AI Recommendations for E-commerce Businesses: How Recommendation Engines Work and Which One Fits Your Store
Recommendations are the part of e-commerce most shoppers never notice and most store owners get wrong. Done well, they raise average order value quietly: the phone case next to the phone, the second pack of diapers at a bundle price, the dress a returning customer would have searched for anyway. Done badly, they show a customer the exact item she just bought, or a random selection that looks like clutter.
This article explains how recommendation engines actually work, the four approaches and their trade-offs, where recommendations should appear (including WhatsApp, which most overseas guides ignore), the small-catalogue and cold-start problems typical of Nigerian stores, a labelled hypothetical example, indicative costs and implementation steps. If you want the business case for recommendations specifically in the Nigerian e-commerce market, the article on AI product recommendations for Nigerian e-commerce covers that; this one is about choosing and building the engine.
What a recommendation engine does
A recommendation engine is a system that ranks your catalogue for a specific shopper in a specific context and returns the top few items to display. The context might be a product page (what goes with this?), a cart (what completes this order?), a home page (what is this person likely to want today?) or a follow-up message (what should we tell this customer about?).
Every engine answers the same question with different evidence:
- Behaviour of similar shoppers: what people with similar histories bought.
- Attributes of similar products: category, brand, price, colour, size, description.
- This shopper's own history: viewed, searched, bought, returned.
- Business rules: margin, stock availability, delivery area, promotions.
The last item is the one owners forget. An engine that recommends out-of-stock products, items that cannot be delivered to the customer's city, or the lowest-margin option every time is working against you. Rules are part of the engine, not an afterthought.
The four approaches: collaborative, content-based, hybrid and LLM
The difference between collaborative filtering and content-based recommendation is that collaborative filtering learns from what shoppers do together (purchase and viewing patterns across customers), while content-based recommendation matches products by their attributes. Collaborative needs lots of interaction data; content-based works from the catalogue alone. Hybrid and LLM approaches combine or extend both.
| Approach | How it works | Strengths | Weaknesses | Best for |
|---|---|---|---|---|
| Collaborative filtering | Finds patterns across shoppers' purchases and views | Discovers non-obvious pairings; improves with volume | Cold start for new products and new shoppers; weak on small catalogues | Stores with thousands of orders a month |
| Content-based | Matches products by attributes and descriptions | Works from day one; handles new products | Can be obvious ("more black shoes"); needs good product data | Small and mid-size catalogues; new stores |
| Hybrid | Combines both, with rules on top | Best overall quality; handles cold start | More complex to build and tune | Growing stores with some history |
| LLM-driven | A language model reads product data and shopper context to reason about fit | Understands descriptions, occasions and natural-language queries; explains suggestions | Per-request cost; latency; needs guardrails against inventing products | Stores with rich descriptions, conversational selling on WhatsApp |
Most platform plugins (for Shopify, WooCommerce and similar) are content-based or simple collaborative ("customers also bought") under the hood. Custom engines are usually hybrids, and LLM-driven recommendations are increasingly used as the conversational layer for WhatsApp and on-site assistants rather than as the sole engine.
Which approach fits your store?
A practical decision framework for a Nigerian e-commerce business:
- Under about 500 orders a month or fewer than 300 products: content-based recommendations via a platform plugin plus hand-curated bundles. Collaborative filtering will not have enough signal to beat a good merchandiser.
- 500 to 5,000 orders a month, growing catalogue: hybrid engine, either a stronger plugin with attribute and behaviour models or a lightweight custom service. Add rules for stock, delivery area and margin.
- Above 5,000 orders a month, or a marketplace with many sellers: custom hybrid engine with real-time event tracking and experimentation (A/B testing of placements).
- Selling conversationally on WhatsApp regardless of size: an LLM layer that draws on your catalogue and the customer's history to suggest products in chat, with strict grounding so it only recommends real, in-stock items. The article on AI WhatsApp chatbots for Nigerian businesses covers the chat side.
Ask a vendor two questions before buying anything: how does the engine handle a product with no sales history, and how does it avoid recommending items that are out of stock or undeliverable to this customer? Vague answers mean the engine will embarrass you.
Where should recommendations appear?
Placement matters as much as the algorithm. Each location has a job:
| Placement | Job | Typical logic |
|---|---|---|
| Product page | Cross-sell and alternatives | Frequently bought together; similar items in stock |
| Cart and checkout | Complete the order | Accessories, consumables, bundle discounts; keep it to two or three items |
| Home page (returning visitor) | Re-engage | Recently viewed, category the shopper favours, restocked items |
| Search results and empty results | Recover intent | Closest matches when exact search fails |
| Post-purchase page and email | Next purchase | Replenishment timing; complementary categories |
| WhatsApp follow-up | Personal nudge | One or two items chosen from history; only with opt-in and template rules |
| Order confirmation and delivery messages | Low-friction add-on | "Add to this order before dispatch" where operations allow |
For Nigerian stores, the WhatsApp placements often outperform on-site ones because that is where customers already talk to the business. A recommendation delivered as "the size 4 nappies you ordered last month are back in stock, and the wipes are ₦300 cheaper this week" reads as service rather than marketing.
