The Summer Ramp That Fell Short

On the first Friday of July, the Bureau of Labor Statistics reported that U.S. employers added 57,000 jobs in June, against a consensus expectation closer to 115,000. Inside that number, food services and drinking places — the industry code that covers restaurants and bars — shed 33,000 jobs on a seasonally adjusted basis. The question that followed was the one every summer poses about the sector that does most of its hiring in summer: are restaurants pulling back, and if so, where.

A corpus of online job postings is one of the few instruments that can, in principle, answer "where" at the grain of a neighborhood. alldone.jobs reads such a corpus. What it shows for the spring is narrow and worth stating precisely: across the one applicant-tracking network that supplies almost all of it, the food-service share of newly posted work fell short of its normal seasonal climb — a modest shortfall, bounded on every side, and not a market-wide collapse. This is an account of that shortfall, of the five companies that turn out to sit behind much of the map, and of what a single-source, single-snapshot record can and cannot certify.

What this corpus can see

The scope, stated plainly, because it governs every reading that follows. In the February-through-June 2026 window, the corpus holds about 5.4 million distinct U.S. vacancy-locations — the unit this publication counts, a single opening at a single place, deduplicated against the same opening re-posted or re-scraped (see the methodology note below). It is not a survey of employers and it is not a census of vacancies. It is a record of what employers chose to post online, where, and when.

One fact dominates. Of those in-window vacancy-locations, 97.4% — 5,293,517 of 5,436,629 — come from a single source: DirectEmployers, an applicant-tracking distribution network used by large U.S. employers. Every other source contributes under 1.5%. So the corpus does not observe "the labor market." It observes direct-employer postings from large U.S. employers, and every claim below is scoped to that record, not restated as a claim about restaurants at large. The thin non-DirectEmployers remainder is not a usable control: it grows roughly sixfold across the window as new feeds come online — from about 7,700 to about 49,000 postings a month — and its food-service share rises as it grows. That rise tracks which feeds connected when, not hiring; it is a moving target, not a second opinion, and it is used here only to disclose the shape of the corpus.

A second fact fixes which months can carry a claim. February and March are not load-bearing. Eighty-seven percent of February's in-window postings arrive on a single ingest day, and March is an ingest burn-in dump — a 2.2-million-row backfill as the pipeline came online, nearly triple any other month. Both are computed and drawn, but greyed; the primary window every claim rests on is April 1 through June 28, three post-burn-in months.

The shortfall

The corpus defines "food service" three independent ways in parallel — a semantic classifier over embedded posting titles, a family of title regular expressions, and the federal SOC major-group-35 food-preparation-and-serving flag — and treats a posting as food service when at least two of the three agree. The three were run side by side precisely so that no single definition's quirks could carry a finding alone. An independent re-derivation reproduced the SOC series and the monthly totals to the digit and reproduced the direction of the other two.

Here is the ≥2-of-3 food-service share of all new postings, by month, on the primary window:

Read the primary window across, and every definition declines from April to June: the semantic net by 0.23 points, the regex by 0.23, the SOC flag by 0.40, the two-of-three intersection by 0.35. The three sit at different levels — the semantic classifier the widest, the regex the narrowest — but they agree in shape, and the shape is downward.

A decline is not by itself news in a seasonal category, so it has to be read against the season. Food-service employment normally climbs into the summer. Using the Bureau's published seasonal factors for restaurants and bars (NAICS 722), anchored to this window's April, the norm scales the posting share by a factor of 1.027 from April to June — a rise of about 2.7% of the share's own value, roughly +0.10 point, from a 3.68% April base to an expected 3.78% in June. This is the correction that matters most to state clearly: the seasonal norm is a small expected rise in the share, not the five-point swing that a naive reading of the employment-level factors would suggest. Against that norm the observed share fell short by about +0.71 point in May and +0.45 point in June — a decline where the season called for a slight rise.

