What e-commerce personalization actually requires behind the scenes

01.1. Clean product data and categories

Recommendations are only as good as the catalogue. If the same item exists as three variants with different names, or half your products have no category, the system cannot tell what goes with what.

Check these first: consistent category trees, complete attributes such as size, colour, material and brand, merged duplicates, accurate stock status and good titles. A model that sees "Black Tee M" and "T-shirt, black, medium" as different products will learn the wrong things.

02.2. Event tracking of views, carts and purchases

The system learns from behaviour. You need a reliable record of what each shopper viewed, searched for, added to a cart, removed and bought. Each event needs a timestamp, a product ID and a session or shopper ID.

Common failures: tracking that misses the mobile app, events that fire twice, purchases that are not linked to earlier views and ad blockers that hide a large share of traffic. Test the tracking before you build anything on it. Compare tracked orders to actual orders. They should be close.

03.3. A customer identity that persists across sessions

A shopper who browses on Monday and buys on Thursday should look like one person. Without that, every visit is a stranger and nothing can be personal.

Use a first-party identifier that persists on the device, and link it to the customer account when they log in or check out. Plan for shared devices and for customers who use both app and web. Be careful with merging. A wrong merge shows one person's history to another.

04.4. Models or rules for recommendations

Start simple. Rules such as "bought together", "viewed together" and "best sellers in this category" are explainable and often work well. They also give you a baseline.

Add machine-learned models when you have enough events and a baseline to beat. Typical approaches look at what similar shoppers did, or what a shopper did recently. Whatever you use, put business rules on top: hide out-of-stock items, skip items already bought, respect brand and margin constraints. Always have a fallback for new visitors and new products, who have no history.

05.5. Testing to prove the lift

Personalisation feels like it works. Feelings are not evidence. Run each placement as an experiment. Keep a control group that sees the old experience, and measure average order value, conversion and click-through against it.

Run tests long enough to cover a full weekly cycle. Check that traffic was split evenly. Beware of measuring clicks alone. A block can attract clicks and still not increase revenue.

06.6. Consent handling

Personalisation depends on personal data, so consent matters. Capture the shopper's choice, respect it in tracking and in the models, and offer a non-personalised experience when they say no. Data protection rules differ by country. Ask your legal advisor what applies to you, and design the system to follow whatever you decide, including deletion on request.

07.How the pieces fit together

Picture a shopper opening a product page. The page asks the recommendation service for five items. The service looks up the shopper's identity, their recent events and the product they are viewing. It scores candidates, removes anything out of stock or already bought, applies the merchandising rules and returns a list. The page shows it and logs which items were displayed. If the shopper clicks one, that click is logged too.

Every step depends on something upstream. A wrong identity gives the wrong history. A missing event gives a thin profile. A dirty catalogue gives odd pairings. Latency matters too: if the service takes too long, the page should show a fallback instead of waiting. Plan for that fallback from the start.

08.Search is personalisation too

Search results are one of the highest-intent places to personalise. A shopper who types "boots" and has browsed waterproof hiking gear should see those first. The same data, product attributes and events feed it. Start by fixing the basics: synonyms, spelling tolerance and sensible ranking by availability. Then add light personalisation, such as boosting categories and brands the shopper has shown interest in.

Be careful with how far you push it. Two shoppers who search the same words and see wildly different results can confuse support and make testing hard. Keep personalisation as a gentle re-ranking on top of a solid base.

09.Who owns it afterwards

Personalisation is not a project that finishes. Catalogues change, seasons shift and models drift. Someone must own data quality, someone the experiments and someone the rules. In small teams these may be one person with part of their week set aside. Name them before launch, and review recommendation quality on a schedule.

10.A short checklist before you start

• Is the catalogue clean and complete?
• Is event tracking tested against real orders?
• Does a shopper keep one identity across visits?
• Do you have a baseline recommendation to beat?
• Is every placement set up as an experiment?
• Does consent status control what is tracked and shown?

11.Privacy, trust and restraint

Shoppers like relevance and dislike feeling watched. Recommendations based on a product they just viewed feel natural. Recommendations that reveal sensitive inferences do not. Review categories such as health, personal care and gifts, and decide where the system should stay neutral.

Give shoppers some control. A simple option to clear history or see why an item was suggested builds trust and is often easy to add. Keep the data you collect to what the system needs, and set retention periods. Less data is easier to protect and easier to explain.

12.A realistic first quarter

Month one is data. Audit the catalogue, fix tracking, test identity and agree the events. Month two is baseline. Ship simple recommendations on two placements, with a control group, and start measuring. Month three is learning. Read the results, fix what looks odd with the merchandising team and decide whether a model is worth building. Many teams find that rules plus clean data take them a long way before a model is needed.

Whatever you build, write down what success looks like before launch: the metric, the size of lift worth keeping and the test length. It stops arguments later and keeps weak ideas from lingering.

13.Common mistakes

• Starting with the model. Without clean data, a better model only learns the mess faster.
• Recommending popular items only. The block then looks like your best sellers page.
• No fallback. New visitors get an empty box.
• No control group. You cannot tell lift from seasonality.
• Forgetting merchandisers. Give them rules and pins. They know things the data does not.

14.What to do next

Pick one placement, such as the product page. Audit the data it would need. Fix tracking and product data. Ship a simple rule-based version. Measure it. Only then consider a model.

Brandsmashers builds the tracking, identity, recommendation and testing layers described here, and we start with the data audit because that is where results are decided. If you would like help judging your readiness, we can review your set-up with you.