A recommendation algorithm has a real job to do. There's more published online than anyone can sort through. The algorithm finds things you might care about, across sources you'd never know to look for, faster than browsing could.
The problem is what it uses to do the job. The algorithm knows what you clicked on. It knows how long you stayed. Engagement isn't the same as interest, but the algorithm treats them the same.
You click on things because they're alarming, provocative, adjacent to something you care about, hard to scroll past. The algorithm reads all of it the same way: this user engages with this type of content. It doesn't distinguish between "I wanted to read this" and "I couldn't stop myself from clicking." Over time, what gets surfaced is whatever you're most likely to engage with again. More urgent. More of the same.
The original web had a different model: people pointed at things. A writer you trusted read a lot and linked to what was worth reading. You followed one and found ten more. The signal was their judgment, not your click history. That's what a blogroll was, and what a "you should read this" still is.
That's how feeds get added in Slow Web. Someone pointed at a site, or you went looking. No algorithm suggested it.
The algorithm surfaces more, across sources you'd never find on your own. What drives it is your past behavior. What drives a recommendation from a person is their judgment about what was worth reading. In Slow Web, your sources publish. You read what arrived.