The Employers Your Hiring Data Can't See: A New NBER Paper Finds Job-Postings Data Overstates Labor-Market Concentration Nearly Threefold — and the Missing Firms Are Your Size

Lightcast's database spans over 2.5 billion job postings, 400 million career profiles and more than a hundred government sources across 150-plus countries (Lightcast, 2025). Almost every comp-benchmarking dashboard, talent-intelligence screen and "who else is hiring this role" view your team opens is built on that layer of scraped job postings data.

A new NBER working paper matched near-universal Burning Glass postings against the representative Job Openings and Labor Turnover Survey and found that the average establishment's postings have to be multiplied by 6.2 to match actual job openings — and that postings-based concentration measures overstate how few employers are competing for a given worker by nearly threefold (Dalton, Kahn & Mueller, NBER, 2026).

Two conclusions, not one. The level and cyclicality of skill demand survive the correction intact. The employer-count reading does not. And the reason it doesn't is that the firms missing from the data are the size of yours.

What One Posting Is Actually Worth

The design matters here because it is the first time anyone has had both sides of the ledger. Michael Dalton (BLS), Lisa Kahn and Andreas Mueller linked Burning Glass postings to JOLTS at the establishment level — the same establishments, the same months — which lets them measure the gap between what gets posted online and what vacancies actually exist.

They decompose that gap into three multiplicative pieces:

• The extensive margin — whether an establishment with an opening posts it online at all. This is where almost all of the historical improvement happened, and by the end of the sample it still sits around 3.
• The intensive margin — how many openings one posting represents. Stable at roughly 2 across the period.
• A nuisance share — postings that don't correspond to any JOLTS opening. Steady at about 0.6, meaning roughly 40 percent of postings come from establishment-months with no matching opening in the survey.

Over 2011–2019, the window that covers most published research built on this data, the pooled adjustment factor averages 6.2. It was about 10 in 2011 and fell to roughly 4 by 2019. Online postings got more representative over time — the authors are explicit about that — but they never became a census.

The structural finding is the one to hold onto: Burning Glass captures the distribution of openings well, and the distribution of hiring establishments badly. Large employers post most of the vacancies and are well covered. Most employers, however, are small and mid-sized, and they are not.

The One Reading Job Postings Data Gets Badly Wrong

Labor-market concentration is measured with the Herfindahl-Hirschman Index — the sum of squared employer shares in a local occupation market, where 10,000 means a single employer and lower numbers mean more competition for the same worker. Azar, Marinescu, Steinbaum and Taska used Burning Glass to show that US local labor markets look surprisingly concentrated, with the typical index above 2,500 (Azar et al., Labour Economics, 2020) — work that moved both academic and antitrust debate toward employer market power.

Apply the adjustment factor and the picture changes by more than half (Dalton, Kahn & Mueller, NBER, 2026):

Those are reductions of 63 and 57 percent respectively. The thresholds are the standard ones from the 2010 DOJ/FTC Horizontal Merger Guidelines, where above 2,500 is highly concentrated and 1,500–2,500 moderately so (DOJ & FTC, 2010).

The slope is the cleanest way to state it: a one-unit rise in concentration measured from postings corresponds to only about a one-third-unit rise once composition is corrected. The HHI squares market shares, which makes it acutely sensitive to the long tail of small and mid-sized employers — exactly the tail that doesn't post online. The result holds when establishments under 10 employees are dropped, when markets are weighted by employment, and when market size is held constant by quartile.

A Head of Operations reading "we are one of only a few employers hiring this role in this metro" off a postings-derived dashboard is reading a number that is roughly three times too concentrated.

The Missing Firms Are the 50–250 Band

The size gradient is the part that should land personally, because it is not monotone in the direction most people assume.

The adjustment factor by establishment size runs: about 6.0 for 2–9 employees, 6.3 for 10–49, 7.7 for 50–249, 6.0 for 250–999 and 4.2 for 1,000–4,999, falling to roughly 3.6 at the largest establishments (Dalton, Kahn & Mueller, NBER, 2026). The mid-sized band needs scaling by roughly double what the largest ones do. The worst-represented establishments in the entire size distribution are not the smallest ones. They are the 50-to-250-employee band.