Two robustness checks decide whether that shortfall is real or an artifact of which batches happened to land. The first is a batch-day jackknife: drop each of the ten highest-volume ingest days one at a time and recompute. The June share moves by only 0.07 point across all ten drops — an order of magnitude inside its 0.45-point shortfall; May's band, 0.33 point, is less than half its 0.71-point shortfall. Unlike February, whose share swings by nearly three points under the same test, the primary window is stable. The second is direction under all three definitions, which holds. The shortfall is not one ingest day and not one definition.

What the shortfall is not, stated with the finding rather than after it. The size of the April-to-June decline depends on one deduplication choice — counting the same listing at two different cities as two vacancies, the standard convention, rather than collapsing it to one. Under a stricter, content-only deduplication the net decline shrinks to about a third of the figure above, and two of the three definitions — the regex and the semantic — no longer decline at all; only the SOC count still falls. The finding survives that stricter test in direction, as a bound, but not in magnitude. And of the 0.45-point June shortfall, only about 0.10 point is the seasonal expectation; the rest is the observed decline itself, over three months of a corpus that is 97% one network, in a share that partly recovered from May to June. This is a modest, direction-only finding about DirectEmployers postings. It is consistent with a soft spring for restaurant hiring intent in this network. It is not a measurement of a summer hiring collapse, and nothing here should be read as one.

Five companies

The pattern that survives deduplication most cleanly is not the trend — it is concentration. On the primary window, five companies account for 47% of all food-service postings in this corpus: Chipotle, Compass Group, Starbucks, Sodexo, and Elior — two fast-service chains and three of the institutional catering and facilities giants that staff cafeterias, campuses, and hospitals. Remove those five and the national food-service posting share nearly halves, from 3.31% to 1.80% — a drop of 1.51 points, about 46% of the whole signal.

The metros move the same way. Most large metros roughly halve when the five are removed — New York from 3.88% to 1.45%, Los Angeles 5.24% to 2.73%, Chicago 3.78% to 1.68%, Boston 2.88% to 1.15%, Philadelphia 4.12% to 1.98%. The reduction is not uniform: it runs from a bit over 40% in Dallas to nearly two-thirds in New York, and a few metros fall by less — Denver drops only from 2.79% to 1.78%, about a third, not a half. But the direction is everywhere the same. What can look like a regional labor-market signal is, in large part, five corporate posting calendars.

Two cautions travel with that number, and neither shrinks it much. The employer identifier is a source-and-slug string, not a canonical registry, so it can split or merge real companies; the 47% is therefore a lower bound on true concentration, and the exact ranking among the five is less robust than the magnitude. And this is a share of postings within this corpus — direct-employer postings from large U.S. employers — not a share of all U.S. restaurant hiring. What is robust is the shape: about half of this corpus's food-service posting volume is five companies, and it holds under the stricter deduplication too.

The neighborhood grain

The reason to read postings at all is geography, and here the corpus is on firmer ground. Locations parse cleanly from a uniformly structured locale field; the ZIP extracts reliably. Every published cell clears both suppression rules this publication binds itself to: no neighborhood-month with fewer than thirty postings behind it, and none where a single employer is more than half of its food-service postings, is shown. What survives is 2,680 CBSA-months and 11,929 ZIP-months of direct-employer food-service posting share.

What the map shows is where the postings are. What it does not show is a clean change signal. Compute a per-neighborhood shortfall against the same April-anchored seasonal baseline and the distribution is a near-null spike at zero: the median CBSA-month shortfall is exactly zero under every definition, the mean a mild positive tilt of about four-tenths of a point. That is the honest bound on the finding — real and directional at the national level, modest, and close to null once it is cut to the neighborhood grain. The map is a picture of where direct-employer restaurant postings sit, ZIP by ZIP; it is not, on this window, a measurement of where hiring moved.

Methodology

Every number above is drawn from a single corpus snapshot dated July 12, 2026, and reproduces from it; a later revisit would ship as a new article at its own vintage, not as a silent edit here. The full method — flow rather than stock, share of total rather than raw counts, the seasonal norm computed in share space and named on the chart, no absolute comparison to government levels data, no causal language, suppression below thirty postings and above a single-employer majority, the single-source scope limit, and the primary-window and jackknife policy — is set out on the methodology page, which this article is written to be checked against rather than taken on faith.

