93.9% of our H-1B records match no employer we can name

We hold 97,662 H-1B sponsorship records, covering fiscal years 2024, 2025 and 2026. We counted them on 2026-09-20. They sit beside the employer records behind our public company pages, and the obvious thing to build from that pair is a sponsor filter: type an employer, find out whether it has sponsored anybody.

It does not work. Only 5,950 of those 97,662 records, 6.1%, can be matched to a company we hold. The other 91,712 match nothing. That is 93.9% of the file sitting unjoined, and scraping harder does not fix it, because a sponsorship record identifies its employer with a name string and a name string is not an identifier.

The 5,950 that did match resolve to 5,813 distinct companies. Of those, 132 carry more than one employer name string. That is 2.3%.

The clusters we can show you

The biggest one is Qualcomm. Four name strings, 3,589 approvals between them, filed as "qualcomm incorporated", "qualcomm technologies inc", "qualcomm atheros inc" and "qualcomm innovation center inc".

Deloitte has three strings and 9,228 approvals. The split is 6,796 under consulting, 1,475 under "and touche", and 957 under tax.

Both figures are real counts from the records we hold, taken on 2026-09-20. What they are evidence of is narrower than it looks.

Our two best examples were not discoveries

Three of Qualcomm's four names, and all three of Deloitte's, were joined by a hand-written alias list. A person wrote that list in advance, because they already knew those companies file under several entities.

Only 42 of the 5,950 matched records were joined that way. Forty-two, out of 97,662. They are almost exactly the famous ones.

Take the alias list away and look at what name normalisation found on its own, and the clusters collapse into companies you have never heard of. The biggest it managed unaided were two firms called Clutch and knit, three name variants each, three and six approvals between them. Not 3,589. Three, and six.

So the honest version of this finding is weaker than the headline anybody would want, and here it is. We cannot measure how common multi-entity filing is. We can only show that it exists in the places somebody already knew to look. Quoting Qualcomm at you is quoting our own lookup table, and a lookup table is not a measurement.

The government's own file has no employer identifier

This is the root cause, and it is stated in plain sight. The USCIS H-1B Employer Data Hub describes its own method in one sentence: "It identifies employers by the last four digits of their tax identification."

The last four digits. Not a company registration number, not anything that survives a rename or a new subsidiary, and not enough to distinguish two employers who happen to share those four digits.

The hub covers fiscal year 2009 through fiscal year 2026 quarter 3. Read on 2026-09-20, it publishes no total record count and no employer count, so you cannot even check the size of your copy against the source. Everyone downstream of that file is matching text, ours included, because the file gives them nothing else to match on.

Where the bigger dataset agrees, and where we part company

h1bdata.info indexes the US Department of Labor's LCA disclosure data and states that it holds "more than 3.8 million records between Apr. 2016 and Jun. 2026". That is a far larger file than ours and it is public. We searched it on 2026-09-20, with the year set to 2026, and got this:

The brand name returns nothing. Three legal entities under that brand return 14,256 records between them.

The tidy lesson would be "search the legal entity, never the brand". Do not learn that, because it is wrong, and the way it is wrong is worse than the original problem.

Change one dropdown on that page, from 2026 to All Years, and run two searches. Same site, same afternoon.

Both examples inverted. The brand search that failed for Qualcomm works once you widen the years, and the brand search that worked for Deloitte returns nothing. We do not know why their index behaves this way and we are not going to guess at somebody else's query layer in public.

What matters is what it does to you. A zero on a sponsor search is not evidence that an employer does not sponsor. It is not even evidence that the string is absent from the file, which is what we assumed before we checked: "qualcomm" returns zero rows in one mode and 16,491 in another, so the zero was a filter artifact both times we saw it. Two settings, two opposite answers, no warning on the page either way.

What to do with a sponsor filter that says no

A sponsor filter returning nothing is telling you one thing, and only one thing: this query, against this file, with these settings, matched no rows. It is not telling you the employer does not sponsor.

Three practical habits follow.

Search the legal entity as well as the brand. Filings carry words a careers page rarely bothers with: "Technologies", "Incorporated", "Services", "USA", "North America", and the name of a company acquired six years ago. Qualcomm needed four tries.

Change the filters before you believe a zero. Widen the year range, then narrow it again, and run "inc" against "inc." while you are there. We got 0 and 16,491 for the same brand on the same site, and the only difference was a dropdown.

Treat a zero as unproven rather than as a no. We cannot tell you which employers this is most likely to bite you on. That would need the prevalence number we have just spent this whole post explaining we do not have.

The question itself will reach you anyway. Sponsorship is a standard question on application forms, a field the software collects rather than something a recruiter asks, so a filter cannot spare you from it. It is a filing question rather than a résumé question, so none of this changes what belongs on the résumé. It changes how much weight to put on a filter before you decide not to apply.

How we counted, and what it cannot tell you

Every number above is a full census of the sponsorship records we hold, not a sample: all 97,662 of them, fiscal years 2024 through 2026, queried on 2026-09-20. The matching works by normalising the employer name string on both sides, plus the hand-written alias list described earlier, which accounts for 42 of the 5,950 matches.

Nothing was excluded. The 91,712 unmatched records stay in the denominator, which is why the headline number is 93.9% and not something flattering.

Here is what the count cannot tell you. An unmatched record might belong to an employer we simply do not hold, or to one we do hold under a name we failed to recognise, and we cannot separate those two cases. That is also why we publish no sponsorship rate for any individual company: 2.3% of matched companies showing multiple names is a floor produced by our matching, not an estimate of how often employers file under multiple entities. The real rate is unmeasured, by us and, as far as the public file goes, by anybody.

For scale, this is a smaller and older file than the posting data we write about most months. The 549,667 live US postings we counted in August 2026 refresh constantly. Sponsorship records arrive in annual batches and describe hiring that already happened.

If you want to check any company yourself, start at the USCIS hub linked above, then repeat the search on h1bdata.info with two or three spellings of the entity name. It takes a minute per company and it is the only way we know to tell a genuine "does not sponsor" apart from a name that never matched.