We Published a Metro Ranking From a Skewed Sample. The Full Corpus Inverts Both Findings.

Bars scaled to the largest value, 11,005

Replaced on 2026-08-13. This chart previously showed the 30,000-row sample ranking, which is now known to be inverted; the sample figures are preserved in the correction section below rather than deleted. These are counts over all 55,564 metro-tagged postings in the 246,518-posting active corpus, snapshot 2026-08-14 00:27 UTC. Counts, not percentages, so nothing is scaled to a denominator you cannot see.

The corpus columns are full-corpus counts over the jobs table and sum to exactly 239,806 across all 20 sources. The sample columns are the 30,000-row matcher cache the metro figures on this page are computed over, and sum to exactly 30,000 across all sources present. himalayas is highlighted because it is the sharpest inversion: 43.0% of the live corpus and 3.5% of the sample. Four sources holding 42,584 live postings between them contribute zero rows to the sample. Snapshot 2026-08-13 03:22 UTC.

Source: Seeker job corpus, 30,000 active job postings, snapshot 2026-08-16. seekerscore.com/insights/job-market-metro-concentration-2026

Correction, 2026-08-13

What this page claimed when it was published on 2026-08-12, and what is wrong with it. CLAIM ONE: 'New York carries 2,092 listings, more than Washington DC (696) and the San Francisco Bay Area (693) put together', a ratio of 1.51 to 1. On the full active corpus this inverts: the Bay Area alone carries 11,005 metro-tagged postings against New York's 6,912, and DC plus the Bay Area together are 14,122, so New York is 0.49x that pair rather than 1.51x. CLAIM TWO: 'Columbus at 360 outranks Seattle at 243 and is 2.63 times Austin at 137'. On the full corpus Seattle carries 4,150 and Austin 1,259 against Columbus's 416, so Seattle has roughly ten times Columbus, not less. WHY THEY WERE WRONG: not arithmetic, but coverage. The 30,000-row sample captured 360 of Columbus's 416 postings (86.5%) and 243 of Seattle's 4,150 (5.9%). It under-sampled Seattle by about 17 times and the Bay Area by about 16 times while capturing Columbus almost completely, so the ranking it produced was a ranking of how well each metro happened to be sampled. WHAT ALSO FAILED, AND WAS ITSELF WITHDRAWN ON 2026-08-16: this page said 'we have not decomposed the metro counts by employer, so we cannot tell you how much of Columbus is one bank', then ran that decomposition and reported Columbus as 91.7% JPMorgan Chase across 11 distinct employers. That decomposition is now withdrawn: 90.5% of the Columbus postings we hold arrive through a single feed which is JPMorgan's own careers portal, so the concentration measured our ingestion rather than the city. See the second correction below. The full decomposition is at /insights/one-employer-metro-job-markets-2026. WHAT STANDS: the source-composition table below, the 72.7% no-metro figure, the refusal to publish a remote rate, and every denominator scoping on this page were correct and are unchanged. The failure was in the headline, not the method section, which is the uncomfortable part: this page named its own sampling problem accurately and then led with a conclusion that problem invalidates.

The two results, and why the sample produced them backwards

The sample said New York 2,092, DC 696, Bay Area 693, Columbus 360, Seattle 243, Austin 137. The full corpus says Bay Area 11,005, New York 6,912, Seattle 4,150, DC 3,117, Austin 1,259, Columbus 416. Line the two up per metro and the mechanism is visible: the sample holds 15.9x fewer Bay Area postings than exist, 17.1x fewer Seattle, 9.1x fewer Los Angeles, 4.5x fewer DC and 3.3x fewer New York, but only 1.16x fewer Columbus, and it actually over-represents Houston at 0.81x. The metros that lost most in sampling are exactly the ones carrying startup and technology inventory from greenhouse, ashby and adzuna, three sources that contributed zero rows to the sample. The metros that survived intact are those served by the enterprise applicant-tracking feeds the sample was made of. The original ranking was not a surprising fact about the Midwest. It was a readout of which feeds got sampled.

The ranking as originally stated, with its denominator

Among the 8,193 rows in the sample that resolve to a named metro: New York 2,092 (25.5%), Washington DC 696 (8.5%), San Francisco Bay Area 693 (8.5%), Dallas 559 (6.8%), Philadelphia 371 (4.5%), Chicago 362 (4.4%), Columbus 360 (4.4%), Houston 282 (3.4%), Seattle 243 (3.0%), Los Angeles 229 (2.8%). The top ten metros account for 5,887 of the 8,193 metro-tagged rows, 71.9%. Read against the whole 30,000-row sample rather than the metro-tagged subset, New York is 7.0% and Columbus is 1.2%. Both denominators are correct and they mean different things, which is why we are giving you both instead of picking the flattering one.

