Commerce discovery and recommendations

A perfect semantic match forty kilometres away is a worse result than a decent one on the next street. Local discovery is one of the few search problems where a hard constraint outranks relevance, and treating distance as just another ranking signal produces results people quietly stop trusting.

Underneath sits the harder question of what to show someone who has not searched for anything yet, which is where collaborative filtering and demand signals do the work.

Geography constrains, it does not compete

Distance is applied as a filter before relevance scoring rather than blended into it. Otherwise a strong semantic match will eventually outrank a nearby one, and the user learns the results cannot be trusted.

Autocomplete needs a fallback path

Suggestions are the most-hit surface in the product and the least tolerant of latency. When the primary path is slow or empty it degrades to a cheaper source rather than returning nothing, because an empty dropdown reads as a broken app.

Notifications are audited or they are spam

Scheduling, delivery status and per-recipient history are recorded, so frequency can be enforced and failures diagnosed. An unaudited notification engine becomes an uninstall driver within a month.