Will a recommendation engine increase our revenue?
It can, and we are not going to quote you a percentage. The uplift figures circulated in this industry come from other retailers' sites with other catalogues and other traffic, and they are marketing material rather than a forecast for you. What we will do is baseline your current placements before building and measure the change through a controlled test, so the answer comes from your data. If the honest answer turns out to be a small effect, you will hear that from us.
Should we fix something else first?
Very often, yes. Site search that understands your product vocabulary, accurate stock availability, and a checkout that does not lose people all typically return more than recommendations, and our ecommerce website development page makes the same point. We assess this at the start and will tell you if recommendations are not your best next investment. Saying so costs us scope and is the reason you can trust the assessment.
How does this relate to your machine learning solutions page?
That page covers the general method: how models are trained, evaluated, deployed, and monitored for drift across any domain. This page covers the retail application, where the hard parts are catalogue data quality, cold start on new lines, merchandising override, and serving inside a page render. If you want a recommender for something other than retail, start from the machine learning page. If you are a retailer, start here.
How much data do we need before this is worth doing?
There is no clean threshold, and any specific number would be invented. What we can say is the shape of the answer: content-based recommendation from catalogue attributes works with no interaction history at all and is where sites with limited traffic should start. Behavioural recommendation needs enough repeat interaction that co-occurrence patterns are not noise, which depends on your traffic, catalogue size, and purchase frequency together. We look at your actual event volume and give you a view rather than a rule of thumb.
Can our merchandising team control what gets recommended?
Yes, and we treat that as a requirement rather than an enhancement. Hard exclusions, pinning and boosting for campaigns, category restrictions, and stock filtering sit outside the model and cannot be overridden by it. Merchandisers who cannot steer the system stop trusting it, and a system the commercial team distrusts gets switched off regardless of its measured performance.
Will it recommend out-of-stock products?
No. Availability is filtered at serve time against the live stock position rather than at the point the candidate set was computed, because a set precomputed overnight will be wrong by mid-morning. This is one of the strongest arguments for the precompute-plus-filter architecture, and one of the quickest ways to destroy customer trust if it is done the other way round.
How do you handle privacy and consent?
Personalisation is applied only where the visitor's consent state permits it, and the system falls back to non-personalised logic otherwise rather than quietly proceeding. What is collected, how long it is retained, and whether it can be tied to an identified customer are decided during design with your legal or compliance function. Beyond the legal position, there is a practical one: recommendations that feel like surveillance produce complaints, and the effect on trust is not recovered by a marginal ranking improvement.
Can you work with our existing ecommerce platform?
Usually yes. Most platforms allow custom placement components calling an external ranking API, which is the integration point. Some platforms ship their own recommendation feature, and where that is adequate for what you need we will say so rather than build a replacement. The reasons to build custom are typically merchandising control the platform does not offer, catalogue characteristics its generic model handles poorly, or the need to combine signal it cannot see.
How often does the model need retraining?
It depends on how fast your catalogue and customer behaviour change. A fashion retailer with seasonal turnover needs it far more often than one with a stable range. Rather than assert a cadence, we monitor the signals that indicate staleness, including coverage drift and declining click-through against the baseline, and set the schedule from what those show. The triggers for an out-of-cycle retrain are documented at handover.
What happens if the recommendation service goes down?
The page still renders. Every placement has a cached fallback showing popular or category-appropriate items, and the ranking call has a strict timeout. A recommendation block is an enhancement to a page, and it must never be able to delay or break the page it sits on. This is designed in from the first version, not added after an incident.
Do we own the model and the code?
Yes. You receive the source code, the pipelines, the trained model artefacts, and the documentation. There is no runtime licence owed to us and nothing that prevents you continuing with your own team or another vendor. Where a project uses a third-party service for part of the pipeline, we identify it clearly during design so you know exactly what is yours and what is rented.