Why the sample cannot be read as the market

The metro figures come from a 30,000-row matcher cache, not from the 239,806-posting active corpus, and the cache is not a random draw. Workday is 52.5% of the sample against 15.7% of the corpus. Oracle Cloud is 29.2% against 6.3%. Together those two enterprise applicant-tracking systems are 81.7% of the sample and 22.0% of the corpus. Going the other way, himalayas is 43.0% of the live corpus and 3.5% of the sample, and greenhouse, adzuna, reed and usajobs hold 42,584 live postings between them while contributing zero rows. So this is a picture of where large enterprises running Workday and Oracle post jobs. That is a real and useful population, and it is not the same thing as the labour market.

Second correction, 2026-08-16: the one-employer decomposition is withdrawn

This section previously reported, as a newly run decomposition, that 363 of Columbus's 396 employer-named postings (91.7%) are JPMorgan Chase across 11 distinct employers, and that removing that one employer leaves Columbus with 33. Those counts are arithmetically correct over our corpus and they do not mean what we said. 421 of the 465 active Columbus-tagged postings we hold, 90.5%, arrive through a single source, and that source is jpmc.fa.oraclecloud.com, which is JPMorgan's own careers portal. 418 of the 421 are JPMorgan. We see 11 employers in Columbus because we ingest one bank's careers system plus a handful of broad boards, not because Columbus has 11 employers. The concentration figure and the connector are the same fact stated twice. Withdrawn, along with the equivalent Philadelphia figure. The companion page at /insights/one-employer-metro-job-markets-2026 carries the full correction.

What we are not telling you about remote work

The same sample reports that 3.0% of postings are remote. We are not publishing that as a market figure, because the sampling above makes it meaningless in that direction. himalayas, a remote-focused board, is 43.0% of the live corpus and 3.5% of the sample, while Workday and Oracle Cloud enterprise feeds dominate what got sampled. A remote rate computed on a sample that systematically excludes the remote-heavy source measures the sample, not the world. The only defensible remote-adjacent statement we can make is a supply-side one: 103,070 of the 239,806 active postings in our corpus (43.0%) come from a remote-focused board. That is a fact about our sources, not an estimate of the share of jobs that are remote.

How to use this page now

Not as a metro ranking. The corrected ranking is in the chart above and it is unremarkable: the Bay Area, New York and Seattle lead, in that order, which is roughly what the conventional map already says. The value left in this page is the mechanism. If you are reading any metro or industry ranking, including ours, the question that decides whether it means anything is what population it was computed over and how that population was assembled. This page had an accurate methodology section describing a badly skewed sample and still led with a headline the skew produced. A caveat further down the page did not protect the reader, and it did not protect us either. For the concentration question this page raised but could not answer, see /insights/one-employer-metro-job-markets-2026, which measures how much of each metro belongs to a single employer.

How we measured it, and what we checked

Metro counts come from the 30,000-row matcher cache sample of our corpus, snapshot 2026-08-13 03:22 UTC. We verified the denominator arithmetically rather than trusting a label: the per-source sample counts sum to exactly 30,000, and the per-source corpus counts sum to exactly 239,806, which is how we know which figures on this page are sample-scoped and which are full-corpus. Metros are assigned by matching a posting's location against a fixed list of named metros and their constituent cities, so a city absent from that list cannot appear in this ranking no matter how many jobs it has, and 21,807 of the 30,000 rows (72.7%) matched nothing on it. Those unmatched rows include remote postings, country-level locations and cities outside the list, and we have not decomposed them. Every metro percentage on this page is stated against one of two denominators, the 8,193 metro-tagged rows or the full 30,000-row sample, and never against the 239,806-posting corpus, because it was never computed over it. CORRECTION TRAIL, 2026-08-13: the full-corpus figures added to this page come from a separate snapshot of the jobs table at 2026-08-14 00:27 UTC, 246,518 active rows of which 55,564 (22.5%) carry a metro. That coverage is not missing at random, and knowing which rows are missing matters for reading the corrected chart: metro resolution runs at 73.8% on apple, 69.1% on google and 65.2% on amazon, but 0.4% on reed and exactly 0.0% on himalayas, jobicy and remotive, which are remote-only boards where a posting has no metro to resolve to. himalayas alone is 44.1% of the corpus. So even the corrected ranking describes located, mostly large-employer hiring, and United Kingdom inventory is badly undercounted through reed, which means London's 3,712 is well below its real size. The original sample-scoped figures on this page have been left in place rather than deleted, because a correction that removes the evidence makes the error unauditable.

Frequently asked questions

Related Data

Guides

• Remote Job Search →
• Hidden Job Markets →
• Job Market Trends →
• Job Search Statistics →