BrewMatch and the language of recommendations
Recommendations are a translation problem
BrewMatch is not trying to make someone a coffee expert. It is trying to help someone buy a better bag of beans with more confidence than they had five minutes ago.
That distinction changes the interface.
Most recommendation flows start by asking users to choose from the vocabulary of the seller: origin, process, roast level, cultivar, tasting notes. Those words can be useful, but they often arrive before the user has a mental model for them.
Start with the user's language
Early BrewMatch flows are organized around plain preferences:
| User phrase | Product translation |
|---|---|
| I like chocolatey coffee | lower acidity and deeper roast notes |
| I want something bright | fruit-forward notes and lighter roast range |
| I drink coffee with milk | body and sweetness matter more |
The important part is not hiding expertise. It is sequencing it.
What changed
The first version tried to explain too much. It treated every recommendation as a tiny lesson. That sounded helpful in a document and felt heavy in use.
The current direction is more restrained:
- ask for taste in ordinary language
- show a short set of matches
- explain why each match appears
- let curious users go deeper
The lesson
Recommendation systems feel better when they respect the vocabulary people bring with them. The product can still be intelligent. It just does not need to announce that intelligence at every step.
