Everything playful stands on this page
How every number on City Vibes is made, what it can't tell you, and the mistakes we caught along the way.
What it's built from
68 US metros (metropolitan areas, not city limits: "Dallas" includes Fort Worth; "San Francisco" includes Oakland), monthly from 2015. Locals' talk: every post in each city's main subreddit since 2015 — 9 million+ of them — via the Arctic Shift archive. News: 120 million+ articles from GDELT, matched to metros by coordinates. Hard data: Zillow (rents, home values, neighborhoods), BLS (jobs, wages), Census (migration, income, inequality), Redfin and Indeed (market and hiring speed), NOAA (30-year climate normals), EPA (air quality), Census business patterns (venues), HUD (homelessness), Opportunity Insights (upward mobility), FBI (police-reported crime).
Fresh as of
Verified local changes
The change timeline follows reviewed milestones in housing, transit, jobs, culture and public space. Each card links inspected project, agency, operator or local reporting sources and shows when the milestone happened and when we checked its evidence. An announced opening stays a plan until an actual opening is documented. Possible effects are described as possibilities.
This is a growing, curated collection. An unlisted city or empty filter does not mean nothing changed. A source link working does not prove its project status is current; active projects need editorial review every 90 days. Discussion trends remain a separate view of attention, with no claim that rising talk establishes a physical or economic change.
Where locals go: how we read 9 million posts without quoting one
Places start from the map, not from words. We take every named restaurant, bar, café, venue, museum, park, hotel and neighborhood that OpenStreetMap lists around downtown, then count how many different people on the city's subreddit named each one over the last three years — so one prolific account can't make a place famous. We publish the count and the place's real details (what it is, its neighborhood, where it is), never what anyone wrote.
Matching names is where this goes wrong, so the rules are strict. National chains are left out (a business name found in four or more of our cities); regional institutions stay. A venue can't claim the name of a town or neighborhood. A name made of ordinary words ("My Place") is a phrase first and isn't counted, and a one-word name only counts if it's distinctive to that city's subreddit. A short name inside a bigger phrase doesn't count either: "Mellon" in "Carnegie Mellon" isn't a place. Real places with ordinary names can be missed; we'd rather miss one than count the wrong thing.
The local press counts too. A subreddit is one window on a city, and a small one in some places. So we also read the sitemaps and feeds of about 300 local outlets — alt-weeklies, food sites, city magazines, Eater and The Infatuation city sites, public-radio arts desks — and count how many of their stories name each place over three years. We keep only each story's link, date and headline, never its text, and every place card links to its latest story. For the cities where headlines are too general ("where to find great donuts"), we also read each story once and keep only the names of places it mentions, honoring publishers who opt out of text and data mining. Each outlet was tested with our collector before it was added; outlets that block automated reading are left out.
The scenes are the subcultures, obsessions, industries and fights that define a city right now — the AI boom, a rationalist community, a fight over a highway, a farmers' market under a freeway. Claude (Anthropic's model) researches each city on the web, reading local reporting first, and writes the cards; our code then checks every source loads and names the city, screens out crime, party politics and claims about who lives somewhere, and measures how much locals actually talk about each scene compared with the other 67 cities. Every card links its sources, and the research is refreshed monthly.
On the rise means lasting: a place's share of this year's posters grew at least 1.8 times faster than the city's typical place, by a margin well clear of chance, across at least four separate months, so one big night never qualifies. A place mostly named in posts about crime, tragedy or national politics is left off the page entirely.
Trends (pickleball, food halls, robotaxis, return to office, new apartment towers…) compare the share of a city's posts mentioning each this year with two years earlier, next to the same change across all 68 city subreddits, so a city's pickleball boom is read against the national one. A trend is shown only when enough people are talking about it. Weather works the same way: of the posts about weather, what share are about snow, storms, heat or smoke.
Show the topic keyword lists (crime share, kindred cities)
The written reads
The headline and three paragraphs at the top of each city are written by Claude (Anthropic's model) from an evidence pack holding exactly the numbers shown on that page, plus its reading list. Every sentence cites its evidence: ↗ jumps to the numbers, [n] opens the article. Before anything is published, a checker rejects any number that isn't in the pack, any citation that doesn't resolve, and any mention of crime or politics. When the data updates, a read that quotes last month's figures is withheld until it's rewritten — the page shows its plain fact list instead.
Souvenirs
The pins, stamps, patches and tickets along each city are its real landmarks: notable places near
downtown from Wikidata, ranked by how many of the world's Wikipedias cover them (with bridges and
monuments favored over lecture halls), each drawn as a stylized glyph. Tap one for its Wikipedia summary,
official site and a map. "Local talk" appears only when a vibe tag names that landmark unambiguously.
Sites of tragedy, graves, military bases and contested-history memorials are never souvenirs; the
remaining editorial calls are listed with reasons in the project's
config/landmark_exclude.csv.
Reading lists
For each city we look for the best recent writing — mostly local outlets — on its culture, neighborhoods, cost and future. Every link was opened and checked before it was kept; summaries are in our own words. Outlets that block automated reading are missing rather than worked around. Reading lists exist for 10 cities so far; the rest are coming.