The reason is mechanical rather than mysterious. The very smallest establishments post online least often, but a large share of the postings they do produce don't correspond to JOLTS openings at all, so the two errors partially cancel. Mid-sized establishments get no such offset.

Two other gradients are steeper still and compound the problem:

• Wages. The lowest-paying quintile of establishments carries an adjustment factor of 17.5; the highest-paying, 2.4. A roughly sixfold gap.
• Urbanicity. Non-core areas require a factor of about 15.9 against 4.4 for large central metros — a 3.5-fold gradient that survives controls for local skill levels and economic health.

If you are a 150-FTE employer, paying market rather than top-of-market, outside a major metro, you are sitting at the intersection of all three. You are close to invisible in the dataset your competitors' comp benchmarks are built from — and so are the employers you are actually losing candidates to.

What Survives, and Why That Matters Too

This is not a paper that invalidates postings data, and reading it that way would be its own error.

Because large establishments generate most openings and are well represented, the level and cyclicality of skill demand hold up after reweighting (Dalton, Kahn & Mueller, NBER, 2026). Measured skill requirements fall modestly, not materially. The authors also replicate Hershbein and Kahn's finding that automation-related skill requirements rose after the Great Recession, and it survives the adjustment despite being benchmarked to 2007, when representativeness was changing fastest.

Rankings survive as well. In a rank-rank comparison of markets, the slope between postings-based and adjusted concentration percentiles is 0.9 — and 1.0 when markets are weighted by employment. Postings data tells you reliably which markets are more concentrated than others. It tells you unreliably how concentrated any of them are.

That distinction is the whole operational payload. Trend and ordering: usable. Level: not without correction.

The Honest Limits

Four, stated plainly.

It is a working paper. NBER working papers circulate before peer review, and the authors say so on the cover. The HHI result is robust across the checks they ran, but it has not yet been through a journal.

The window is 2011–2019. Representativeness was still improving through that period, and the adjustment factor had fallen to about 4 by 2019. Nobody has measured what it is now. Post-2020 remote postings, AI-written job ads and aggregator consolidation all plausibly moved it — in unknown directions.

Establishment is not firm. The analysis is establishment-level. If your competitive question is about enterprise employers operating many sites, the mapping to your situation is looser than the headline suggests.

Most workers were never in highly concentrated markets. Once you weight by employment, concentration is substantially lower even before adjustment, consistent with earlier findings on labor market power. The threefold overstatement is real; the thing being overstated was smaller than the unweighted statistics implied.

Also worth naming: the mismatch between postings and survey openings is large in both directions — many openings without postings, many postings without openings. The authors attribute that to measurement error and definitional gaps concentrated in small firms, not to a clean signal with a known bias.

What to Change Before Your Next Comp Band Closes

Four moves, none of which requires buying a different data product.

1. Demote every "employer count" and "concentration" reading to directional. If a dashboard tells you how many employers compete for a role, treat the level as unusable and the ranking as usable. Compare this metro to that metro; don't price off the absolute.
2. Stop letting concentration readings justify a lower band. "There aren't many other employers for this role here" is the inflated conclusion, and it is the one most likely to be quietly funding a below-market offer — and a counter-offer loss you'll book as a candidate-quality problem.
3. Rebuild your real competitor set from local ground truth. Exit-interview destinations, offer-decline reasons and the firms named in your own candidates' histories will surface mid-sized employers no scraped feed contains.
4. Keep using job postings data for what it measures. Which skills are appearing in requirements, and how fast — that is the finding the adjustment left standing. Point the tool at skill demand, not at market structure.

The underlying discipline is knowing which part of a number carries weight and which part is an artefact of how it was collected. That is the same reason Scovai's scoring model separates the half computed arithmetically from the role's stated requirements — skill coverage weighted by criticality, seniority fit, compensation realism — from the half returned as labelled model judgement, rather than publishing a single blended figure. A number you cannot decompose is a number you cannot audit.

Open the last comp band you signed off. Find the sentence that justified it with how few other employers were hiring that role. That sentence is carrying three times more weight than the data behind it can bear.