How arkerah decides what a corporate job is

What counts as a corporate job on arkerah?

We use one test, and it is the same for every advert in every country we cover. The job has to be done at a desk, with a laptop, by someone with a professional or graduate background. And it has to own a piece of the business, not only support one.

The second half matters as much as the first. A receptionist, an administrative clerk and an executive assistant all sit at desks, but they service a function rather than run it, so they are not in scope. We judge the role, never the employer: a marketing manager at a hospital is in scope, and a nurse at a software company is not.

Managing the work is not doing the work. A logistics planner belongs to supply chain. The forklift driver in the same warehouse is not a corporate role, however skilled the job is. The same goes for a maintenance engineering manager and the technician holding the spanner.

Store managers are out of scope on purpose. A store manager runs a shop floor, however senior the title, and arkerah is built for corporate careers, so we leave those adverts out. That is a choice about what arkerah is for, not a gap in how it reads them.

How does arkerah read a job advert?

Every advert passes three steps, in this order. The first is free, the second costs one model call, and the third is a decision taken from the model's answer.

An advert that names no work at all, such as an open application or a talent pool sign-up, is out of scope with the reason no role. One whose function is none is out of scope as not corporate. One whose title names a trade or a shop-floor role is out of scope as a trade title. Everything else is corporate.

Nothing is deleted. An out-of-scope advert keeps its row, its text and its reason, so that a rule we change can be checked against everything it removed.

Why does arkerah use title rules as well as a model?

Because the model is good but not perfect, and its wrong answers are as confident as its right ones. Its instructions already say that hands-on trades are not corporate, and it still gave some of them a business function: an electrician's advert, Elektromonteur Installatietechniek, came back as engineering.

The title rules are a backstop for exactly those cases. They are deliberately narrow, made of words that almost never mean a desk job, in the languages of the five countries we cover.

Running the floor is still the floor, so a store manager or a site foreman is out of scope however senior the title sounds. A bare title such as Teamleader counts as the floor only when LinkedIn files the advert under retail, because the same word names an office team everywhere else.

Some titles stay in on purpose, because the words alone cannot tell. On the same day, Sales Assistant was a shop assistant at one company and a desk job selling insurance by phone at another.

How is each rule measured before it ships?

A rule that removes a job from everyone's list does it without asking, so it ships only with a measured precision figure. We take every advert the rule would remove from what arkerah serves, read each one, and count how many really are out of scope. We report the Wilson 95 per cent lower bound: the precision we can be confident of, not the best case.

We grade distinct pairs of title and company, never raw postings. One recruiter's advert, copied 22 times, once looked like a 4.7 per cent error rate. It was a single mistake, counted twenty-two times.

A rule that misses the bar does not ship. A filter that hid jobs in a language you might not read scored 24 per cent precision and was rejected on that number.

For kinds of role the question is narrower than whether a label is right. Leaving out jobs labelled marketing costs someone who wants finance something only when one of those jobs is really a finance job. So each label is measured against each kind of role you might pick: of the 58 graded jobs labelled marketing, how many were really finance. A label is left out of your list only when, at the 95 per cent lower bound, at least 90 per cent of its jobs are none of the kinds of role you picked. On the 4 October 2026 grading, 99 of the 110 pairs clear that bar. Product is never left out for someone who wants engineering, because 5 of the 56 graded product jobs were really engineering.

The floor words added on 4 October 2026 matched 98 of the 9,710 jobs arkerah served that day, 1.0 per cent. Most of these figures come from a single grading pass, done by an AI model reading each advert against a written rubric, so read them as evidence rather than proof. Every graded row is kept, so anyone can check them.

Do the same rules apply to everyone?

Yes. The corporate decision is taken once for each advert, before any person sees it, and it is the same for a student and for a director. An advert out of scope is in nobody's list, and it is not in Every job here either, because it is not part of what arkerah serves.

What changes from person to person is a second layer: filters that come only from what you told us.

Which filters come from what you told us?

Only facts you stated can narrow your list, and an answer you have not given excludes nothing. If we do not know which languages you read, no advert is hidden for its language. If you never told us your level, no advert is hidden for its level.

• Languages: adverts in a language you did not say you read are left out.
• Level: you see your own level and the one above it.
• Companies and industries you turned down: their jobs are left out until you undo it.
• Sponsorship and nationality: an advert that says it will not sponsor a visa, or that requires a nationality you have not told us you hold, is withheld only when your answer arms the rule. We name the kind of requirement, never a count.
• Kinds of role: once you confirm the kinds of role you want, a job labelled as another kind is left out only when that label was measured to be almost never one of yours, and only for adverts read by the version of the model that was measured. Each one is listed in Why not, where This is my kind of job brings it back to your list.

On a ranked list, the jobs that score lowest for you in each country, the bottom tenth by default, are left out too. That score is measured against your own other scores, never against a bar shared between people.

How does arkerah use your level, and who sees internships?

Onboarding asks your level once. We then show you adverts at that level and one above it, judged by the top of the range the model read from the advert. A junior sees junior and mid roles. A senior job whose range starts at mid is not shown to a junior, because the top of the range is what a junior would have to clear.

Internships go only to people who said they are an intern or a student. A title that names a student role, such as Internship, Werkstudent, Stage, Praktikum, Beca or Alternance, is read as an internship whatever the model read. A traineeship or a graduate programme counts as junior. On 4 October 2026 we read all 56 student titles this took out of juniors' lists, and 56 of 56 were student roles.

If you state a level, an advert the model read but could not place is not shown. An advert we have not read yet is shown. If you never stated a level, nothing is withheld for level at all.

Is anything hidden without telling you?

Not by your own filters. The Why not tab lists the jobs at a place that you are not shown, and names which of your filters stands in the way of each one. Every job here shows every job at the place you are looking at, newest first, with none of your filters, no ranking and no floor.

When a breaker withholds jobs, it says which kind of requirement it was, a nationality or visa sponsorship, and never how many.

The corporate rules are different, because they decide what arkerah serves at all, so a trade or shop-floor advert is not in Every job here either. That is why every one of them is measured before it ships, kept narrow, and recorded with its reason each time it removes an advert.

What are the limits of this approach?

• Before the model, the title gate removes shop-floor titles in all five languages, but its rules for trades, warehouses, cleaning, care, teaching and kitchens are in Dutch, with a few English words. The same rules in French, German and Spanish are written and are being measured on real adverts, and they switch on only once they clear 98 per cent. Until then, such a title goes straight to the model.
• The model reads up to the first 12,000 characters of an advert. A requirement stated only at the end of a very long advert can be missed.
• Some titles are ambiguous and stay in on purpose: Sales Assistant, Client Advisor, Retail Consultant.
• Most precision figures come from a single grading pass, done by an AI model reading each advert against a written rubric. Where a second, independent pass read the same adverts, the two agreed on 76 of 81 function labels (27 September and 4 October 2026) and on 230 of 231 visa adverts (30 September 2026). No person has graded a full sample yet, and each trade and floor figure still rests on one pass.
• arkerah covers five countries today: the Netherlands, Belgium, France, Germany and Spain.

Questions

Read next

• Connect your AI assistant to arkerah
• Privacy