Faceted filter

This is the control surface, not the technique. Faceted search is the idea that a result set can be cut down along several independent attributes at once, and that the cuts combine; a faceted filter is the panel you actually click, the one with Category, Price and Brand stacked down the side of a results page. The distinction is worth keeping because the technique can be delivered other ways (a row of dropdowns, a query language, a set of tabs), and because the panel has craft problems of its own that have nothing to do with how the index is built.

The counts are what separate a faceted filter from a pile of checkboxes. A number beside an option is a promise: tick me and you will have this many results. Keeping that promise means each count is measured with every other group’s constraints applied but not its own, which is why ticking a second brand usually adds results while ticking a second price band often does not. It also means an option that would leave nothing must say so. Showing a zero and disabling the row teaches the shape of the data; hiding the row makes the panel jump around under the reader’s hand, and leaving it clickable sends them to an empty no results state the panel could have predicted.

Each group needs a decision about how many of its options can be true at once, and the control should say which it is. Attributes a thing can have several of (colour, feature, brand) are checkboxes and combine with OR inside the group; attributes it can only have one of (a price band, a delivery speed) are a radio group and replace rather than add. Between groups the combination is AND, which nobody has ever needed explaining but which quietly makes the third or fourth tick feel unpredictable. What fixes that feeling is not an explanation, it is the count.

Everything applied has to be visible outside the panel too, because a reader who has scrolled the results is looking at a list without knowing why it is short. That is what the applied filters row of removable chips above the results is for, together with a Clear all that puts everything back in one action. The last question is when the results update: filtering as each option is ticked is right when the response is instant, and an explicit Apply is right when it is not, or on a small screen where the panel arrives as a bottom sheet over the results it is about to change.