Which AI assistant should you ask about places? The answer depends on what you're asking. PlaceBench, our test of six assistants on 150 everyday questions about real places, splits its questions into five kinds. Here is how each assistant scored on each kind.
Scores by kind of question (out of 100)
| Kind of question | Questions | Dots | Muse | Claude Duet | Instinct | Grok Bot | Ask Maps |
|---|---|---|---|---|---|---|---|
| Finding a place nearby | 54 | 88 | 74 | 76 | 67 | 57 | 33 |
| Place details | 41 | 100 | 99 | 84 | 74 | 89 | 76 |
| Planning a route or a time | 33 | 91 | 73 | 77 | 77 | 83 | 43 |
| Booking and contact details | 11 | 100 | 98 | 93 | 100 | 100 | 64 |
| Sharing with friends | 11 | 86 | 86 | 66 | 98 | 39 | 30 |
| All 150 | 150 | 92 | 83 | 79 | 76 | 73 | 49 |
Same scoring as the main result: 1 point for a right answer, 0.75 for a small error, 0.25 for partly right, 0 for no answer and −1 for a wrong one. We computed the per-kind numbers from the public grades file. The groups with 11 questions are small, so treat those rows as a hint, not a ranking.
What stands out
- "Find me a place nearby" is the hardest kind. Most questions name a day and a time, so the assistant has to check that a place is open then, not just that it exists. Dots scores 88 here. Ask Maps scores 33.
- Place details are easier. Dots and Muse are near perfect on facts about a named place (99 to 100).
- Route and time planning splits the field. Dots scores 91 and Grok Bot 83. Ask Maps scores 43.
- Instinct does best on sharing with friends (98), a kind where Grok Bot (39) and Ask Maps (30) struggle.
Questions in other languages
46 of the 150 questions are written in the city's own language, the way a local would type them. On those, Dots scores 95 and Instinct 93. Grok Bot drops to 55 and Ask Maps to 51.
How the answers were checked
Every question has an answer key built from the place's own website first, then the operator's timetable, then its Google Maps listing, then other public sources. Claude Opus 5.5 graded each answer against its key. OpenAI's gpt-6.1-sol regraded 60 answers blind as a cross-check and made the same right-or-not call on 48. This is one run, from 11 to 14 November 2026, and assistants change.
Data
Questions, keys, sources and every grade, under CC BY 4.0: github.com/voygr-tech/placebench. Full write-up: voygr.tech/placebench.
Top comments (0)