13 December 2025 · 4 min read
Start with rules and popularity
Trending, recently added, and category-based rules produce reasonable results immediately and give you a baseline to beat.
Many products never need more than this. Measure before building a model.
Content-based fills the gap
Item similarity from attributes and text works without user history and handles new items immediately.
This also solves the new-item cold start that collaborative filtering never fully escapes.
Measure against the baseline
A model that does not beat 'most popular in category' is a cost with no benefit. Test it honestly.
Hold out a control group permanently so you can always answer whether the system is helping.
Where to go from here
If you are weighing recommendation systems for an upcoming project, the useful next step is usually a short scoping conversation rather than a longer article. Bring your constraints — budget, timeline, existing systems, and the team you already have — and the decision usually resolves itself in an hour.
We run these as a fixed-scope discovery session with a written recommendation at the end, including the cases where our advice is that you do not need us.
Working on something like this?
Send us the brief. We will tell you what we would build, what we would cut, and roughly what it costs.
Talk to us