Search as an Argument
Type a dancer’s name into YouTube and you get thirty uploads of the same three dances, with a viral cumbia on top. This is how TangoTube reads what you meant, answers with dances instead of files, ranks by what people who love this dance value, and lets every click argue back.
- distinct things people searched for
- 13,190distinct things people searched for
- dances, the unit of a result
- 70,321dances, the unit of a result
- where view counts stop mattering
- 300kwhere view counts stop mattering
- real queries checked on every build
- 11real queries checked on every build
- empty result pages allowed
- 0empty result pages allowed
Part 01
Read what they meant before looking anything up
Search here has three jobs YouTube cannot have, because they depend on the catalog: understand what tango thing you mean, answer with dances rather than uploads, and rank by what enthusiasts value. All three are tested on every build.
Part 02
The result is a dance, not a file
Uploads are grouped into dances by channel, dancers, and date, then by look-alike thumbnails for reposts. Multi-part sets are one result with numbered part pills; switching parts happens inside the card. There is no toggle for the flat list, because the flat list is a defect, not a preference.
Part 03
Rank by what enthusiasts value, and muzzle the crowd
A cap of two results per channel was also tried and rejected: it demoted well-tagged festival sets in favour of rougher uploads. Curator prominence is a quality signal, and capping it threw quality away.
Part 04
Let every click argue back
This is how users shape the ranking without being asked: by what they click and what they do not. A person keeps the judgement; the golden set keeps the memory.
Part 05
Never show an empty room
The result count is always shown, even when it is one, so a thin page reads as thin rather than broken. The empty state is the catalog’s most useful form, because it is where it learns what it is missing.
Take-aways, without the tango
- 01Interpret before you retrieve, in one place, and show the interpretation so it can be corrected.
- 02Define the result as the real-world thing, not the file, and never offer the flat view as a preference.
- 03Rank by signals the experts would name, and cap the one signal everyone else optimises for.
- 04Log what was shown and what was clicked from day one, and turn real failures into tests that gate every change.
- 05Never return zero. Tell “nothing exists” from “you filtered it out”, and let the empty page feed the catalog.