Beyond 'Customers Also Bought': How AI Personalisation Is Reshaping Retail and E-commerce

Online catalogues are vast, and attention is short. A shopper on a mobile phone sees a handful of products on the first screen, and whatever appears there disproportionately shapes what they buy. That makes ranking, the decision of what to show and in what order, one of the most commercially consequential pieces of software in retail.

In markets like Nigeria, the stakes have a local twist. Many shoppers are mobile-first, data-conscious and price-sensitive, and they often discover products through social media and messaging apps before ever reaching a store's website. Personalisation therefore has to work with sparse histories, cheap Android devices and intermittent connectivity, and it has to earn trust in an environment where scams make people wary.

Collaborative filtering and embeddings

The classic approach learns from patterns of behaviour: people who interacted with similar items probably share tastes. Matrix factorisation turned this into learned vectors (embeddings) for users and items, and modern systems extend it with deep networks. A common production pattern is the two-stage architecture described in Google's paper on Deep Neural Networks for YouTube Recommendations: a fast candidate-generation model narrows millions of items to hundreds, and a richer ranking model orders them using many features.

Hybrid and sequence-aware models

Pure collaborative filtering struggles with new products and new users, the cold-start problem. Hybrid models such as Wide and Deep Learning combine memorised feature crosses with generalising deep embeddings and can incorporate product attributes, text and images. Sequence models treat a session as a sentence of interactions; transformer-based recommenders like BERT4Rec predict the next item from the order of recent clicks, which captures intent far better than a static profile.

Search as personalisation

For many retailers, search drives a large share of revenue. Learning-to-rank models re-order search results using query features, product quality signals and user context. Semantic search with text embeddings handles synonyms, misspellings and local vocabulary, which is critical where shoppers mix English, Pidgin and Yoruba or Igbo terms in the same query.

Generative and conversational commerce

Large language models are now being used to write product descriptions, summarise reviews and power shopping assistants that answer questions in natural language. The best of these are retrieval-augmented: the LLM grounds its answers in the actual catalogue, stock levels and policies rather than inventing details. Agentic assistants that compare options and fill carts are an emerging frontier, and an area where reliability and consumer protection matter a great deal.

Bandits and experimentation

Because recommendations change behaviour, offline accuracy is a weak predictor of online impact. Contextual bandits balance exploiting known good items with exploring new ones, and disciplined A/B testing remains the only reliable way to know whether a model actually helps.

Large streaming and e-commerce platforms have openly published much of the research above, which is why the field moves quickly. Marketplaces use personalisation not only for product recommendations but also for email and push timing, homepage layout and promotional targeting. Grocery and fashion retailers increasingly use demand forecasting alongside personalisation, so that what is recommended is also what is in stock nearby.

In African e-commerce, a notable pattern is social and conversational commerce: small merchants selling through WhatsApp and Instagram, with AI tools helping them draft responses, manage catalogues and identify repeat customers. Here personalisation is less about massive models and more about giving small businesses capabilities that only large retailers used to have.

• Privacy and consent. Personalisation runs on behavioural data. Laws such as Nigeria's data protection framework, overseen by the Nigeria Data Protection Commission, require lawful bases, transparency and data minimisation. Collect what you need, explain why, and honour opt-outs.
• Manipulation and dark patterns. Optimising purely for clicks or conversion can reward urgency tricks and misleading scarcity cues. The OECD's AI policy work highlights consumer protection as a core concern with algorithmic systems.
• Personalised pricing. Charging different people different prices based on inferred willingness to pay is legally and reputationally risky, and can discriminate by proxy.
• Filter bubbles and popularity bias. Recommenders tend to amplify what is already popular, squeezing out small sellers and new products. Marketplaces should measure exposure fairness across sellers.
• LLM hallucination. A shopping assistant that invents a warranty or a delivery date creates real liability. Ground every factual claim in authoritative data.

1. Get the event data right. Clean, consistent tracking of views, add-to-carts, purchases and returns, with stable product IDs, is the foundation. Many teams skip this and spend months debugging models instead.
2. Start simple. Popularity by category and location, plus "frequently bought together" from co-purchase counts, is a strong baseline. Move to embeddings and sequence models when you can show the uplift.
3. Solve cold start with content. Use product text and images to embed new items, and use context such as location, device and referral source for new users.
4. Optimise for long-term value. Measure repeat purchases, return rates and customer satisfaction, not just click-through. A recommender that drives returns is destroying value.
5. Design for low bandwidth. Keep payloads light, precompute recommendations where possible and make sure pages remain useful when personalisation services are slow.
6. Give users control. Let people see why something was recommended and dismiss categories they do not want. Transparency is a trust feature.

When I mentor developers building their first e-commerce products, I encourage them to treat personalisation as a product capability with ethical boundaries, not just a model to tune. The question to ask is whether the shopper would thank you for the recommendation if they knew how it was made.

That test also turns out to be good business. Retailers that personalise in ways customers find genuinely helpful, such as remembering sizes, surfacing items back in stock or recommending complementary products, build loyalty that discount-driven targeting rarely achieves. In crowded markets, trust compounds.

• Modern retail personalisation spans recommendations, search ranking, messaging and conversational assistants.
• Two-stage candidate-and-ranking architectures, hybrid models and sequence-aware transformers are the core techniques.
• Retrieval-grounded LLMs are reshaping product content and shopping assistance but must not invent facts.
• Privacy, dark patterns, personalised pricing and seller fairness are the major ethical risks.
• Strong event data, simple baselines and long-term metrics beat premature model complexity.

1. Deep Neural Networks for YouTube Recommendations — ACM RecSys 2016 (Covington, Adams and Sargin)
2. Wide & Deep Learning for Recommender Systems — arXiv (Cheng et al.)
3. BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer — arXiv (Sun et al.)
4. OECD.AI Policy Observatory — OECD
5. Nigeria Data Protection Commission — NDPC

Cover image: original illustration created for this article. Cited works remain the property of their respective authors and publishers.