# How AI is improving product recommendation engines AI is transforming product recommendation engines. Learn how ecommerce businesses are improving conversion, AOV and catalogue discovery with smarter AI systems. Published: 2025-12-29 Author: Rodan Analytics Most recommendation engines are not doing what their owners think they are doing. They surface popular products, reward repeat purchases and bury anything that does not already have velocity. The result is a system that amplifies what already works and ignores everything else — including the products most likely to delight a new customer or grow a marginal one. The mistake most ecommerce businesses make at this stage is treating their recommendation engine as infrastructure rather than strategy. It gets implemented, handed to a platform team and left to run. Conversion rates look acceptable. No one investigates what the engine is actually optimising for, or what it is costing them in suppressed discovery, poor personalisation and abandoned journeys. AI is changing what recommendation engines can do — not by making them faster, but by making them genuinely predictive rather than merely reactive. This article explains how that shift works in practice, where the real commercial gains sit and what separates businesses that extract value from it from those that do not. ## Why most recommendation engines underperform The dominant model in ecommerce recommendation has been collaborative filtering: recommend what similar users bought. It works well enough at scale to look credible in an A/B test. The problem is that it is retrospective. It tells you what people like you have already done. It cannot tell you what you would have wanted if you had been shown something different. For a mid-market fashion retailer with 600 SKUs and a seasonally rotating catalogue, collaborative filtering means new arrivals take weeks to accumulate the signal needed to appear in recommendations. By then, the peak selling window has passed. The engine is always running two steps behind the actual opportunity. The second failure mode is context blindness. A customer browsing at 11pm on a Sunday on a mobile device is not in the same purchase mode as the same customer clicking through a promotional email on a Tuesday morning. Traditional engines do not model context — they model identity. That distinction matters enormously for conversion. The third is cold start. New customers have no history. Most engines either default to popularity lists or ask customers to self-identify preferences through onboarding flows that most people abandon. Neither is a solution. ## What AI-powered recommendation actually looks like The meaningful upgrade is not a better algorithm in isolation — it is the combination of richer signal, better modelling and faster iteration. Modern AI-powered recommendation systems draw on a broader input set: browsing behaviour (dwell time, scroll depth, hover events), basket composition, return history, search queries, price sensitivity signals and real-time session context. They model these inputs together rather than treating purchase history as the primary variable. A useful example: a health and wellness ecommerce brand with a catalogue of approximately 400 products introduced session-context modelling to its recommendation layer. Rather than recommending based on past purchases alone, the engine weighted current session signals — what the customer had searched for, which category pages they had visited in that session, how long they had spent on specific product pages. Conversion on recommended products increased because the engine stopped interrupting a consideration journey with prompts to rebuy products the customer had already bought. The second meaningful shift is the move toward large language models for semantic product understanding. Rather than relying on category tags and metadata (which are only as good as whoever wrote them), LLM-enabled systems can understand product descriptions, customer reviews and search queries in natural language. A customer searching for "something for a weekend away that doesn't crease" does not match a keyword. But a semantically capable recommendation engine can match it to the right product cluster without a human ever building that rule. Quantsole, Rodan's LLM-enabled business intelligence product, reflects this kind of thinking — bringing natural language understanding into commercial data workflows so that insight generation does not depend on whoever configured the system three years ago knowing what questions would matter today. ## Where the commercial gains actually sit The case for improving recommendation engines usually gets made on click-through rate or conversion uplift. Those are real, but they are not the most interesting numbers. The more significant gains are in average order value, second-purchase rate and catalogue utilisation. For a business with a long catalogue tail, the ability to surface relevant products that customers would not have discovered otherwise is a revenue line that most current systems leave entirely on the table. Consider a direct-to-consumer furniture business with 2,000 SKUs. Their top 200 products generate 80% of recommendation-driven revenue. The remaining 1,800 products exist on the site but are effectively invisible to anyone who does not search for them directly. An AI-powered recommendation system with semantic product understanding and contextual signals could realistically shift 10-15% of that tail into active recommendation rotation — not by guessing, but by matching product attributes to demonstrated customer preferences in real time. Second-purchase rate is equally important. Most recommendation logic optimises for the immediate transaction. AI systems can model purchase propensity across time — identifying customers who are three weeks post-purchase and statistically likely to be receptive to a specific complement or replenishment prompt. That is a retention play dressed as a recommendation play. ## What separates businesses that get this right The gap is rarely technical. It is organisational. Businesses that extract genuine value from AI-powered recommendations treat the recommendation engine as a commercial asset that requires active management. Someone owns it. Someone monitors what it is recommending and why. Someone is asking whether the optimisation objective — usually click-through or conversion — is aligned with the actual commercial objective, which might be margin per order, lifetime value or new-to-brand category penetration. The businesses that struggle have a different pattern: the recommendation engine was implemented during a platform migration, the vendor promised it was intelligent out of the box and no one has looked at the configuration since. The engine is technically live and operationally ignored. Three questions every growth leader should be able to answer about their current engine: - What is the engine optimising for, explicitly — and is that the right objective? - How does it handle new arrivals, and how long before a new product enters active recommendation rotation? - What percentage of your active catalogue received at least one recommendation impression in the last 30 days? If the answers are vague, the engine is not working as hard as it should. Building the capability also requires honest assessment of data readiness. AI-powered recommendations need clean, consistent event tracking, a product data layer that is rich enough to enable semantic matching and an infrastructure that can serve recommendations in real time without latency killing the experience. Most businesses have gaps in at least one of these areas. Identifying where the gaps sit before commissioning a new system is the work that determines whether the investment pays off. An AI readiness assessment — the kind of diagnostic work Rodan undertakes as a structured engagement — typically surfaces two or three foundational data issues that would otherwise undermine a more sophisticated recommendation layer. That work is worth doing first. ## The cost of leaving this alone Every month a recommendation engine runs on outdated logic is a month of suppressed discovery, missed second purchases and revenue that went to a competitor whose system was more attentive than yours. The businesses that will separate themselves in ecommerce over the next three years are not the ones with the largest catalogues or the biggest marketing budgets. They are the ones that know, at the individual session level, what a customer needs next — and put it in front of them at the right moment. That capability is not reserved for the largest players. It is accessible to mid-market and growth-stage ecommerce businesses today. But it requires treating the recommendation engine as a strategic system, not a platform feature. If you want to understand where your current recommendation infrastructure has gaps and what a more capable system would require, Rodan's diagnostic engagement is the right starting point. It is a structured, fixed-scope piece of work that gives you a clear picture of where you are and what to do next. Book a diagnostic at rodan.io. --- **Meta description:** AI is transforming product recommendation engines. Learn how ecommerce businesses are improving conversion, AOV and catalogue discovery with smarter AI systems. HTML: https://rodan.io/insights/how-ai-is-improving-product-recommendation-engines