Ranking that lifts revenue without eroding trust.

A recommender that only shows people more of what they already clicked will look excellent for a quarter and then quietly narrow your catalogue and your customers' interest in it.

01 / The problem

Where this usually goes wrong.

Recommendation is the area where offline metrics lie most convincingly. A model can score beautifully on held-out historical data purely by learning which items were already popular and already shown — the log records what the old system chose to display, not what customers would have preferred.

This is a feedback loop, and left unexamined it produces a system that gets more confident and less useful at the same time, while the dashboard shows improvement.

02 / What we build

The parts of the system.

Event collection that is actually reliable: Views, clicks, add-to-cart, purchase, return and dwell time, captured consistently across web and app. Recommendation quality is bounded by event quality, and fixing the tracking is frequently the highest-return work.
A two-stage architecture: Fast candidate retrieval from the full catalogue, then a heavier ranking model on a few hundred candidates. This is what keeps response times viable at catalogue scale.
Explicit cold-start handling: Content and attribute-based fallbacks for new items and new customers, so that new stock is not invisible until it accidentally accumulates interactions.
Business rules as a first-class layer: Margin, stock position, contractual placement and exclusions applied transparently on top of the model rather than smuggled into training data where nobody can audit them.

03 / How it is measured

What we agree to be judged on.

Set before the build starts, against a measured baseline, and reported honestly afterwards — including where the numbers are disappointing.

Online, against a holdout: A permanently held-out control group seeing the previous experience. Offline metrics guide development; only a live comparison settles whether it works.
Revenue per session: Not click-through. Click-through is trivially gamed by promoting whatever is most tempting to click and least likely to be kept.
Catalogue coverage: What proportion of stock is ever surfaced. A recommender that only ever shows the top 200 items is quietly deciding your merchandising strategy.
Return rate: Recommendations that lift purchases and lift returns have moved cost around rather than created value.

04 / Honest limits

What this will not do.

Recommenders amplify existing behaviour. Without deliberate exploration they converge on a narrow, self-reinforcing slice of the catalogue, so a small budget for showing uncertain items is a design requirement rather than an inefficiency.

They also cannot fix a catalogue or a pricing problem. If the right product is not stocked, or is priced wrongly, better ranking simply surfaces the disappointment faster.

05 / When it applies

You probably need this if:

A large catalogue where most items are rarely seen
Merchandising is manual and cannot keep pace with the range
Search and browse convert far worse than direct navigation
An existing recommender that nobody can explain or tune

Recognising two or more of these is a reasonable trigger for a diagnostic. Recognising none of them is a reasonable trigger for not spending money here yet.

06 / Related

Often scoped together.

These capabilities share data, infrastructure or evaluation approach with personalised recommendations, and are frequently part of the same engagement.

Next step

Is personalised recommendations the right instrument for your problem?

The diagnostic exists to answer exactly that, including the possibility that the answer is no. It is time-boxed, fixed-fee, and the written assessment is yours.

Request a diagnostic All capabilities