We read 202,073 UK job adverts with an LLM: 1 in 6 from visa-sponsor employers rule out sponsorship

Baris Asa — Mon, 02 Nov 2026 10:19:18 +0000

I run SponsorJobs, a UK job board for people who need a work visa. Every advert we collect from an employer on the Home Office sponsor register is read by a language model when it arrives, and the ones that rule out sponsorship are hidden. That also gives us a running count of how often "licensed sponsor" and "this job will sponsor" disagree.

Over 30 days (3 October to 1 November 2026) we read 202,073 unique adverts. Here's what came out, and a few things I learned about doing this with an LLM.

The numbers

Measure	Share
All adverts that ruled sponsorship out (or required UK nationality, residence, clearance or an existing right to work)	16.1%
Full-text adverts only	19.1%
Same employers we already read in September	23.4% (September: ~22%)
Live adverts that explicitly offer sponsorship	0.9%

For every advert that offers sponsorship, about 40 rule it out. Early-career roles are hit hardest: 48% of internships and placements, 24% of entry-level roles.

How the classifier works

Each advert gets one of four labels: offers sponsorship, might, says nothing, rules it out. A "rules it out" verdict only counts if the matching wording is actually in the text, so the model has to point at evidence rather than guess from the employer or the job title.

We check it by hand: in a random sample of 80 refusals, 91% hold under our rules. If you don't count a bare "you must have the right to work in the UK" or "this role may require security clearance" as a refusal, it's 69%. Of 80 adverts we kept, none was a clear refusal we'd missed.

Lesson 1: snippets lie by omission

Some sources only give us about 500 characters of an advert, and the "no sponsorship" line is usually near the end. Of 1,802 identical jobs we had in both forms, the snippet was ruled out 33 times and the full advert 362 times. If you classify job adverts, find out how much of the text you actually have before trusting the rate.

Lesson 2: coverage changes the headline more than behaviour does

Between September and November we started reading many more employer career sites directly. The headline moved, but on the employers we already covered the rate was flat (22% to 23%). Comparing like with like was the only way to say anything about employers rather than about our own pipeline.

Lesson 3: most "no"s are the rules talking

About half of refusals just say sponsorship isn't available. About a quarter set a condition (security clearance, UK nationality, years of residence), common in defence. About 1 in 5 only say you must already have the right to work. A lot of this reflects UK visa rules since July 2025, not employer whim.

Data

The aggregate data is CC BY 4.0 on GitHub: barisasaa/uk-visa-sponsorship-adverts-2026. The full write-up, method and limits are on the original study page.

Happy to answer questions about the method in the comments.