Recommendation Systems: Personalization, Cross-selling and Homepage Layout — Artificial Intelligence in e-commerce

Gains:

Understanding the logic of the recommendation engine (bought together, similar products, personalization) and its impact on business metrics
Ability to construct product relationships, package and cross-selling scenarios with artificial intelligence support and turn them into testable hypotheses
Ability to see the risks of filter bubble, bias and stock/margin-neutral automatic recommendation and link recommendation rules to commercial control

A significant portion of sales on large e-commerce sites comes from the product recommended to the customer, not the product he is looking for. You've all seen the blocks "Those who bought this also bought this", "Our picks for you", "Cheaper together". The system behind these blocks is called a recommendation engine: it is software that produces answers to the question "What should I show this person now?" by looking at the customer's behavior, the relationship of the products and general patterns. Artificial intelligence (AI) is useful in this field in two ways: it helps you generate scenarios and rules when constructing recommendation logic, and it analyzes your own product relationships in small stores and comes up with package/cross-selling ideas. But beware: a good recommendation system has to look not just at “how many clicks” but also at inventory, margin, and customer experience.

Let's clarify the two terms. Cross-selling is offering another product that complements a product to a buyer (a case for a phone buyer). Up-sell is offering a higher/expensive model that meets the same need (256 GB instead of 128 GB). Both are legitimate ways to increase average basket (AOV); But it works when it is truly beneficial to the customer, and it disturbs when it is forced.

Three basic logics of the recommendation system

Recommender systems generally use one or a mixture of three approaches.

1) Content-based: It looks at the product's own features. "Other running shoes similar to this running shoe". It even works with new clients because it doesn't require behavioral history.

2) Collaborative filtering: Looks at the behavior pattern. "People who bought this product also bought this." It's powerful, but requires a lot of data and suffers from a "cold start" (new product/new customer).

3) Personalization: It tailors the showcase to a single customer based on their past behavior. It has a powerful effect but carries a risk of privacy and "filter bubbles".

The following table summarizes the recommendation blocks and their purpose:

block

logic

Purpose

example

Recommendation setup with AI in a small store

Even small and medium stores that cannot establish a corporate recommendation engine can intelligently structure product relationships using AI. You can anonymously give the real "products sold together" data (order items) to AI and have it analyze "which products are bought together, which packages can I recommend?" AI generates cross-selling and bundling hypotheses for you; You filter them through stock and margin, and then you actually test them. The key word here is hypothesis: the AI's suggestion is not evidence, but an idea to be tested.

Tip: Turn each recommendation idea into an “if-then” hypothesis: “If we recommend Y to someone who buys X, their basket average will increase.” Then measure this with a small experiment (A/B test or period of time). A recommendation without measurement is just a guess.

Commercial audit of recommendation rules

A recommendation system can be dangerous if left to completely automated “most clicked”. Stock: recommending a product that is out of stock will disappoint the customer. Margin: the product with the most clicks may have the lowest margin; The system may direct you to unprofitable products. Brand safety and suitability: irrelevant or inappropriate recommendation in sensitive categories (baby, health) loses trust. Filter bubble: constantly showing the customer the same type of product kills discovery and narrows sales. That's why automatic recommendations are necessarily framed by business rules: "recommend what's in stock", "don't feature the product whose margin is below this threshold", "maintain category compliance".

Bias is also a real risk. Because the recommendation system is based on past behavior, it can reinforce past inequalities; For example, constantly recommending cheaper/lower quality products to a certain group. This is both an ethical and commercial issue. It is necessary to regularly audit the recommendation outputs with a view to "what is being recommended to whom".

Four copyable templates

1) Analysis of products sold together:

Your role: e-commerce analyst assistant. Below are anonymous order items (item codes and number of times taken together). Data: [table]. Task: find the 10 pairs of items most frequently bought together; For each pair, write a cross-selling hypothesis (“Offer Y to buyers X”). Note: These are hypotheses; I will also check the stock and margin.

