Activity Recommendations Now Come From a Curator, Not a List

Ask the rest of the internet for things to do in a city and you get a popularity-sorted list, dumped at you with no awareness of the trip it was supposed to fit. Fortrip used to do the same thing. A Tokyo–Kyoto–Osaka plan came back with eighteen Tokyo picks and nothing for the other two cities. A traveler who had already said she wanted slow mornings still got a top-of-list pick for the 4:30 a.m. Tsukiji outer-market walk. The trip was being treated as a search query, not a shape.

The problem was always the same. Popularity, on its own, is what a stranger would recommend to another stranger. It doesn't know that you are spending three days in Kanazawa, not three hours; it doesn't know that the day before is a long shinkansen; it doesn't know that you said, on message four, that crowds wear you out by lunch. A list ranked by popularity is a list ranked for the average tourist, and the average tourist is not a real person. Real itineraries don't survive contact with averages.

So we rebuilt it as a curator. The mechanics break into three moves.

Cross-city search runs the candidate pass over every city in the trip at once, not city by city. A 12-day Spain plan with Madrid, Granada, Sevilla, and Barcelona now produces one ranked pool of candidates with their city attached, so the curator can compare a Granada flamenco tablao against a Barcelona modernisme walk on the same shortlist. Treating each city as its own world is what produced the eighteen-in-Tokyo result. Treating the trip as one decision is what fixes it.

Per-city quotas allocate a budget of picks per city before the curator finalizes anything. A 3-city trip no longer gets 18 picks in one city. The budget for that city closes by the third or fourth pick, and the curator spends what's left elsewhere. Quotas scale with how many nights you're staying in each place, so Kyoto on a 4-night stay earns more slots than Hakone on a 1-night stay.

Review-depth weighting breaks ties. The curator favors signal that has been tested by volume: a 4.3 with 8,000 reviews beats a 4.8 with 12, because a high score on a thin sample is mostly noise. This is the single change that has cleaned up the most surprising junk. Tourist-trap restaurants with stacked early reviews drop out. Long-running neighborhood museums that locals actually use rise.

After all three moves run, a finalize step picks a small curated set — usually six to twelve activities for a week-long trip — and writes a one-line rationale for each. The output is a short slate the assistant can defend, not a list to skim.

A list ranked for the average tourist is gone. What replaces it is ranked for the trip in front of it.