every product pitch in this category opens with the same line. "the ai-powered friend who knows the city." we wrote it ourselves on a deck once. then we sat with it and realized we had no idea what we meant.
"friend" is doing a lot of work in that sentence. friends aren't search engines. they aren't yelp averages. they aren't the median of a thousand reviews. when you text the friend who knows the city, you're not asking for information. you're asking for opinionated, full-time attention — and what comes back is an answer, not a list.
so we tried to enumerate what that friend does that yelp doesn't. it took us eighteen months. here's what we found.
i.they tell you what to skip.
yelp lists 40 places. the friend tells you which 38 to skip. that subtraction is the entire product. the value isn't in surfacing more options — calgary already has more bars than any one person can visit in a year. the value is in a confident no.
our venue cards lead with "skip if" before "perfect for." it scares some people. they'd rather hear the upside. but the friend who knows the city saves you from the wrong room before they ever sell you on the right one. anti-recommendations are the trust contract.
"don't go to that place — you'll have a fine time but it's not for you" is more useful than ten five-star reviews.
ii.they sequence, they don't list.
nobody asks the friend "what's a good restaurant in inglewood." they ask "we're going out tonight, where should we go?" — and the friend answers with three rooms, in order, walking distance from each other. opener, main, close. they don't hand over a search result; they hand over a night.
the order matters more than the rooms. cold garden → fine print is a different night than fine print → cold garden. the first is "two people getting comfortable." the second is "two people winding down from somewhere intense." both are correct sequences for different couples on different evenings. the friend knows which one you are.
iii.they walk the room.
the friend has been there. they know the bartender's name. they know which patio empties at 7 and which one fills at 9. they know that the bathroom line at this place gets brutal after 10 on saturdays. none of that is on yelp. none of it can be derived from reviews, hours, or photos. you have to put your feet in the room.
this is the hard part of the engine to automate, and it's why we walk every venue ourselves. our curators sit at every bar for at least one full evening. they order. they overhear. they leave with notes that look like this:
- cold garden, sunday 4pm: three families on the patio, two dogs, one couple in their 30s on what looks like a third date — relaxed, no phones out, both leaning in. perfect first-date energy.
- cold garden, friday 9pm: wall-to-wall, conversation died at the next table, took fifteen minutes to flag a bartender. do not send first dates here on a friday night.
that's the difference. the second note can't come from data. it has to come from someone who was there.
iv.they know who you are.
the friend doesn't recommend the same room to everyone. they remember that you like wine but hate sweet cocktails. they remember that you brought your last date to the same place and want something different this time. they have memory of you, not just memory of the city.
the engine learns this from your voice intake, your past plans, what you rated up or down, what you skipped without telling us why. over time, the recommendations narrow. that's the goal — not a wider net, but a tighter one.
v.they handle the logistics.
this is the part that surprised us. the friend doesn't say "here's the number, good luck booking it." the friend holds the table. they call the venue, they get you in, they text you the time. you don't open eight tabs and pray.
that's why sloane exists. she's the voice agent that handles the booking layer for venues that aren't on opentable — which is most of the good ones in calgary. you outsource the call, the wait time, the awkward "do you have a reservation for two on saturday at 9?" loop. logistics is part of the friendship. we forgot that for the first year and the product felt thin. we added it and the product clicked.
so why is this hard to automate?
most ai products in this category fail at one of those five things — usually the first three. they list instead of subtract. they recommend instead of sequence. they aggregate instead of walk. they generalize instead of remember. they suggest instead of book.
we got the first iteration wrong. we built a recommendation engine that said "here are the top 10 cocktail bars in calgary by rating." it was useless. the friend who knows the city wouldn't talk like that. nobody asks "what's the highest-rated cocktail bar." they ask "where should i take her tonight." different question, different answer.
so we threw it out and built the format. three rooms. opener, main, close. walkable. sequenced for the night you described, not the average night anyone might want. and we built the curation layer underneath, where every venue card is written by a person who walked the room and wrote down what they saw.
we're still teaching the engine, every week. it's not perfect. but we know what we're aiming at now: not a search engine, not a recommendation feed. the friend you'd text at 6pm if you knew somebody who actually knew the city.
that's it. that's the whole thing.