underperforms a simple model on good data.

Traditional site search matches strings. A shopper searching "warm jacket for hiking in the rain" against a keyword index gets whatever happens to contain those words, which is frequently nothing useful, and a null result page is one of the highest-abandonment moments in e-commerce.

Semantic search replaces string matching with vector embeddings, so the query is matched on meaning rather than exact wording. A waterproof insulated shell surfaces even if its description never uses the words "warm" or "hiking." The underlying retrieval machinery is the same family of technology covered in RAG in generative AI, applied to a product catalogue rather than a document store, and retrieval quality depends heavily on how well products are described rather than on the embedding model alone.

This is also where search stops being purely a UX concern. McKinsey's data shows AI-generated product recommendations converting at 4.4 times the rate of traditional search, which means your catalogue's retrieval quality has become a selection criterion for which products get surfaced at all, not just a nicety affecting how easily humans browse.

Demand forecasting is the least visible of the three pillars and arguably the most consequential, because personalization and search both depend on it silently. Recommending a product that's out of stock converts a good recommendation into a bad experience, and surfacing unavailable inventory in search results wastes the highest-intent traffic you have.

AI forecasting improves on traditional methods by incorporating signals classical models ignore: seasonality interactions, promotional lift, weather, regional variation, and cross-product substitution effects. On the physical side, computer vision for inventory covers how camera-based stock tracking closes the gap between recorded and actual inventory, which is the gap that quietly corrupts forecasting inputs in the first place.

Treated as three projects, personalization, search, and inventory get budgeted separately, staffed separately, and evaluated separately. Treated accurately, they're three applications of one asset: structured, accurate, complete product and behavioural data.

The practical implication for sequencing: catalogue data work is not a prerequisite chore to get past before the interesting AI projects start. It is the project with the highest leverage, because it's the only investment that improves all three outcomes simultaneously.

The reason this foundation suddenly matters more than it did two years ago is that a growing share of your traffic isn't human. IBM's January 2026 research across more than 18,000 consumers in 23 countries found 73% now use AI somewhere in their buying journey. Morgan Stanley projects up to $385 billion of US e-commerce moving to agentic channels by 2030.

AI shopping agents don't browse. They query, compare structured attributes, and select, which means the product data that a human shopper would forgive as incomplete is simply disqualifying to an agent that can't parse it. This is the same capability distinction covered in agents vs chatbots: an agent acts rather than converses, and acting requires machine-readable inputs.

The adoption curve also has a clear gap worth planning around. Research across 4,500 shoppers found 58% use AI to research products, 37% begin a purchase journey through an AI assistant, but only 13% complete a purchase after an AI referral. The discovery layer has moved to AI considerably faster than the transaction layer, which means the near-term priority is being findable and selectable by agents, not building agent checkout.