AI for e-commerce: increase revenue with smart product recommendations

AI product recommendations can help an online shop show the right item at the right time: the accessory for the device just bought, the alternative to a sold-out model, the refill pack after a few weeks. This article explains how such recommendations work, what your shop needs to have in place and how to measure properly whether they pay off.

How AI product recommendations work

Most recommendation systems rest on two basic ideas, which are often combined. The first looks at the behaviour of many customers: people who bought item A often also bought item B. That produces the familiar “Customers also bought” block. The second compares the products themselves by their attributes, such as material, size, price range or intended use, and suggests similar ones.

Modern systems also use behaviour within the current session. Someone who looks at three cordless hedge trimmers in a garden shop is more likely to be shown matching spare batteries and gloves than potting compost. Language models add another option: customers describe in their own words what they are looking for and get a selection with reasons. “I need walking boots for day hikes, fairly wide fit, no more than 150 euros” is hard for a classic filter bar, but a normal request for an AI adviser in the chat. How this kind of real-time personalisation can be used in sales beyond the shop is described in real-time personalisation in sales with AI.

Where recommendations make sense in the shop

Not every spot works equally well. These placements have proved their worth in particular:

• Product page: alternatives for the undecided and accessories that really go with the item.
• Basket: small additions such as batteries, care products or protective cases, without getting in the way of checkout.
• Sold-out items: an available alternative straight away instead of a dead end.
• After purchase: reminders about consumables when they are likely to be running low.
• In the advice chat: answers to specific product questions with a reasoned recommendation.

A bike dealer with an online shop can, for example, offer helmets in the right head size when someone buys a children's bike. A pet supplies shop can remind customers to reorder food. In both cases, what matters is that the recommendation saves the customer effort rather than simply padding the basket. Unsuitable suggestions quickly come across as pushy and damage trust. You can read more about approaching existing customers sensibly in the article on cross-selling and upselling with AI.

Prerequisites: data, product range, technology

A recommendation system is only as good as the data it works with. Three points decide whether it succeeds.

Clean product data

If sizes, materials and compatibility aren't maintained, the AI will suggest the wrong charging cable. Maintaining product attributes is the least glamorous but most important groundwork. The article Preparing sales data for AI also covers this topic.

Enough sales history

Behaviour-based recommendations need lots of orders to spot patterns. A shop with a few dozen orders a month doesn't have that data. In that case, rule-based recommendations you define yourself are often the better choice.

Consent and the legal framework

Personalisation based on browsing behaviour usually relies on cookies or similar technology, which requires consent under the GDPR and the TDDDG (German Telecommunications Digital Services Data Protection Act). Recommendations by email to existing customers are only permitted without separate consent under narrow conditions, for example for similar products and with a clear option to object. Have your consent solution and email sequences checked by a lawyer before you personalise.

Worked example: are product recommendations worth it for your shop?

You can roughly estimate in advance whether AI recommendations will pay off for your shop. The following figures are purely assumptions for illustration, not empirical values.

Assumption	Value
Orders per month	2,000
Orders that also include a recommended item	100
Average price of the additional item	15 euros
Additional revenue per month	1,500 euros

Of these 1,500 euros, only the contribution margin matters, meaning revenue minus cost of goods, shipping and payment fees. Set this amount against the system's monthly costs and the maintenance effort. If the contribution margin is well above them, a test is worthwhile. If it is just below, there are usually cheaper levers.

Measure the result with a proper comparison: some visitors see the recommendations, others don't. That is the only way to tell whether the additional revenue really comes from the recommendations or would have come anyway. The article on A/B testing in AI sales describes how to set up such tests.

Recommendations in conversation: the AI adviser

Many abandoned purchases happen not because the right product is missing but because a question remains unanswered. Does the case fit the older model? Is the mattress suitable for side sleepers? When will the delivery arrive? An AI adviser in the website chat or on WhatsApp can answer such questions in the evening too and suggest suitable items based on your product data. It should identify itself as an AI and hand over to your team for complaints or complicated cases. An example scenario on AI customer service in an online shop walks through what this can look like in practice.

At Neurobots, AI employees handle exactly these advice conversations. They reply in the website chat, on WhatsApp, by email or on the phone, at night too, and pass interested contacts on to your CRM. Data is stored in GDPR-compliant data centres in Frankfurt.

When it isn't worth it

With a small range of a few dozen items, your customers usually know the offer already, and a hand-maintained list of matching products is enough. It is similar with one-off items such as art or unique pieces, where there are no patterns from repeat purchases. And if your product data has gaps, start there before you put money into technology.

Frequently asked questions

Can't my shop system do this already?

Many shop systems offer simple recommendation features or extensions for them. That is often enough to start with. A dedicated AI system is more worthwhile if the range is large, there are many orders or you want to use recommendations in the chat and in emails too.

How quickly can you see whether it works?

That depends on order volume. A test needs enough orders in both groups so that differences aren't down to chance. For smaller shops this can take a few weeks, and you should allow for seasonal fluctuations.

Don't personalised recommendations feel creepy to customers?

They can, if suggestions feel too personal or come from data the customer doesn't expect you to have. Stick to links to the current purchase that make sense, and show why an item is being recommended where needed.

Do I need consent for AI product recommendations?

For recommendations based on tracking via cookies or similar technologies, usually yes. Recommendations based only on the current basket are often less critical. The exact assessment belongs in the hands of your data protection officer.

Conclusion

AI product recommendations don't run themselves, but they are a useful tool for shops with a sufficient range and order volume. What matters are well-maintained product data, suitable placements, proper consent and an honest test that shows what the recommendations really bring.

The page AI sales assistant for online shops shows how a digital adviser answers product questions and helps potential buyers along.