Full public record

Employer attribution methodology

Every filing count on this site answers one question: which filings belong to which company? The government does not answer it for us. DOL publishes an employer name as free text, exactly as the filer typed it, and a single company files under dozens of legal entities across years. This document states the rules we use, why each one exists, and what we deliberately exclude.

The rules are maintained as a single machine-readable ruleset that every part of the build consumes. Nothing attributes a filing by any other means.

An employer is attributed to a company when its name matches that company's include pattern and does not match its exclude pattern.

Both patterns are applied to the employer name after normalization: uppercased, with every run of non-alphanumeric characters collapsed to a single space and the result trimmed. Wal-Mart Associates, Inc. becomes WAL MART ASSOCIATES INC. The SQL and Python implementations compute this identically, so a rule that matches in one matches in the other.

Patterns are anchored to the start of the name

Almost every include begins with ^. A company's filings lead with its name; an unrelated business that merely contains the word does not.

This is the single most important rule, and it exists because the earlier unanchored version was wrong in ways that mattered:

Subsidiaries are listed explicitly

A company gets credit for a subsidiary's filings when the subsidiary is wholly owned and its work is the parent's work. These are named one by one rather than inferred, so each is auditable:

Exclusions are recorded, not silent

Where a company shares its prefix with an unrelated business, the exclusion is written down in the rule's exclude field with the reason in notes:

• BlackRock excludes BlackRock Microsystems (neurotechnology), BlackRock IT Solutions, BlackRock Construction Services, BlackRock Resources, BlackRock Automation and BlackRock Energy.
• Micron excludes Micron Products (medical devices), Micron Systems and Micron Solutions.
• General Motors excludes GM Financial and Cruise, which are tracked as separate businesses.
• HCL America excludes HCL Global Systems, an unrelated staffing firm with 2,956 filings that a bare HCL prefix would have absorbed.
• Ernst & Young excludes Ernst Metal Technologies.
• Intel excludes names beginning INTELLI.

Judgment calls we made

Three cases had no obviously correct answer. Our choices:

1. Apple is restricted to Apple Inc. and Apple Payments Services. Entities such as Apple American Group (an Applebee's franchisee) and Apple Processing LLC are excluded despite leading with the word, because they are not Apple Inc.
2. AMD is matched only as Advanced Micro Devices. The bare initialism is too common to be safe, so a small number of genuine AMD entities filing under an AMD prefix are given up in exchange for excluding seven unrelated firms.
3. Adobe is restricted to Adobe Inc. and Adobe Systems, excluding businesses named after the Spanish loanword.

In each case we chose to undercount rather than over-attribute. A number that is slightly low is defensible; a number that credits a company with someone else's filings is not.

DOL publishes some quarterly files as cumulative year-to-date rather than quarter-only. FY2021 Q2 and Q3 both begin at 2020-10-01; FY2023 Q2 begins at 2022-10-01. Loading them naively counts the same case up to four times.

We keep one row per case number per fiscal year, preferring the record with the latest decision date so a case that was certified and later withdrawn reports its final status rather than a stale certification.

Deduplication is scoped within a fiscal year. 102,084 case numbers legitimately appear in two fiscal years, typically certified in one and withdrawn in the next. Those count once in each year rather than being reassigned, because each year's file reflects a real decision in that year.

Two signals on employer pages and rankings come from an automated evaluation model, Jev (typesafe-ai/jev, run through the Vercel AI Gateway). It answers fixed, typed questions; it does not write any text we publish. Every answer is computed once when the data is refreshed, never while you browse.

• Business model. For every employer with 10 or more filings, Jev reads the filer name, its most common NAICS industry code and its five most common job titles, then picks one of: IT staffing and outsourcing, technology product, finance, healthcare, university or research institute, or other. A label is published only when Jev's confidence is 0.75 or higher; below that, the employer gets no label and no peer rank. Peer rank is the employer's position by total filings among employers given the same label.
• Title seniority. Jev scores every job title filed 20 or more times since FY2024 on a 0 to 4 scale: entry level, mid level, senior, lead or principal, manager and above. A title scoring 2 or more counts as senior-titled. Only explicit words count: "senior", "lead", "principal", "manager" and the like. Numbered grades such as Engineer 3 or Analyst IV count as mid-level, because grade ladders differ from company to company. A generic "staff" title (staff pharmacist, staff scientist), "member of technical staff", and a title where "manager" names the job itself (product manager, project manager, account manager) count as mid-level too; managing people does not. The score reflects the title as written: some employers, banks especially, use "Vice President" for roles other firms would call mid-level.
• Senior titles at entry-level pay. Among an employer's certified filings since FY2024 that carry a senior title and a stated prevailing-wage level, the share at Level I, the lowest of the four tiers DOL defines, the one intended for entry-level positions. Shown only with 25 or more such filings, ranked only with 100 or more.

A senior title at Level I is a signal worth a closer look, not proof of a violation: the wage level describes the job's requirements as the employer certified them, and titles are not standardized across companies.

Before publishing we hand-checked random samples and the filings behind the top of the ranking. Title seniority: 40 of 40 consistent with the scale above, after three revisions found by reading the filings behind the top-ranked employers: generic "staff" titles, numbered grades, and job-function "manager" titles had each been counted as senior. Business model: 36 of 40 correct; the misses were a food-service company and a homebuilder labeled finance, a debt collector labeled IT staffing, and a mattress retailer labeled technology product. In the 50 largest IT staffing and outsourcing filers, low-confidence labels had swept in a card issuer, a payments processor, a payroll company, a software vendor and a management consultancy; the 0.75 threshold removes them. One remaining judgment call, Optum Services, a healthcare services company, still carries the label. Treat a peer rank as approximate.

State these before anyone else finds them:

1. USCIS attribution uses a separate pattern list. USCIS publishes employer strings in a different format from DOL, so its pattern list is maintained independently and is not yet held to the anchoring rule above. Unifying the two is outstanding work.
2. Coverage is 87 companies. The underlying dataset holds every H-1B filing from 193,788 employers; the site reports on a curated subset.
3. We attribute by filer name, not by corporate ownership. A subsidiary we have not identified files under its own name and is not credited to its parent. New subsidiaries appear as companies restructure.
4. Employer names contain typos. 1 Tata Consultancy Services Limited and similar malformed entries are not attributed. The effect is single-digit.

Every published figure can be reproduced from the government files without trusting our database: download the DOL disclosure workbooks, apply the normalization, anchoring and deduplication rules stated above, and compare the result to what the site shows.

We do exactly that on every quarterly refresh. An independent verification pass reparses the raw workbooks from scratch — separate from the production build — and the refresh fails if any published number does not reproduce. Structural checks run alongside it, asserting that deduplication holds, that every complete fiscal year has all four quarterly files, that no company drops to zero mid-history, that no two companies claim the same employer, and that published wages are plausible.