One part of the method is worth naming here, because it is the reason to trust the rest. An earlier version of this analysis reported that the food-service share was flat across the window — no visible slowdown. An adversarial review of our own work took that headline apart: the flat series had been computed on a deduplication rule that collapsed a few national chains thirty- to sixty-fold and did so unevenly by month, which quietly propped June up; its baseline was the unstable February batch; and its seasonal comparison had mistaken a five-point swing in employment for a five-point swing in posting share. Every number in this article is the rebuilt measure that survived that review — a distinct-vacancy-location flow, an April-anchored primary window, and the seasonal norm carried correctly in share space. We found our own measure's flaws before publishing on them; that the corrected finding is smaller and more hedged than the first draft is the point, not a footnote to it.

A displacement story was also searched for and set aside. Because summer in this sector is where an "automation is replacing entry-level restaurant work" account would first appear, the corpus was tested for it — neighborhoods where food-service postings fell short and automation-role postings grew in the same ZIP across consecutive months. The candidates that cleared the raw thresholds did not clear the bar: the "automation" postings in them were warehouse-robotics, manufacturing, data-center and pharmacy roles, not machines standing in for line cooks, and the classifier behind that layer is not precision-validated. No AI-displacement finding ships from this corpus. The same bar that rejects a weak story is the one a strong one would have to clear.

Limitations

• The shortfall's magnitude depends on the deduplication convention. The April-to-June decline survives in direction under both the distinct-vacancy-location unit used here and a stricter content-only deduplication, but under the stricter unit its magnitude is about a third the size, and only the SOC definition still declines (the regex and semantic reverse to a slight rise). The vacancy-location unit — collapsing identical content at the same location, counting identical content at different locations separately — is the deliberate mid-point between raw rows and content-only clusters, and the headline decline is sensitive to that choice.
• The seasonally-expected rise is small. The CES factor scales the April share by about 2.7% to June — roughly +0.10 point, not five points. So the ~0.45-point June shortfall is dominated by the observed decline, not by the seasonal adjustment; the adjustment only establishes that the season called for a rise rather than a fall.
• Scope is one applicant-tracking network. 97.4% of in-window vacancy-locations are DirectEmployers. Every claim is scoped to direct-employer postings from large U.S. employers, not "the labor market." The non-DirectEmployers panel grows sixfold over the window and cannot arbitrate; its rise tracks feeds onboarding, not hiring.
• Single snapshot, three load-bearing months. The window is reconstructed from one 2026-07-12 snapshot using each posting's declared post date. February (87% one ingest day) and March (burn-in backfill) are greyed context only; claims rest on April–June, and June postings had only about two weeks to accumulate before the snapshot.
• Employer concentration is a lower bound, and rankings are less robust than the magnitude. The five-company figure is a within-corpus share, not a share of U.S. restaurant hiring; the employer identifier is a source-slug, not a canonical registry, so it can split or merge companies. The 47% holds under stricter deduplication (48.8%); the specific ordering of the five is the softer part of the claim.
• Coding coverage is partial. About a quarter of postings carry no occupational code, and that fraction moves across months. For food service specifically the SOC flag is well-measured — food-service titles are about 88% covered and the major-group-35 boolean is right about 95% of the time on resolved cases — and it is used only as one of three definitions, never alone.
• The neighborhood shortfall is near-null. At the CBSA-month grain the shortfall distribution is centered at zero (median zero, mean about +0.4 point). The finding is a national-level, direction-only shortfall; it does not resolve to a clean per-neighborhood change signal, and no cell-level change claim is made.
• The displacement fingerprint is exploratory and non-publishable. The surviving co-location cells rest on an unvalidated automation classifier, and the automation postings in them are warehouse, manufacturing, data-center and pharmacy roles — a category error for a restaurant-automation story. No AI-displacement framing ships.
• ZIP is treated as ZCTA. Postal ZIP codes are used directly as Census ZCTAs; the two do not coincide exactly, a documented approximation in the location design. The ZIP map omits 842 surviving point/PO-box ZIPs that have no Census polygon.