The Same Industries, Six Other Cities

The last San Francisco page found that only 0.4 of the 8.3-point payroll gap since February 2020 is the industries the city has; 6.8 points is the same industries doing worse here than elsewhere. This page makes the elsewhere explicit. Two of the six were chosen for their overlap with the city's technology economy, and the other four were chosen so that similar policy and different policy, similar geography and different geography, would each have a case in the set. The page is allowed to end without a clear cause, because a set of seven cannot carry one; what it can carry is which differences line up with the order of recovery and which do not.

• SeattleSeattle-Tacoma-Bellevue: the tech-heavy coastal peer with similar politics that recovered faster.
• AustinAustin-Round Rock-San Marcos: the tech destination under the opposite state policy regime.
• New YorkNew York-Newark-Jersey City: the other dense, expensive, transit-dependent tech-and-finance core.
• BostonBoston-Cambridge-Newton: tech and biotech, expensive, slow: the closest structural twin.
• MiamiMiami-Fort Lauderdale-West Palm Beach: the largest Sunbelt magnet for the firms and people that left.
• PortlandPortland-Vancouver-Hillsboro: the governance twin, with SF-like policy and an SF-like outcome.

Each metro is the Census-defined statistical area, keyed by its code rather than its name, because the names changed in the 2023 delineations. Outcomes are measured on the metro; most policy is set by a city or a state, and the ledger below records the level of each rule and the share of the metro it reaches, since a city rule can describe a small fraction of the population whose recovery is being ranked.

Nine dimensions, each with a stated direction and its own base period, which differ: payroll and rents run from February 2020, population from the April 2020 census base, downtown jobs and establishments from 2019, permits from a five-year average before the pandemic. The ordering is therefore called change since the pandemic began, not recovery from one date. Six dimensions enter the ordering; three are shown as companions of a partner they would otherwise double-count, and where San Francisco's rank differs between a companion and its partner the difference is itself about mechanism.

Rising rents are read as demand returning, not as a good outcome for renters. A metro can rank high on that row and be worse to live in for it; affordability is the housing page's question, and this page reads rents as a sign that people want to be here again.

The headline is a tally rather than a score, because a tally is a thing anyone can recount: San Francisco is last in three of the six ordering dimensions (payroll employment, population, housing permits), sixth in the other three (downtown jobs, business establishments, rents), and in the bottom two on all six. Its single-dimension ranks run from 6 to 7, and under any weighting that treats every dimension as at least neutral its combined rank would fall inside that range, which is the reason the range rather than the mean is the claim.

Every dimension counts once. That is a choice, not a fact, made before any number was seen because no other weighting could be defended. The row below shows what removing each dimension does to the order. The mean rank, with each dimension counted once, is one point inside the range: 6.50 for San Francisco against 1.83 for Miami. Removing any one dimension leaves San Francisco in position 7 of 7; without rents the order runs Austin, Miami, Boston, Seattle, New York, Portland, San Francisco.

Three groups of differences, one row per dimension: policy, economy and environment. The 11 economy rows are measured from the same public records as the dimensions above and are read now, and two more rows are measured from the record rather than coded. The other 21 policy and environment rows were coded from the model's own knowledge as of 2026-06: seven have every cell confirmed by a document and are read, nine are confirmed for some metros and shown with those named, and five are shown greyed as not yet verified. A row is read only once every cell is confirmed, each confirmed cell links to its document, and rows that stay grey are themselves information about what the public record reaches from here.

Each read rests on two yes-or-no facts fixed before any value was entered: whether San Francisco sits at the extreme of the seven at the end a commentator would predict, and whether the other six line up the same way with San Francisco left out. A row with no defensible direction is read as differs or does not differ. Beside each row is the rank correlation across the seven and its exact probability from all 5,040 orderings, printed as a direction and not as a verdict. Across the 11 directed rows read so far, about 0.79 findings would appear by chance if the rows and the order were unrelated; the record shows four (state income tax, top marginal rate; state tax on long-term capital gains, top rate; observed rent, February 2020; typical home value, February 2020).

A read on seven metros is a story; a read on fifty is a test of it. The 49 largest metros with every measure present are ordered here on the four dimensions the record holds for all of them, population, domestic migration, rents and establishments, by the same device as above, and San Francisco stands 49 of 49. Each state-level rule in the ledger is then read across them: the rank correlation with the order, a probability from 10,000 shuffles of the order with a seed per row, and where San Francisco sits among the metros that share its setting. Cleveland left the panel: the 2023 delineation added Ashtabula County to Cleveland, so the 2019 establishments file (code 17460) describes a different geography.