2) Bundle recommendation:

Product: [main product]. Complementary products: [list, with margin notes].Task: Suggest 3 sensible packages; For each package, explain the benefit to the customer in one sentence. Rule: all products in the package must be in stock; making the low-margin product the main draw.

3) Top selling scenario:

Customer is looking at: [product, price, features].Upscale models: [list, with price/feature differences].Task: draft an honest upsell message: how much to pay for which additional benefit.Rule: no exaggeration; Explain the difference through concrete features.

4) Suggestion rule check:

Check the following recommendation rules.Rules: [list].Task: evaluate each rule against the following risks: stock consistency, margin protection, category fit, filter bubble, bias.List the risky rules and your recommendation for correction.

three mini cases

Case 1 — Cross-selling gain. An electronics store analyzed order data with AI and found a pattern of "40% of those who bought a phone also bought a case within 30 days." Added a suitable case recommendation block to the phone page (filtering out those in stock). Basket average increased measurably. AI gave a hypothesis, the store tested and verified it.

Case 2 — Margin trap. One store abandoned the homepage entirely to "most clicked" products. The most clicked products were the campaign products with the lowest margins, some even sold at a loss. Turnover increased but profits fell. Solution: "Feature the product whose margin is below the threshold on the home page" filter has been added to the recommendation rule.

Case 3 — Filter bubble. On a book site, because the customer was constantly recommended the same type of book, the customer never discovered new genres and his interest decreased over time. When the diversity rule "at least 20% of the suggestions should be from different categories" was added to the suggestion rule, discovery and sales recovered.

Weak prompt / Strong prompt

Weak prompt:

What can I recommend with this product, give me some ideas.

The model produces generic, perhaps out-of-stock recommendations without knowing your actual data and margin.

Powerful prompt:

Your role: e-commerce analyst. Main product: wireless headset (margin 28%, in stock). Complementary candidates (with stock/margin): carrying case (45%, in stock), set of earbuds (50%, in stock), charging cable (20%, in stock).Task: propose 2 cross-selling packages only with in stock and with margins over 30%; Write the customer benefit in one sentence for each package.

Common mistakes

Mistaking the suggestion for evidence. The output of AI is the hypothesis; It does not become a decision without being tested.
Ignoring inventory and margin. Automatic recommendation can highlight unprofitable or sold out product.
Just attribute it to popularity. The most clicked is not always the most profitable or most convenient.
Creating a filter bubble. Constantly showing the same product kills discovery and sales.
Not checking bias. The system can reinforce past inequalities; Regular inspection is required.
Forced upselling. Pushing an expensive model that has no benefits will cause you to lose confidence.

In summary

Recommendation systems increase basket average and loyalty by showing the customer the right product at the right moment. There are three basic logics: content-based (similar products), collaborative (bought together), and personalization. Even in small stores, AI can generate cross-selling and bundling hypotheses from order data; But these are ideas to be tested, not proof. Be sure to frame recommendation rules around stock, margin, category fit, and variety; Check the filter bubble and bias risks regularly. Automatic recommendation creates value when adhered to commercial control.

Application task

Extract the 5 product pairs most frequently purchased together from your own order data (with anonymous product codes). Ask AI for cross-selling hypotheses with the “Co-sold product analysis” template. Turn each hypothesis into an “if-then” statement and filter it through stock and margin. Write a test plan for the 2 most promising hypotheses (how long, what metric to measure). Finally, evaluate your recommendation rules for risks with the "Recommendation rule audit" template.

checklist

[ ] I treated suggestion ideas as hypotheses to be tested, not as evidence.
[ ] I filtered each suggestion through stock and margin.
[ ] Besides popularity, I also considered profitability and suitability.
[ ] I avoided the filter bubble with the diversity rule.
[ ] I checked for bias what the recommendations show to whom.
[ ] I planned to measure the promising suggestions with a real test.