Methodology

We recompute these figures from the source releases listed below. A source link opens the publisher’s data page; it does not identify a single record behind an aggregate. Source units, periods and employer grouping can differ, so the figures require those definitions as well as the files to reproduce them.

Coverage and row counts are read from the database. Loaded means loaded into this dataset, not the publisher’s release date, website deployment date or a manual review date. Row counts have different units across sources and are not comparable coverage scores.

The publisher may replace a download. The local release identifiers below describe the files used for this dataset; byte-for-byte equivalence with the current download has not been rechecked. This inventory applies to data loaded on2026-08-28. It is not a new download or a review date.

Our denial rates come from USCIS petition adjudications. That is a different dataset from the Department of Labor's LCA certification status, which is approved at close to 100% for nearly every employer. The two are often reported interchangeably; they measure different things.

We do not blend petition types. Application categories describe the petitions in the source data. They do not determine which category or outcome applies to an individual. We lead with historical change-of-employer decisions and show each application type separately.

The headline denial figure combines 5 fiscal years, FY2021 through FY2025. One employer's petitions in a single year are often a handful of cases, and a percentage drawn from a handful reads like a probability. Combining years raises the count — but that only helps if the years belong to the same period. Across all employers in our data, the share of change-of-employer petitions denied ran at 9.54% in FY2018, 9.42% in FY2019 and 4.99% in FY2020, and has stayed between 1.28% and 1.99% in every one of the 5 years from FY2021 through FY2025. Averaging the earlier years in with the recent ones would not give a steadier estimate; it would give a number that describes no period at all. So the window covers one period, it stops short of the incomplete current year, and every figure taken from it carries the years it covers — on the page, in the tables, and in the search result summary. Each employer page also lists its own petitions year by year, so the combined figure never stands alone.

The window is not "the last 5 years" computed automatically. A window that rolls forward on its own would, sooner or later, merge two different adjudication periods without anything drawing attention to it. Its boundaries are set deliberately, and each time we reload the data we re-check that the years inside it still resemble one another; if they stop doing so, the refresh stops and the window gets re-decided by a person.

A threshold is a decision about what not to show you. We use 10 of them. Two decide whether a figure is a percentage or a raw count; the rest decide whether a page exists, whether we ask search engines to list it, and how long a list gets before we cut it. Each one is listed below with what it excludes, because the useful question is not which constants we chose but which trade-offs we made.

The first two are different quantities gating different things, so they are configured and described separately — one is about wages, the other about petition outcomes, and merging them into a single “minimum sample” would hide that. The reasoning behind the denial-rate one:a number that a single case can move by ten percentage points should not be called a rate. At 20 decisions one case moves it by five points, which we think a reader can work with — and that is why every rate on this site is printed next to its n, at the same size. Screening out the worst samples is the first job; letting you judge what is left is the second.

Below a threshold we show raw counts rather than a percentage with a warning label attached. A small-sample percentage gets read as a probability no matter what the label says, while “8 approved, 2 denied” carries its own uncertainty.

Wage summaries use medians of certified filings with a reported wage, never averages. A row with one filing shows that single filed value and is not labeled a median. For samples from 2 to 9, the median is withheld; at10 or more we show it with the wage sample size. We exclude filings whose wage is implausible against its own prevailing wage — almost always a unit entered wrong, such as an hourly rate filed as a yearly one. Those records stay in the database, flagged; they are kept out of summary statistics.

A filed wage is what the employer committed to the Department of Labor. It excludes bonuses and equity, and it does not confirm what any individual was paid.A filing is also not a hire — employers file for roles they do not always fill.

The same company files under many spellings. We merge them only by rules that aredeterministic — identical after normalising case and punctuation, identical after dropping a legal suffix, or a name that literally contains the other as its “doing business as” prefix. We do not merge by similarity scores or edit distance.

The reason is asymmetric cost: wrongly merging two companies puts two different pay distributions on one page, which is not fixable by a reader; failing to merge only costs an extra page. Where we know from public fact that two filing entities are one company, that judgement is recorded by hand, with its evidence, rather than inferred.

Job titles work the same way, and the limit is visible: 28.6% of certified filings currently reduce to one of our 29 role pages. The rest carry titles we cannot group without guessing, so we leave them ungrouped rather than force them somewhere. They remain visible on their employer's page.

We also never flatten seniority. Staff and non-staff filings for the same role differ by tens of percent; a single blended figure overstates one end and understates the other.

Cities are grouped by the worksite on the filing, not the employer's address, and get a page at 50 certified filings (2,578 qualify; 2,298 of those also clear the content threshold and are listed in the city index). Where we cannot confidently place a city inside one federal statistical area, we show the city's own figures and make no comparison to a wider area — an area average printed under a city heading is two different populations in one number.

WARN coverage and employer matching are separate limits. WARN records are published state by state, and the loaded date ranges are listed above. “Not covered” means we have not loaded the relevant state’s records. “No matching records” means a source is loaded but no observations have been matched to that employer. Neither state establishes that no notices were filed or that no layoffs occurred. Employer pages identify loaded and missing worksite states; matching may be incomplete even in a loaded state.

The most recent fiscal year is incomplete in the loaded LCA, USCIS and PERM releases.WARN coverage uses the dates of loaded observations, not a fiscal-quarter release claim. Where we know how far into the year the data runs, we say so; where the publisher gives us no way to tell, we say “partial year” rather than invent a quarter. A partial year always reads as a decline if it is not labelled.

Nothing here is legal or immigration advice, and none of it predicts what will happen to your petition. A high denial rate does not mean an employer will not sponsor you; a low one does not mean your petition will be approved.