The four wide dimensions are growth measures, and the states with no income tax, right-to-work law and rent-control preemption are largely the same states, which are also the ones people moved to for weather and price; a wide read cannot separate the policy from the region it belongs to, and says so beside every row. A state's fact is the verified value where a document confirms it and the model's coded value otherwise; a wide read is provisional, and not counted, while any panel state's fact is unverified; 4 of 5 reads are.

The closest thing to a control this record has. Alameda, Contra Costa, Marin and San Mateo are the other four counties of San Francisco's metro, and Santa Clara sits next door; all of them live under California's taxes, laws and courts and draw on the same industries, and none of them lives under the city's own rules. On employed residents, February 2020 as a hundred, San Francisco county stands at 89.59 at 2026-07 against 93.46 for the rest of its metro and 97.35 for Santa Clara. On jobs at the employer's address, the city's fell 3.0 percent from 2019 to 2023 while the rest of the metro's rose 1.1 percent.

finding Read as registered: finding if San Francisco county sits at least 2.0 index points below the rest of its metro on employed residents and its jobs change trails the rest by at least 3%; no finding if it is within half of either gap on both; neither between. A gap is evidence that something specific to the city, not the state or the region, is at work; it does not say which rule. Employed residents are counted where people live, so a San Franciscan who kept a job across the county line still counts for the city; jobs are counted where the employer is; the two are different questions and both point the same way.

Payroll growth from 2020-02 to 2026-08 in nine industries, San Francisco against three of the six. Against Austin the city trails in nine of nine, a gap of 26.9 points weighted by the city's own industry mix, of which information carries 5.1 percent; against New York the city trails in six of nine, a gap of 5.0 points weighted by the city's own industry mix, of which information carries 2.8 percent; against Seattle the city trails in six of nine, a gap of 5.2 points weighted by the city's own industry mix, of which information carries 4.3 percent. A shortfall that runs through most industries points at conditions in the city; one carried by information alone would point at the office and remote work.

neither Read as registered: finding if San Francisco trails in at least 7 of nine industries against every one of the three metros; no finding if it trails in fewer than 5 against any of them, or information alone carries more than half of the gap; neither between.

Levels cannot show cause; timing can get closer. For each dated change in a city's rules, the series it was meant to move is taken as that city minus the mean of the other six, and the slope over the twelve months after the date is compared with the slope over the twelve before. A bend after a date is consistent with cause and with coincidence; a series that did not bend is the stronger statement, because it says the rule did not show up where it was meant to. Each event names the date it is anchored to (election, signing, effective, announced) and is never re-anchored. A date is verified when a cached document from an approved host states it; a read on an unverified date is provisional and not counted; 3 of 6 are.

What the measured rows say is narrower than the question and worth stating plainly. The two rows that line up with the order are the price of being here before the pandemic: the metro that was the most expensive to rent and to buy in February 2020 recovered the least, and the cheaper ones recovered more, with the other five falling between in roughly that order. Two policy rows that documents have confirmed line up the same way, and they are the same fact twice: the state's top income tax rate and the state's tax on capital gains sort the seven with San Francisco at the costly end, and across the largest metros the income tax still sorts them once latitude and Census division are held fixed, which is the one policy result on this page that a region does not explain away. Two other confirmed policy rows do not line up: the local income tax, where the city levies none and New York and Portland do, and the minimum wage, where Seattle sets a higher one than San Francisco and recovered better. Remote work, the usual explanation, does not put San Francisco at the extreme of the seven in 2024, and its rise since 2019 does not either, so on this record it is a change the city shared rather than one that singled it out. Four further reads narrow where a cause could sit without naming it, and one of them narrowed the city's own story: San Francisco trails the rest of its own metro on both employed residents and jobs, but so does every one of the seven cities on employed residents, and the cities of Austin and Portland trail theirs by more, so the gap is real and not unusual. The shortfall against Seattle and New York runs through six of nine industries and information carries under a twentieth of it, so it is not a technology story either. The three dated rules that looked like bends in the payroll gap share one upturn, and a turn of that size appears in a substantial share of the placebo months, so the record does not single them out. And the ordering's own verdict depends on how wide a downtown is drawn: neither at the registered radius, a finding at twice it. What San Francisco could do differently is therefore still a question about the city's own rules, and the rows that would answer it are the ones the record reaches only in part: rent regulation, zoning reform, the business tax and the possession law are each confirmed for some of the seven and none for all.

This section is the author's reading rather than a measurement: the rows above are read by rules fixed before their values were entered, and this paragraph is what one reader makes of them, which would change as the grey rows are verified and read.