← Back to the article

The roles behind “88% of Engineering Job Postings Don't Say If They Sponsor Visas. 91% of the Application Forms Ask”

Read on 13 October 2026 from the public Greenhouse, Ashby and Lever job boards. 387 board names were tried and 201 boards answered; where one name answered on two systems the larger board was kept, which left 187 boards and 23,009 postings of every kind. A posting qualified when its title described a software, data, machine learning, security, platform, mobile or infrastructure engineering role, with the same filter as the September count: 4,916 postings. Postings of one company with the same title and the same opening 400 characters are one role.

The main sample is 2,718 roles open in the United States at the 136 companies based there. Where each company is based, with the source, is in companies.json.

The posting text. Every sentence that names visas, sponsorship, H-1B, immigration, work authorization, citizenship, permanent residence, export control or security clearance was kept: 310 distinct sentences. The application form. The questions a candidate answers when applying are public on all three boards: Greenhouse through its job board API, Ashby through the request its public job page makes to load the form, Lever in its public apply page. Every form was read, and every question on the same topics was kept: 535 distinct questions. Only question text, answer options and whether a question is required were kept; nothing was submitted.

Labels. A language model (Claude) labeled each distinct sentence and question under written instructions. 5 labels were changed on review; each change and its reason is in overrides.json and in the tables below. A second run of the same model, which did not see the first, labeled a random 100 of the first extract and every question added when the Ashby and Lever forms were read, 329 items in all. The two runs agree on 329 of 329 for the categories the article counts (offers, refuses, asks); both labels are in agreement-sample.json. No person checked every label, which is why every one is printed.

A role’s stance is everything its postings say about the United States or about no named country: a refusal outranks a transfer-only rule, which outranks case by case, which outranks a plain yes. A form asks about sponsorship if any posting of the role has a question labeled as asking about it, for the US, the job’s location or no country.

The same rows as data: role-stances.json, labels.json, companies.json. Every figure in the article, with the checks behind it, is in results.txt, the unedited output of the analysis script. The H-1B rules as checked on 13 October 2026, with sources: legal-timeline.md. The two scripts that decide the counts are here too: extract.py picks the sentences and questions, analyze.py counts them. Long dashes inside quoted text are printed here as hyphens; the data files keep them.

What the posting says, and what the form asks

Application forms read: 2,718 of 2,718 roles. The form asks about sponsorship in 2,466, 90.7%.

By company

Companies based in the United States, their software engineering roles open in the US, what the postings say, and how many of the forms ask.

Every sentence, with its label

All postings in the sweep, not only the main sample. Labeled other: not about this candidate’s visa or right to work, mostly non-discrimination paragraphs.

Every form question, with its label

Question text, then the answer options after ||. All postings in the sweep.

Every role

The main sample. Sentences and questions are the keys in the two tables above.