What data does a recommendation engine need?
A recommendation engine needs a clean product catalogue with attributes, and event data linking shoppers to products: views, add-to-cart, purchases and, ideally, searches and returns. Content-based engines can start with the catalogue alone; collaborative engines need months of interaction history across many shoppers.
Checklist:
- Product catalogue with consistent categories, brand, price, variants (size, colour) and stock status.
- Product descriptions and images, which LLM and content-based approaches rely on.
- Order history linking customer identifier to products, with dates.
- On-site events: product views, searches, add-to-cart, checkout abandonment, tied to a session or customer.
- WhatsApp and offline sales imported against the same customer identifier where possible.
- Stock and delivery-area data available to the engine at request time.
- Margin or priority data for rules.
The most common gap in Nigerian stores is variant and attribute quality: products uploaded with inconsistent categories, missing sizes and "N/A" descriptions. A content-based engine is only as good as the attributes it can read, so catalogue clean-up is usually the first task.
What changes for Nigerian online stores
Small catalogues and thin traffic. Many Nigerian stores have a few hundred products and a few hundred orders a month. Collaborative filtering struggles at this scale; content-based rules, curated bundles and LLM reasoning over descriptions do the work. Do not pay for a "machine learning" engine that needs data you do not have.
Sales outside the website. A large share of orders come through WhatsApp and Instagram DMs and never touch the site. If those are not recorded against the customer, the engine sees half the picture. Log them in the order system.
Delivery constraints. Recommending a heavy or fragile item to a customer outside your reliable delivery zones creates a sale you cannot fulfil well. Feed delivery-area rules into the engine.
Price sensitivity and bundles. Recommendations that lead with a bundle saving or free-delivery threshold ("add ₦2,500 more for free delivery") tend to land better than pure "you may also like" lists.
Stock volatility. Import-dependent stores run out and restock unpredictably. Real-time stock checks before display are essential, and "back in stock" recommendations are unusually effective.
Data and page weight. Recommendation widgets that load slowly on a mid-range Android phone over mobile data cost more conversions than they gain. Keep them light and lazy-loaded.
Privacy. Behavioural tracking for recommendations is personal-data processing under the Nigeria Data Protection Act 2023; disclose it in your privacy notice and apply the same consent discipline as for marketing.
Example (hypothetical): an online baby and toddler store
Example (hypothetical): an online store in Lagos sells nappies, formula, feeding gear, clothing and toys, with about 900 products and roughly 1,200 orders a month, 40% of them via WhatsApp. Recommendations are currently a static "popular products" block.
A proportionate build:
- Catalogue clean-up (weeks 1 to 3). Standardise categories, add age ranges and sizes as attributes, fix descriptions, mark consumables with a typical replenishment interval.
- Content-based plus rules (month 2). Product pages show complementary items by category pairing (nappies with wipes and rash cream; bottles with sterilisers), filtered by stock and the shopper's delivery zone. Cart shows a "reach free delivery" suggestion.
- Replenishment recommendations (month 3). For consumables, a WhatsApp template goes out shortly before the predicted run-out date, with the last-ordered size and one complementary item. Opt-in is recorded at first purchase.
- Hybrid layer (month 6 onwards). With on-site events and WhatsApp orders now logged against customers, a lightweight collaborative model re-ranks the content-based candidates. An LLM assistant on WhatsApp answers "what do I need for a six-month-old starting solids?" by grounding on the catalogue.
- Testing. Each placement runs with a control variant so the team can see whether recommendations change order value or just decorate the page.
This is a hypothetical scenario, not a Linestech client result.
How much does a recommendation engine cost in Nigeria?
For a Nigerian online store, the main cost drivers of a recommendation engine are the approach (plugin, hybrid, custom, LLM), the state of the catalogue and event data, the number of placements including WhatsApp, and whether real-time stock and delivery rules must be integrated. Indicatively, plugin-based recommendations with catalogue clean-up cost ₦400,000 to ₦1,500,000; a custom hybrid engine ₦2,000,000 to ₦5,000,000 or more; LLM conversational recommendations on WhatsApp ₦1,500,000 to ₦5,000,000; plus recurring costs.