Editorial rules
- No quotes from Reddit. The site publishes aggregates only — how many people named a place, trend shares and counts. No post titles, no links to threads, no usernames.
- No crime, tragedy or national politics in anything we derive from posts. A candidate tag is dropped when more than a third of the posts using it are about crime or politics, judged three ways: keyword lists, subreddit flairs ("News", "Politics", "RIP"…), and a small neural model that compares each post title to prototypes like "political protest" or "a person was killed" versus "delicious food" or "hiking trail". That's how politicians' names and incident-related words stay out without anyone maintaining a list of names.
- Crime appears once, as fear vs reality: how much of a city's news and Reddit talk is about crime, next to what its police department reports. No danger rankings, no neighborhood crime maps.
- No political lean. Politics shows up only as a share of what people talk about.
- Booms aren't good for everyone. "Fastest repricing" neighborhoods are ones where prices race ahead of long-time residents, and we say so.
Definitions
- Typical rent: Zillow Observed Rent Index (all home types, smoothed). Typical home: Zillow Home Value Index, middle third of homes.
- Rent bite: typical rent ÷ the average monthly wage (BLS QCEW). Averages are pulled up by top earners, so a typical paycheck stretches less than this suggests.
- Patio month: normal daily high between . Normals are NOAA's 1991–2020 averages at each city's main weather station.
- Net movers: Census net domestic migration — people who moved in from elsewhere in the US minus those who left — per 1,000 residents.
- Attention: a city's share of all geolocated US city news ÷ its share of population. 1.0 = covered exactly in proportion to its size.
- Cost of living: BEA regional price parities, shown as what a $75,000 salary buys compared with the average US place. Each year is relative to that year's average and history is revised, so we show levels, not changes.
- Typical metro: the median of the 68 metros. Every number on a city page is shown next to it, and where higher or lower is plainly better for someone living there (rent, commute, clean air — not summer heat) a chip says whether the city is better or worse than most. Charts draw the typical metro as a grey line or marker.
- Going out (compare, match, map): Census business counts per resident vs the typical metro for restaurants, bars and breweries, museums and stages; outdoor spots are the named parks, preserves, beaches and viewpoints OpenStreetMap lists within 11 miles of downtown.
- Maps: places and outlines © OpenStreetMap contributors (via Overpass and Nominatim); base map Esri, HERE, Garmin. Place descriptions are from Wikipedia (CC BY-SA).
- Commute: average one-way travel time and how workers get there (Census ACS 1-year). Walkable living: share of residents in block groups the EPA's National Walkability Index rates "most walkable" (2021 vintage, the latest published).
- Upward mobility: Opportunity Insights' measure of where children raised in lower-income (25th percentile) families end up in the national income distribution as adults.
- Kindred cities: "talks like" = most similar subreddit talk mix; "statistical twin" = nearest city on rent, pay, inequality, news mix and size.
- Generative posters: the sun grows with summer heat; peaks get snowcaps where the normal winter low is below freezing; snow flecks and rain streaks follow annual normals; the skyline rises with population. Palettes, typefaces and landscapes are drawn from each city's culture and geography.
Bugs we caught (and why you can trust the rest)
Pittsburgh had no snow. NOAA publishes no snowfall normal for 23 of our 68 weather stations, and adding up "missing" gave "0 inches". Missing now stays missing, on the page and in the posters.
2020 wasn't a year. The Census's 2020 population estimates cover April to July only. Charted next to full years, it made 2020 look like a migration drought. It's excluded.
Austin stopped being looked up overnight. 38 cities' Wikipedia titles were redirect pages ("Tampa, FL" gets 3 views a day; "Tampa, Florida" gets 1,700), so every attention number for them was near zero from the day the redirect took over. All 68 now point at the real article, history re-fetched.
A data feed froze without telling anyone. Redfin's old market file stopped updating in June 2026. We moved to its current feed and now flag the 11 big metros it reports only as their core area.
Baltimore was 12× too famous. Our metro circles overlapped: every article located in Washington DC also counted as Baltimore. Articles now belong only to their nearest metro.
New York had zero news. GDELT has no city-level entry for NYC — only boroughs and landmarks — so matching by name silently dropped it (and Austin, and Kansas City). We match by coordinates.
Everyone was "rising" about the same thing. The first version of rising terms gave every city the World Cup, data centers and a new automated-moderator message. Comparing each city against the national trend fixed it.
Can online talk predict where a city is headed?
The backtest hasn't been run yet.
Honest limitations
Subreddits are not cities: Reddit skews young, male and online, and each sub has its own rules (some delete every "should I move here?" post, which we still count as a question asked). News coverage has its own appetites. Metro numbers hide neighborhood variation. Police data covers only the principal city's department, and national reporting changed in 2021. Correlation in any of this isn't causation; nothing here predicts the future yet — the backtested "boom radar" will publish its accuracy, whichever way it comes out.