AI Product Recommendations for Nigerian E-commerce: A Merchandising Guide

Most Nigerian stores discuss recommendations as a feature to install. It is more useful to treat it as merchandising: deciding what to put in front of a shopper who has already shown intent, then measuring whether they bought more because of it.

This article covers the applied side: placements, the rules-versus-AI decision, recommending inside chat where many Nigerian sales happen, and how to measure lift without fooling yourself. AI Recommendations for E-commerce Businesses explains how the underlying engines work if you want the mechanics.

What product recommendations actually do

A product recommendation is a suggestion of what to view or buy next, generated from product attributes, other customers' behaviour, or the current shopper's own history. In a store it does three separate jobs, and they should be designed separately:

• Discovery. Helping someone find the item they came for when your catalogue is too big to browse.
• Attachment. Adding the accessory, refill or matching piece that belongs with the main purchase.
• Return. Bringing a past buyer back at the right moment with the right item.

Only the second job reliably raises order value in the short term. Discovery raises conversion, and return raises repeat purchase rate. Decide which you need before choosing a solution.

The five placements that earn money

Indicative guide. Test on your own store; placement value varies by category.

Two placements deserve caution. Homepage recommendations rarely pay for a small catalogue, and post-checkout upsells that delay the confirmation message erode trust in a market where buyers are already nervous about paying online.

Rules or AI? Let your catalogue decide

The honest answer for most Nigerian stores is that you should start with rules and graduate.

A useful middle path: use an AI assistant to analyse a year of order data and propose the pairings and bundles, then implement those as fixed rules in your store. You get the insight of the model without the cost and maintenance of a live engine.

The data you already have, and what to start collecting

Recommendations run on data your store probably holds in fragments:

• Order history: which items appear in the same order, and in which sequence over time.
• Product attributes: category, brand, size, colour, material, price band, occasion.
• Customer identity: phone number is the most reliable key in Nigeria, because the same buyer may use different names and several delivery addresses.
• Behaviour: page views, searches and cart adds on your website, if analytics is set up properly.
• Chat history: what customers ask for by name in WhatsApp and Instagram, which is a goldmine most stores throw away.

Start collecting now, even before you need it:

• Tag every product with structured attributes, not just prose descriptions
• Store orders with line items, not as a single total
• Deduplicate customers by phone number
• Record the channel each order came from
• Log out-of-stock requests, which tell you what to recommend and what to restock
• Keep returns data, so you stop recommending items people send back

Recommending inside WhatsApp and Instagram

For chat-led Nigerian stores, this is where recommendations produce most of their value, and where most tools ignore them.

Patterns that work:

• Two or three options, never a catalogue. When a customer asks for a product, return the closest match plus one alternative at a different price point, each with price and availability.
• The natural attachment. Phone with case and screen protector, gown with matching headwrap, blender with spare jar, generator with service kit. One suggestion, stated as useful rather than pushy.
• Budget branching. "If your budget is around ₦45,000, this one; if you can go to ₦60,000, this one lasts longer." Nigerian buyers respond well to explicit price laddering.
• Post-delivery follow-up. Two days after delivery, suggest the accessory or refill.
• Restock alerts. When a customer asked for something out of stock, message them the day it returns. It is the simplest recommendation and among the most effective.

An AI assistant connected to your catalogue can do all of this automatically. AI Chatbots for Nigerian Online Stores covers building that assistant.

Bundles, upsells and the naira maths

Work out the value before building anything. Take your last 90 days:

• Average order value (AOV), total revenue divided by orders.
• Attachment rate: percentage of orders with more than one item.
• Contribution per extra item: the margin, not the price.

Suppose a store has 500 orders a month at an AOV of ₦38,000, and an attachment rate of 18%. If well-designed recommendations lift attachment to 26%, that is 40 extra attached items a month. At an average add-on margin of ₦4,000, the monthly gain is ₦160,000, or about ₦1,900,000 a year. That maths tells you whether a ₦2,000,000 engine makes sense or whether a week of manual pairing work is the right answer this year. Run it with your own numbers; do not assume the percentages above.

Bundle design rules for Nigeria:

• Bundle for convenience and price clarity, not just discount. "Complete set, one price, one delivery" is persuasive when delivery costs are separate.
• Keep the add-on well below the main item's price. Add-ons above roughly a quarter of the basket value require a new decision, and new decisions cost conversions.
• Never bundle in an item you cannot reliably keep in stock.

What changes for Nigerian online stores

• Chat is a recommendation surface. In markets where nearly all sales pass through a website, recommendations are visual modules. In Nigeria, a significant share must work as sentences in WhatsApp.
• Delivery cost changes basket logic. Because delivery is usually charged separately and can be significant, "add this now and pay one delivery fee" is a stronger argument here than a small discount.
• Stock volatility. Importers face lead-time and FX variability. Recommending items that are frequently out of stock trains customers to ignore recommendations entirely. Filter by live stock, always.
• Thin behavioural data. Many Nigerian stores have modest website traffic, so collaborative filtering has little to learn from. Order history and attributes carry more weight than browsing data.
• Price sensitivity and laddering. Showing a cheaper alternative alongside the item viewed is not cannibalisation; it often saves the sale.
• Mobile and data cost. Recommendation modules with heavy images slow pages on mobile data. Keep images light and limit the number of items shown.
• Personal data. Recommendation systems process purchase histories and phone numbers, which are personal data under the Nigeria Data Protection Act 2023. Keep access restricted and verify current obligations with the Nigeria Data Protection Commission.

Example (hypothetical): an Ibadan fashion store

This is an illustrative scenario, not a Linestech client result.

