For most e-shops, the best product recommendations still come from classic methods, not from a language model ranking products. "Customers who bought this also bought" computed from your own orders, plus a ranking model trained on views and sales, is cheap, answers in milliseconds, and in 2026 evaluations still beat language models asked to rank products directly on most test domains. AI earns its place elsewhere: cleaning and enriching the product catalogue that the ranking reads, surfacing new products that have no sales history yet, and understanding vague requests such as "a warm gift for a two-year-old under 40 euros". In an online test at the Lazada marketplace, better product attributes alone raised the value of goods sold through recommendations by 7.89%, with no AI model running when the page loaded.
Do e-shops need AI for product recommendations?
E-shops need recommendations, but they rarely need a language model to produce them. Two classic methods do most of the work. Item-to-item collaborative filtering counts which products are bought or viewed together and fills the "also bought" slot. A gradient-boosted tree model orders a list using signals such as views, sales, price and stock. Both answer in well under a millisecond. A 2026 benchmark across five domains found that language models used as rerankers "fail to consistently outperform strong collaborative and content baselines" (arXiv 2606.29947), and a second 2026 study found that no open-weight model beat a tuned classic baseline under a fair protocol (arXiv 2609.00086). The reason is simple: a recommender's strongest signal is what similar customers did, and a language model has no access to it.
Where does AI actually improve product recommendations?
AI improves recommendations in four places, and the first pays most reliably:
- The catalogue, offline. A model extracts and normalises attributes, fills in missing categories and writes absent descriptions, and the classic ranker, the search and the filters all improve because their input improved. At Lazada, attributes extracted this way raised gross merchandise value by 3.81% in badges, 5.32% in search and 7.89% in recommendations in online A/B tests (AutoPKG, arXiv 2604.16950).
- New products. A product with no sales has no "also bought" data, but a model can read its description. The measured bottleneck is getting the new product into the candidate list at all, so the fix is a content-based search, built on embeddings, that runs alongside the behavioural one.
- Vague requests. A chat or search box where the model turns "something waterproof for hiking, size 42" into filters that your existing search then runs. The model outputs filters, never a product list, so it cannot invent products.
- Explanations, such as "because you bought the tent", provided they are generated from the real reason, as explained below.
Why not let a language model rank the products?
Because the page cannot wait, and the gain is smaller than it looks. A product-page carousel has roughly 50 to 150 milliseconds on the server, with no room for a model call. The cost adds up too: 20,000 product views a day with a model call on each carousel comes to roughly 50 to 150 euros a month on a small model, and 10 to 20 times that on a frontier model, for a carousel whose extra revenue nobody has measured. The published gains need careful reading as well. On one benchmark the best proprietary model reranked a fixed list 59% better than the strongest classic method, but changing only the method that produced the candidate list moved the results by more than 50% with the same model, so much of what looks like model quality is the candidate list. The same model scored almost twice as high again when allowed to name any item freely, which a shop that must recommend real, in-stock products cannot allow. And a language model's ranking depends on the order in which products appear in its prompt, which anyone who controls that order can exploit.
What does a small Greek e-shop need first?
A small Greek e-shop needs clean product data and working search before any AI. When I measured 505 Greek e-shops in August 2026, only 5 passed five elementary technical checks (the Greek e-shop technical audit). The order that pays:
- Complete product attributes, such as size, colour, material and compatibility, in fixed values. This is where a model helps most, run once over the catalogue and checked by a person.
- Search that understands Greek. Shoppers type without accents and in Greeklish ("papoutsia"), and Greek words change form ("μαύρο παπούτσι", "μαύρα παπούτσια"). Index a transliterated form next to the Greek one and use a Greek stemmer, as covered in RAG in Greek.
- "Also bought" from your own orders, recomputed nightly. With few orders, start with best sellers per category and similar products by attributes.
- Only then a model at request time, on one surface, with its cost and its delay written down before any code.
Between 2019 and 2022 I built customer segmentation and demand forecasting models for e-commerce clients, and the lesson from that work applies here: a model earns its keep only when it fits the way the team already works and is measured after launch.
How much does AI for recommendations cost?
Enriching the catalogue costs little. A pass over 50,000 products, at about 400 tokens in and 150 out each, costs roughly 8 euros on a small model at September 2026 prices. The limit is review capacity, not money, so a person checks a sample rather than every product. Model calls on every page view are where costs grow, as the numbers above show, and precomputation removes most of them: run the model on a schedule, store its output as numbers per product or per customer segment, and serve pages without calling it. Cache by segment and context, never by individual customer, because a per-customer cache almost never hits. How token costs add up in general is in what an AI agent costs to run.
How do you know recommendations work?
You know from an online A/B test, with revenue per session and conversion rate as the primary numbers. Offline scores decide what to test, never what to ship: one 2026 recommender improved offline recall by 5.0% to 8.8%, and its main metric by 0.792% in a five-day online test (arXiv 2608.21012). Watch the guardrails too: page load time at the slow end, how many distinct products get shown across all customers, and the share of impressions that goes to the long tail. A change that lifts conversion while showing the same twenty products to everyone turns into a stock and supplier problem within a quarter. When you test offline, split the data by time, training on older orders and testing on newer ones, because a random split lets the model see the future.
Are AI explanations like "because you bought" trustworthy?
AI explanations are trustworthy only when they are generated from the real reason. If a ranking model chose the product and a language model is then asked to explain the choice, the explanation is a plausible story written after the fact, fluent and possibly wrong about the cause. It matters legally as well: in the EU, marketplaces must describe their main ranking parameters under the Platform-to-Business Regulation, and online platforms must state the main parameters of their recommender systems under the Digital Services Act. A generated sentence is not that description. The safe pattern is to pass the model the actual signal, such as "often bought with item 4471", and let it do nothing more than phrase it. To see AI working on a shop's own catalogue and orders, the running demos include an agentic checkout.
Frequently asked questions
Can an AI chatbot replace the filters in my e-shop?
No. A shopper who knows what they want finds it faster with filters. A chat earns its place for vague requests, gifts and questions close to customer support, as covered in AI customer support for an e-shop.
How many orders do I need for "also bought" recommendations?
There is no fixed number: the signal becomes useful once the same product pairs appear together repeatedly. Until then, combine the few co-purchases you have with similarity by attributes and best sellers per category.
Do AI shopping agents change this?
They raise the value of clean, structured product data, because an agent that buys on a customer's behalf reads attributes, not photos. That is covered in agentic commerce.
If you run an e-shop and wonder whether AI would improve your recommendations, send me your shop's address and a rough count of products and monthly orders. Whether the next step is the catalogue, the search or the recommendations usually shows from those three.