How we measure ATS usage in Sweden

Job ad data from public job boards and career sites.

Usage is measured from publicly advertised jobs. Employers who do not advertise publicly are invisible here, and systems sold mainly to them are under-represented.

Researched fields are checked per system against the vendor's own material, and the most widely used systems were researched first.

199 systems in the registry · 38 with measured Swedish usage · 17 with researched detail

We can name the system for 2,732 of the 9,796 employers who advertised publicly over 90 days to 11 Aug 2026 - a little over a quarter. Employers who did not advertise during the period are not here at all.

At ad level: 42,577 of 63,096 ads can be tied to a system at level A or B. 10,659 are applied to by email, 815 have no application route, 8,084 lead to something that is not a recruitment system, and 670 have a route we cannot read.

The measurement period is 90 days to 11 Aug 2026. The data runs from 9 Jul 2025 to 11 Aug 2026.

Fetching policy

We fetch at most one page per apply host, and only when the ad data was not enough. We honour robots.txt and Content-Signal per host, identify ourselves with a contact address, never fetch behind a login, and never submit a form.

90 days is the same period our registry of staffing and recruitment firms uses, on the same ad feed. A shorter period swings with the season; a longer one makes the answer out of date.

• Installed base and shares are measured over 90 days.
• The registry can be ordered two ways, and they answer different questions. Employers using it counts how many distinct Swedish employers we observe running a system, so it measures how widely a system was chosen. Job ads through it counts the ads themselves, so it measures how much hiring flows through it and is dominated by a few very active recruiters. Bullhorn sits 21st on the first and 9th on the second, because nine staffing firms post about a thousand ads between them.
• The shorter comparison period is 30 days.
• A lookup on a single employer may reach back 12 months. INVENTED period length.
• An observation older than 120 days loses one confidence level. INVENTED threshold.
• A researched field older than 180 days is marked «may have changed». INVENTED threshold.

Shares are not published below 20 ads or 5 employers, and below 5 employers no installed-base figure is published at all. Both thresholds are ESTIMATED.

Every observation gets a level, and the level decides which sentence the page is allowed to write. The levels are precision thresholds: how often a claim at that level is right.

A - confirmed, threshold 98 % precision

The system is named in the ad's own application route. At most 2 errors in 100 claims, and at least 95 % for each individual system, or that system is moved down to B. Measured in the latest run: 100%.

B - confirmed, threshold 93 % precision

The same kind of sentence as A, but the signal sits one step further from the application point. Up to 7 errors in 100 claims. Measured: 100%.

C - probable, threshold 80 % precision

Up to one error in five. Published as its own marked figure, never counted into the headline numbers, and switched off entirely if the level falls below its threshold. Measured: 100%.

D - weak signal

Weaker signals are not an answer. They are shown so you can judge them yourself. It is never an answer and is counted nowhere.

U - no answer

No application route, application by email, a route that is not a recruitment system, or a route we could not read. Each one is a result, not an empty field.

Eight cases decide whether a figure is true. Each has a fixed rule, and where the rule rests on a threshold we chose ourselves it says INVENTED.

1. H1, a staffing firm advertising for a client. The system belongs to the legal entity named as the employer in the ad, never the end client. The consequence is published rather than hidden: systems sold to staffing and recruitment firms are over-represented in ad counts and under-represented in employer counts, which is why both counting units stand side by side and are never mixed in one sorted list.
2. H2, a system behind the employer's own domain. White-labelling is the norm in Sweden, not the exception. Detection therefore starts from the markers the vendor stamps into the ad itself and only then looks at the host domain. Detection that started from vendor domains would find under a third of the largest vendor's footprint.
3. H3, chains of redirects. More than 5 hops is abandoned and the ad becomes U rather than guessing from the last hop. INVENTED threshold.
4. H4, one employer with two systems. Far rarer than people assume: of 720 employers with several ads in the research data, exactly one had two distinct systems. A second system is published only if it has at least 3 ads and at least 20% of the employer's attributed ads. Both thresholds INVENTED. The rule counts systems, never host domains, because several hosts is common and several systems is not.
5. H5, an employer switching systems. A switch is recorded only once the new system has at least 2 ads within 30 days. One recent ad does not outweigh the history. INVENTED threshold.
6. H6, an ad distribution intermediary. A link through an advertising tool gives U on the question of which system, unless the chain can be followed all the way. The relationship with the intermediary is published as its own fact rather than read as a recruitment system.
7. H7, a route that is not a system. Job platforms, aggregators and generic forms are a positive answer, not a failure. Attributing an aggregator as a recruitment system is the easiest category error available here, so those actors are listed individually.
8. H8, unclear employer identity. The key is the organisation number. 98.8 % of ads carry one, and no organisation number in the data pointed at two names. Where it is missing a normalised name is used, which is exactly why the fallback exists instead of the row being discarded.

