AI Product Recommendations vs Traditional Tools: When to Switch

Before AI-based engines became common, stores relied on a few methods, many of which still work well.

• Manually curated related products: a merchandiser picks accessories or alternatives for each item.
• Rule-based widgets: “other items in this category”, “best sellers”, “new arrivals” or “recently viewed”.
• Bundles and kits: fixed sets defined by the team.
• Email segments built on simple conditions, such as last purchase category.

These approaches are transparent. You know exactly why an item appears, and you can change it instantly.

AI engines learn from behaviour across all shoppers. They weigh views, purchases, search queries, time spent and product attributes to rank items for each visitor or context. They update as patterns change, so seasonal trends or a viral product can move up without anyone editing a list.

Some newer systems use language and image embeddings to judge similarity, which helps with products that have little sales history. Vendors such as Nosto, Klevu, Algolia, Dynamic Yield, Bloomreach, and cloud services like Amazon Personalize or Google’s Vertex AI Search for commerce offer different mixes of these methods.

Several signals suggest manual and rule-based approaches have hit their limit.

Signals that point to AI

• The catalogue has hundreds or thousands of active products, and curated lists are out of date.
• Products change often, such as fashion collections or marketplace sellers adding items daily.
• Returning visitors are a large share of traffic, so personalisation has history to work with.
• The team spends hours each week maintaining related-product lists.
• “Best sellers” widgets show the same few products everywhere, starving the long tail.

If three or more of these describe your store, an AI pilot is likely worth the effort.

Switching is not always an improvement. Some stores lose control without gaining much.

• Small catalogues: with 30 products, a merchandiser can curate better links than a model with thin data.
• Low traffic: models need interaction data; without it, recommendations default to popularity anyway.
• Strong expert logic: technical products where compatibility matters, such as spare parts, require exact rules that an algorithm may break.
• Regulated categories: pharmacy or alcohol, where certain suggestions must never appear.
• Tight budgets: some AI engines are priced by traffic or revenue share, which can outweigh the lift for small shops.

In these cases, improve the traditional setup first: fill in missing related products, build sensible bundles and tidy category structure.

Run an A/B test on one placement, such as the product page “you may also like” section. Version A keeps your current rules or curated lists; version B uses the AI engine. Keep everything else identical, including the widget design and position.

Measure revenue per visitor as the main metric, and look at click-through, add-to-cart from the widget, average order value and return rate. A widget that increases clicks but also returns can be worse overall. Run the test across complete weeks and make sure both versions receive similar traffic sources.

Most mature stores end up with a blend. AI handles the bulk of personalised placements, while merchandisers keep manual control where expertise matters. Common patterns include pinning compatibility-critical accessories on technical products, excluding certain categories from algorithmic slots, and using rule-based “recently viewed” alongside AI “recommended for you”.

A useful principle: let the algorithm rank, let humans set the boundaries.

Traditional tools are usually free within your platform or inexpensive apps. AI engines may charge a monthly fee, a share of attributed revenue or usage-based pricing. Implementation also costs time: feed preparation, tracking checks, design changes and testing. Check current vendor pricing and ask how they define “attributed revenue”, because generous attribution windows can make a tool look better than it is.

Demos always look good because vendors pick the catalogue and the shopper. Ask questions that reveal how the engine will behave on your store.

• How does the engine handle new products with no sales history?
• Can we exclude categories, brands or margin bands from algorithmic slots?
• How is attributed revenue defined, and over what time window?
• What happens to page speed when the widget loads, and is it loaded asynchronously?
• Can we export recommendation logs to check what was shown to whom?

Clear answers to these questions are a better predictor of success than the demo itself.