A fashion store carries about 320 products across ready-to-wear, fabrics and accessories. Orders come roughly half from the website, half from Instagram and WhatsApp. Average order value is modest and most orders contain a single item.

Phase one, no engine: the team exports twelve months of orders, uses an AI assistant to find which items are bought together and in what sequence, and implements the top 25 pairings as fixed related-product rules on the website and as saved suggestions for the chat team. They also add a restock-alert list.

Phase two, six months later: with cleaner data and more traffic, they add a recommendation engine that personalises the product page and cart by browsing and purchase history, and they automate a 45-day follow-up suggesting accessories for past buyers.

Plausible outcome, not a guarantee: more multi-item orders, a measurable increase in repeat purchases from the follow-up flow, and clearer insight into which fabrics drive accessory sales. The important point is that phase one cost a week of work and informed phase two.

What recommendations cost in Nigeria

Indicative 2026 ranges; actual quotes vary with scope, catalogue size, vendor and exchange rate. Compare two or three written quotations on identical scope.

Budget 15–25% of build cost per year for maintenance on custom work, and remember that an engine needs someone to watch its output as the catalogue changes.

How to implement recommendations in eight steps

1. Define the job. Discovery, attachment or return. Pick one to start.
2. Measure the baseline. AOV, attachment rate, repeat purchase rate, for the last 90 days.
3. Clean product attributes. Category, price band, colour, size, brand, occasion, in fields.
4. Analyse co-purchases. Export order line items and find what sells together, by product and by category.
5. Implement the top rules on the product page, cart and confirmation message, filtered by live stock.
6. Extend into chat. Give the team or the assistant the same pairings and the price-ladder script.
7. Add the return flows. Replenishment reminders and restock alerts.
8. Review after 60 days. Compare against baseline and decide whether an engine is justified.

How to measure lift honestly

Recommendation vendors report clicks. Clicks are not money. Measure it this way:

• Compare periods properly. Watch for seasonality: December, Ramadan, back-to-school and salary week all distort a simple month-on-month comparison.
• Split traffic where you can. Show recommendations to half your visitors and not the other half for a defined period. If your traffic is too low for that, run a clean before-and-after on a stable period instead and be honest about the uncertainty.
• Track incremental revenue, not attributed revenue. An item the customer would have bought anyway is not lift.
• Watch returns. If recommended items are returned more often, your lift is smaller than it looks.
• Report margin, not revenue. Accessories often carry better margin than headline products, which can make a modest revenue gain a large profit gain, or the reverse.

Mistakes to avoid

• Recommending out-of-stock items. One disappointment trains a customer to ignore the module permanently.
• Recommending near-identical products at checkout. At the cart, the customer has decided. Show complements, not alternatives.
• Ignoring chat. For many Nigerian stores, most recommendation opportunities happen in a conversation, not on a page.
• Buying an engine for a 70-product catalogue. There is not enough variety for personalisation to add anything a good merchandiser cannot.
• Letting the model pick blindly. Exclude clearance, damaged, discontinued and low-margin items explicitly.
• Heavy modules on mobile. Slow pages cost more conversions than recommendations recover.
• Never reviewing it. Seasonal relevance changes. A recommendation set left untouched for a year sells last season's stock.

Conclusion

Recommendations for a Nigerian store are a merchandising decision first and a technology decision second. Define whether you need discovery, attachment or return; measure your baseline AOV and attachment rate; implement rule-based pairings from your own order history across the product page, cart, confirmation message and chat; then decide whether the naira maths justifies an engine. Filter everything by live stock, keep modules light for mobile, and measure incremental margin rather than clicks.

If you want product recommendations working across your website and your WhatsApp conversations, with live stock filtering and honest measurement, Linestech builds recommendation and personalisation systems for Nigerian e-commerce businesses.

Frequently asked questions

How many products should I show in a recommendation block?

Two to four on a product page, one or two at the cart, and no more than three in a chat message. Long lists push the decision back to the customer, which defeats the purpose, and they slow mobile pages on metered data.

Can I do recommendations if my store runs on WooCommerce or Shopify?

Yes. Both support related products, manual pairings and apps for upsells and frequently-bought-together logic. Custom engines can also plug into them through APIs. The constraint is usually data quality, not the platform.

Will recommendations work if I sell only 40 products?

Manual merchandising will, and it will beat any algorithm at that size. Choose the two or three natural companions for each product yourself, show them consistently, and revisit quarterly. Spend the budget on traffic and conversion instead.

How is this different from personalisation?

Recommendations suggest products; personalisation adapts the whole experience, including content, offers, search results and messaging, to a specific customer. Recommendations are usually the first component of personalisation and deliver most of its early value.

Can recommendations work in email and SMS too?

Yes, and in Nigeria WhatsApp and SMS typically reach customers more reliably than email. Use purchase history to drive replenishment and win-back messages. Respect consent and provide an opt-out, in line with the Nigeria Data Protection Act 2023.

What data privacy issues should I consider?

Purchase history, phone numbers and browsing behaviour are personal data. Collect what you need, state the purpose in your privacy notice, restrict internal access, set retention periods, and check current requirements with the Nigeria Data Protection Commission before sharing data with third-party tools.

How soon should I expect results?

Rules-based pairings and restock alerts can show a difference within weeks because they are immediate and specific. A learning engine needs enough traffic and orders to train and tune, so allow a full quarter before judging it, and only judge it against a measured baseline.

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.

• Nigeria Data Protection Commission (NDPC)
• OpenAI platform documentation
• Anthropic developer documentation
• Nielsen Norman Group UX research
• Baymard Institute e-commerce UX research