Two detection rules are worth naming separately. A signal that requires corroboration gets its stated level only when another named rule also matched the same ad; alone, it drops two steps. And when two systems each give their strongest signal on the same ad, the one closest to the application point wins, but the answer is lowered to C and the ad is written to a review queue, because one of the two claims is wrong and we do not know which.

The only form of claim that is strictly true is this one: the employer's job ads published on Platsbanken between two dates contained apply links to the vendor's system. Every sentence on this site reduces to that.

We never conclude that an employer has no system. Absence of signal is absence of evidence.

The registry never claims that an employer does not use a particular system. We can name the system for well under half of employers, so a claim like that would often be wrong.

Roughly one Swedish job ad in six is applied to by email. There is then no system to read, and that is a finding about the market rather than a fault in the method.

Every section on a system page says where its content came from: «Measured in Swedish job ad data» is measured by us, «From the vendor's own website» is the vendor's own claim and is published as a dated quote, «Checked against the vendor's material» was checked against their material on a stated date, and «Company records and public announcements» comes from company registers and public announcements. Superlatives, customer counts and analyst placements are published only as attributed quotes, never as the registry's own findings.

Two kinds of check, and how to tell them apart

A vendor statement reaches this registry one of two ways, and every one of them says which. Some pages were read by a person, who decided what the sentence meant and what it did not. The rest were read by a script that recorded the vendor's own description of their product, verbatim, with the URL and the date - and nothing else. Where that is what happened, the evidence panel says so in plain words.

Both are the vendor speaking about themselves, so neither is more true than the other. What differs is how much was interpreted. The automated read never produces a compliance credential, a data-residency answer or a size band, because those need someone to decide whether a page actually supports the claim. A footer that mentions ISO 27001 has not shown a certificate, and that is the one inference this registry will not make automatically.

«Nothing published found» means we looked on the vendor's website and in their public security material on the date shown and found nothing. That is not the same as it not existing. Ask the vendor.

A system without these fields has not been researched yet. That says nothing about the system.

No paid placement, no affiliate commission and no sponsored slots. The order on this page follows measured data or the alphabet, and nothing else.

Every published row is a product we could verify exists and is sold, or explicitly is no longer sold.

A row is included when the product can be verified as a recruitment system sold on the Swedish market, or as one explicitly no longer sold. Rows that fail the Swedish market presence test are kept, marked «outside the geographic scope», rather than deleted, and they are never published. The reason is practical: deleting a name is how a registry loses the ability to answer «what happened to ReachMee», which is the most useful question it gets about a dead brand.

Canonical names follow one rule: owner prefixes are stripped. Owners change every two to four years in this category, and a name that encodes the owner rots and takes the address and every inbound link with it.

The selector weighs six dimensions. The weights are ESTIMATED and are published here and on every shortlist, because a weight nobody can inspect is an opinion dressed up as a calculation.

• S122 points
• S218 points
• S324 points
• S412 points
• S510 points
• S614 points

A field we have not researched scores half the points for that dimension, neither full marks nor zero. A system with fewer than 8 of 12 researched fields is not ranked at all.

No vendor has paid to appear here and none can. Talendary sells ATS integrations, which makes this category commercially adjacent for us. No commercial relationship of any kind is an input to this ranking.

If the system you already have is the one that fits your answers, we say so.

Not sure what you need? Run the selector

No measured figure is published until the check has been run and reported. The thresholds are fixed and are never lowered to make a run pass. A run that fails them withholds the figures across the whole site; it never withholds the pages.

• Precision A overall: at least 98 %.
• Precision A per system: at least 95 %.
• Precision B: at least 93 %.
• Precision C: at least 80 %, or C is not shown at all.
• Coverage (recall) overall: at least 60 %, and at most 15 percentage points of spread between systems. That spread is blocking for any ranked or comparative display, because a system that white-labels harder than its competitors would otherwise look smaller than it is.
• Stickprov: minst 400 annonser. INVENTERAT golv.

Latest run, 9 Aug 2026

• Stickprovets storlek: 915 annonser.
• Precision A: 100%.
• Precision B: 100%.
• Precision C: 100%.
• Coverage overall: 100%.
• Criteria not met: none.
• Ranked and comparative claims: not permitted.