| Component | Indicative 2026 range | Notes |
|---|---|---|
| Catalogue and attribute clean-up | ₦200,000 – ₦1,000,000 | Usually the first and most valuable task |
| Platform plugin setup and rules | ₦200,000 – ₦800,000 | Plus subscription, often US$-priced |
| Event tracking (views, carts, searches) | ₦300,000 – ₦1,200,000 | Needed for collaborative and testing |
| Custom hybrid engine and API | ₦1,500,000 – ₦4,000,000+ | Includes stock and delivery rules |
| WhatsApp replenishment and follow-up flows | ₦500,000 – ₦2,000,000 | Plus per-message fees |
| LLM conversational recommendation layer | ₦1,000,000 – ₦4,000,000 | Grounding, guardrails, catalogue sync |
| A/B testing and reporting | ₦300,000 – ₦1,000,000 | Measures real uplift |
| Hosting and model usage | ₦150,000 – ₦800,000 per year plus US$ API fees | Exchange-rate sensitive |
| Maintenance and tuning | 15 – 25% of build per year | Catalogue changes, seasonal retuning |
All figures are indicative 2026 ranges; actual quotes vary with scope, vendor and exchange rate. Separate one-off build from recurring plugin, hosting and model fees, and compare two or three written quotations on the same scope. The related article on e-commerce website cost in Nigeria places these figures within an overall store budget.
Step-by-step: adding recommendations to your store
The first step is to clean the catalogue so products carry consistent attributes; the second is to deploy content-based recommendations with stock and delivery rules in two or three placements; the third is to instrument events and measure against a control. Collaborative and LLM layers come after the data exists.
- Audit the catalogue: categories, variants, descriptions, stock accuracy.
- Choose placements: start with product page and cart; add WhatsApp replenishment if you sell consumables.
- Deploy content-based recommendations with rules for stock, delivery zone and margin.
- Log all orders (site, WhatsApp, Instagram) against a customer identifier.
- Add event tracking on the site and app.
- Set up a control for each placement and measure order value and conversion.
- Introduce a hybrid model when interaction data supports it.
- Add an LLM conversational layer for WhatsApp or on-site assistance, grounded strictly on the catalogue.
- Review monthly: uplift, out-of-stock recommendations shown, customer complaints, page speed.
The article on how to build a mobile app with AI recommendations covers the app-specific version of these steps.
Mistakes to avoid
- Recommending what the customer just bought. Exclude recent purchases unless the item is a consumable due for replenishment.
- Ignoring stock and delivery. Every recommendation must be checkable against availability and deliverability at display time.
- Buying collaborative filtering for a small store. It will show noise; use content-based rules and curation.
- Letting an LLM invent products. Ground it on the live catalogue and block anything not in it.
- Cluttering pages. Three good suggestions beat twelve random ones, especially on mobile.
- No control group. You will not know whether the engine earned its cost.
- Forgetting WhatsApp. For many Nigerian stores it is the highest-performing placement.
- Slow widgets. Test on a mid-range Android phone over mobile data before launch.
Conclusion
Recommendation engines are not a single product you buy; they are a choice of approach matched to your catalogue size, order volume and channels. For most Nigerian online stores the right sequence is clean catalogue data, content-based recommendations with stock and delivery rules, WhatsApp replenishment flows, honest measurement against a control, and only then collaborative and language-model layers. Get the sequence right and recommendations become quiet, dependable revenue rather than page decoration.
If you are deciding whether a plugin or a custom engine fits your store, or want recommendations that work inside WhatsApp as well as on your site, Linestech can help you assess your catalogue and order data and build the right layer.
Frequently asked questions
Do I need a lot of traffic before recommendations are worth it?
No. Content-based recommendations and curated bundles work from day one because they use the catalogue, not traffic. What needs volume is collaborative filtering, which learns from patterns across many shoppers. Start with attribute-based rules and let interaction data accumulate for a later hybrid layer.
Can recommendations work on WhatsApp rather than a website?
Yes, and for many Nigerian stores they work better there. Replenishment reminders, restock alerts and one or two complementary suggestions can be sent as approved WhatsApp Business Platform templates with opt-in. A grounded language model can also suggest products in live chat. Log WhatsApp orders against customers so the engine sees them.
What is the difference between upselling and cross-selling in recommendations?
Upselling suggests a higher-value version of what the shopper is considering (a larger pack, a better model). Cross-selling suggests complementary items (a case with a phone). Recommendation engines handle both; cross-selling is usually the safer default in price-sensitive markets, while upselling should be reserved for clear value differences.
Will an AI recommendation engine slow down my store?
It can if built carelessly. Well-designed recommendations are computed in advance or cached, loaded after the main content, and kept to a few lightweight items. Test on typical Nigerian devices and connections before going live, and prefer fewer, faster suggestions.
Can I use ChatGPT-style AI to recommend products?
A language model can recommend products if it is connected to your live catalogue and constrained to it, which is the grounding work a developer does. Used on its own, it will happily invent products, prices and availability. It is best applied to conversational contexts (WhatsApp, on-site assistant) and natural-language questions, layered on top of a rules or hybrid engine.
How do I measure whether recommendations are working?
Run each placement with a control group that sees no recommendations or the old static block, and compare average order value, conversion rate and revenue per visitor over a few weeks. Also track how often the engine shows out-of-stock or undeliverable items, and whether page speed changed. Judge on these numbers, not on click counts alone.
Sources and further reading
Figures, platform rules and regulations change. These are the primary references behind this article and the places to check before you act on it.
