AI Hyper-Personalization: Increasing E-commerce Revenue by 40%

Almost every headline personalisation figure is measured on the surface rather than on the business. Click-through on a carousel. Revenue per visit for sessions that saw the widget. Conversion rate among users who engaged with a recommendation. Each is a real measurement of something, and none of them is a measurement of what personalisation added.

The problem is selection, and it is severe. Sessions that saw the recommendation widget are sessions that scrolled far enough to reach it, which are sessions that were already going better than average. Users who clicked a recommendation are users who were already browsing with intent. Comparing them to everyone else measures how engaged people differ from disengaged people, and attributes the entire difference to the model.

Then there is the phrase revenue influenced by personalisation, which is the most quoted figure in the category and an attribution artefact rather than a result. It counts every order in which a personalised surface appeared anywhere in the journey. On a site where personalisation is switched on everywhere, that number converges on total revenue, and it has been presented to more boards than any other statistic in retail technology.

Personalisation is not one thing, and the five things it usually means have wildly different returns. The order below is close to the reverse of the order most programmes attempt them in.

Start by measuring the null-result rate: the share of searches returning nothing, on a catalogue that contains the thing. In most retailers nobody has looked, and the figure is uncomfortable. Every one of those is a customer who told you precisely what they wanted to buy and was told you do not have it.

The causes are almost always mundane. Attribute coverage is thin, so colour, material and size are in the product photography and not in the data. Synonyms are absent, so trainers finds nothing on a catalogue full of sneakers. Typo tolerance is off or too strict. Query understanding cannot tell a brand from a category from an attribute, so a three-word query is treated as one string.

Fix retrieval before ranking, always. Personalisation on search is re-ranking a correct result set towards this particular shopper; it cannot rescue a result set that was wrong. A personalised ordering of twenty products that do not match the query is still twenty products that do not match the query, and the customer leaves either way.

Teams buy or build a recommender and then place it wherever there is room on the page. The model is rarely the constraint. What separates a recommender that earns money from one that decorates is the filtering layer between candidate generation and display, and that layer is business logic rather than machine learning.

Collaborative filtering needs history, and on a typical retail site the majority of sessions are anonymous, first-time, or too short to have produced any. Treating that as the fallback path gets the proportions exactly backwards: it is the main path, and it deserves the design attention.

What works there is content-based: product attributes, category structure, the referring query or campaign, and well-constructed popularity priors scoped to the current context rather than the whole site. Popularity within this category, this week, for this price band beats a global bestseller list by a wide margin and costs almost nothing to build.

Which points back at the catalogue again. Content-based recommendation is only as good as your attributes, and attributes are the thing nobody wants to fund.

Six months of a personalisation programme typically goes here, and that is the right place for it to go. Every model above degrades to noise without these five, and no vendor demonstration will tell you so.

There is an important distinction that gets blurred, often deliberately. Personalised promotion is showing different offers to different people at a published price — ordinary commerce, understood by customers, and the mechanism behind every loyalty scheme. Personalised pricing is showing a different price for the same product to two people at the same moment based on what the model believes they will pay.

The second carries exposure that has very little to do with the mathematics. Consumer protection regulators in several jurisdictions have taken an active interest. The optics are poor in a way that does not survive explanation, because the honest explanation is that we charged you more because we thought you would pay it. And a screenshot of two prices side by side travels further and faster than any amount of uplift.

The alternatives capture most of the value with none of the exposure: segment or market-level dynamic pricing, targeted promotions against a published price, and margin-aware ranking — putting the better-margin item first among genuinely relevant results, which is merchandising and has been for a century.

This is the part that determines whether you end up with a result or an anecdote, and it costs a small, permanent slice of revenue to do properly. Pay it.

The sequencing below is close to what we would propose for a mid-sized retailer, and it is deliberately back-loaded on the things vendors demonstrate first.

Three of these only appear after six months, which is why they survive the pilot and damage the programme.

Not forty per cent. A realistic programme, sequenced as above and measured against a genuine holdout, tends to produce a mid-single-digit improvement in revenue per user in the first year, concentrated in search, with the compounding arriving in the second — and that is a good outcome, obtained honestly, on a number you can put your name to.

The forty per cent is available too. It is just measured on a carousel.