Coverage per system

The spread is measured over the systems with at least 20 labelled observations. Below that floor a figure is too uncertain to block a publication with, but it is reported anyway, because a system we detect worse than average looks smaller than it is.

• ReachMee: 100% coverage, 36 labelled. Included in the spread measure.
• Teamtailor: 100% coverage, 34 labelled. Included in the spread measure.
• RecMan: 100% coverage, 31 labelled. Included in the spread measure.
• Higher: 100% coverage, 25 labelled. Included in the spread measure.
• Oracle Recruiting: 100% coverage, 25 labelled. Included in the spread measure.
• Ponty: 100% coverage, 25 labelled. Included in the spread measure.
• Recruto: 100% coverage, 25 labelled. Included in the spread measure.
• SAP SuccessFactors Recruiting: 100% coverage, 25 labelled. Included in the spread measure.
• Talent Recruiter: 100% coverage, 25 labelled. Included in the spread measure.
• Varbi: 100% coverage, 25 labelled. Included in the spread measure.
• Visma Recruit: 100% coverage, 25 labelled. Included in the spread measure.
• Ashby: 100% coverage, 24 labelled. Included in the spread measure.
• Intelliplan: 100% coverage, 24 labelled. Included in the spread measure.
• Visma EasyCruit: 100% coverage, 23 labelled. Included in the spread measure.
• Onecruiter: 100% coverage, 23 labelled. Included in the spread measure.
• Greenhouse: 100% coverage, 21 labelled. Included in the spread measure.
• Workday Recruiting: 100% coverage, 21 labelled. Included in the spread measure.
• SmartRecruiters: 100% coverage, 16 labelled. Too few labels to be part of the spread measure.
• Webcruiter: 100% coverage, 12 labelled. Too few labels to be part of the spread measure.
• Bullhorn: 100% coverage, 9 labelled. Too few labels to be part of the spread measure.
• Zerolime: 100% coverage, 7 labelled. Too few labels to be part of the spread measure.
• Emply People: 100% coverage, 5 labelled. Too few labels to be part of the spread measure.
• Jobvite: 100% coverage, 4 labelled. Too few labels to be part of the spread measure.
• Recruitee: 100% coverage, 4 labelled. Too few labels to be part of the spread measure.
• Personio: 100% coverage, 3 labelled. Too few labels to be part of the spread measure.
• Lever: 100% coverage, 1 labelled. Too few labels to be part of the spread measure.
• softgarden: 100% coverage, 1 labelled. Too few labels to be part of the spread measure.
• Workable: 100% coverage, 1 labelled. Too few labels to be part of the spread measure.
• Jobylon: 100% coverage, 0 labelled. Too few labels to be part of the spread measure.
• Oracle Taleo: 100% coverage, 0 labelled. Too few labels to be part of the spread measure.

Strata in the sample: measured_vendors (725), unattributed_url_present (60), white_labelled_hosts (58), long_tail_vendors (32), email_or_no_route (20), known_hard_cases (20).

What the check does not show

The sample was labelled by an independent automated signal, not by two people with a 20 % overlap as the plan specifies. The difference is not a formality. A single labeller cannot disagree with itself, so inter-labeller disagreement is unmeasured rather than zero, and siffran 0% i datan betyder «not measured», not «complete agreement». A human pass over the same sample is outstanding. That is a real weakness in the check and it belongs here.

Page content and DNS are now run, but once per apply host rather than once per ad. It is the same signal at a far lower cost, because thousands of ads share a few hundred hosts, and it is why coverage rose. The price is that a host which changes system between two runs is judged as it was at the lookup, not vid annonsen.

Ingen precision per system publiceras under 20 labelled observations. INVENTED floor.

We find no recruitment system in the registry that commits in writing to storing all candidate data in Sweden. Several store it within the EU/EEA. That is a result, not a gap in our research.

No system in the registry publishes a commitment to store data in Sweden only, as of 11 Aug 2026. If you choose this we will show systems that store within the EU/EEA and mark the difference.

Corrections go to a person. We do not remove a correctly measured figure on request, but we do fix one that is wrong and say that we have.

An employer who does not want to be included is removed without argument, and we record it so the page is not recreated.

Goes to a person here. We publish nothing automatically.

Weights, thresholds and rules change from time to time. Every change is dated here, with its reason, so a figure that moved can be explained afterwards.