I would not trust a restaurant discovery startup just because it says AI-driven insights on the deck. Food apps are brutal. People forgive a slow website. They do not forgive a bad dinner recommendation twice.
Discovery sounds like a model problem. It is usually a data and trust problem wearing a model hat.
What users actually want at 7 p.m.
They want a short answer:
- open now
- fits my budget
- not a 45-minute wait surprise
- matches what I mean by "casual" or "date night"
That is part taste, part logistics, part neighborhood knowledge.
An LLM can write a charming blurb about a bistro. It cannot fix stale hours, dead listings, or reviews from three years ago.
The mistake is ranking on vibes
A lot of teams rank places using scraped reviews, social posts, and embedding similarity.
Looks smart. Fails in the real world.
A spot with great photos and mediocre food climbs. A place locals love but rarely review disappears. A new opening gets hyped before the kitchen finds its rhythm.
I would rather show fewer results with clear reasons than a long list that feels personalized and lies.
Where AI can help without faking magic
Useful angles I have seen:
- summarizing long review text into plain pros and cons
- grouping places by what people actually mention: quiet, loud, fast service, big groups
- spotting outdated menus or hours if you verify against another source
- helping someone describe a craving in normal words
None of that replaces feet on the ground. Someone still has to keep the dataset honest.
Repeat use is the only scoreboard
Downloads are noise. Repeat Friday-night opens are signal.
If locals keep the app, you are doing something right. If tourists install once and delete, your AI story is wallpaper.
Build for one city, one cuisine, one user type first. Win trust in a small radius. Expand after the ranking survives your own friends being mad at a bad pick.
That is slower than a launch post. It is also how discovery products stay alive.
