Count the recommendations you’ve accepted today. The playlist Spotify assembled, the video queued next, the products under “you may also like”, the articles your news app surfaced, possibly the person you’re dating. We used to choose from what friends suggested and reviews praised. Now a substantial share of daily decisions arrive pre-ranked by algorithms, and it’s worth knowing how the ranking happens.
Under the hood there are really three ideas, endlessly remixed.
Collaborative filtering ignores what things are and watches what people do: people whose history overlaps yours chose this next, so you might too. It powers the classic “customers also bought”, and its known weakness is the cold start: a new user or new item has no history to match on.
Content-based filtering works from attributes instead: you liked these dark Scandinavian thrillers, here’s another with the same fingerprint. No crowd needed, but it can trap you in more-of-the-same.
Hybrids blend both, which is what serious systems actually run: Spotify folding your listening history together with audio features and playlist co-occurrence; Steam mixing what you’ve played with what a game is; Tinder ranking by stated preference and by revealed swiping. Even newsrooms run them: the article you just read was probably followed by three chosen this way.
I got to build one of these where the stakes were unusually concrete: commercial hotels. At JLL, one of the world’s largest commercial real-estate services firms, a hotel coming to market meant finding the hundred likeliest buyers in a book of two hundred thousand investors, a search that took a broker up to three months by hand. The original sketch of the problem still says it best:
The twist that made it hard: buyers don’t rate hotels the way viewers rate films. Our interaction matrix was 99.35% empty, so the signal had to come from behaviour — enquiries, bids, stated mandates — rather than ratings. The full case study covers the bake-off, including the baseline that embarrassed a fancier algorithm.
So should you be glad these systems exist, or worried? Both, and specifically. The genuine gift is discovery at scale: the back-catalogue song, the niche game, the buyer nobody thought to call. The genuine costs are narrowing (optimising for what you did before quietly shrinks what you see next), and the gap between engaging and good for you — an engine tuned for clicks will happily feed any appetite that clicks.
My practical advice is small but real: notice the ranking. A recommendation is a prediction about you with a business model attached, and simply remembering that, then deliberately wandering outside the feed now and then, keeps the algorithms where they belong, working for you rather than the reverse.