Job Opportunities API

Check the data. Then trust it.

Job Opportunities API

Changelog

What changed, when, in plain language. Most recent first. This is a record, not marketing copy. See /facts for what the ledger actually holds today.

2026-11-16

  • A job you were already charged for is not charged again for 24 hours. On GET /v1/jobs, a job that was charged to your key in the last 24 hours now costs 0 records when it comes back, whether in a later page, a new search or an overlapping since window. The new header X-JOA-Records-Repeated says how many rows of a response were free for this reason. /v1/changes is unchanged: each event is one record.

2026-11-16

  • Terms updated (16 November 2026): plans belong to your account. Every key of your account runs on its plan; a new customer's first key and its secret are shown once payment completes; cancelling moves the account, with all its keys, to the free Explore plan. The wording in force before is kept at /terms/archive/2026-11-16.
  • Privacy updated (16 November 2026): one sentence under What we collect on /privacy. If you start a checkout but do not pay, the account created for it is deleted after 30 days. The wording in force before is kept at /privacy/archive/2026-11-16.

2026-11-15

  • Work visa text on jobs. Every job now carries work_visa_text: what the advert itself says about visa sponsorship, work permits, relocation and the right to work, as statements with a code and a quote copied word for word from the employer's advert, and an evidence_level (stated_open, stated_closed, no_information or not_assessed). Read so far: jobs in Germany, France and Italy; every other job reads not_assessed. New filter visa_text=stated_open or visa_text=stated_closed on /v1/jobs, /v1/jobs/similar, /v1/jobs/closed and /v1/export. It describes the advert: not advice and not a decision on any person's case. A job whose work_visa_text carries a statement carries the legal note immigration_information. Metering unchanged.

2026-11-14

  • Terms updated (14 November 2026): new sentence under Money on /terms. If you subscribe again while a cancelled subscription is still running, the new subscription becomes your account's plan at once and the cancelled one is not charged again. The wording in force before is kept at /terms/archive/2026-11-14.

2026-11-13

  • Country corrections reach a country filtered copy. When we correct the country of a job (for example a career site job we had in Germany that is in India), a /v1/changes poll filtered by the old country, with country= or a key country lock, now receives the job once as withdrawn with correction: true, carrying only job.id and job.status. It is not billed. It is sent once, at the change, and not again when the job changes or closes later. A poll filtered by the new country receives the job as updated with correction: true, also free.
  • New withdrawn_reason value country_corrected on /v1/jobs/withdrawn. With country or a key country lock, a job whose country changed away from your filter (most often our correction) is listed once, at the time of the change, even when the job is still served under its new country. Without country and without a lock this value does not appear. Clients that switch on withdrawn_reason should treat an unknown value as "drop the id", as for every other reason. Free, like the whole feed.

2026-11-12

  • Terms updated (12 November 2026): new sentence on the November 2026 Black Friday deal on /terms (/terms#the-api). Until 23:59 Athens time on 30 November 2026 a new Starter plan costs EUR 14.50 a month billed monthly instead of EUR 29, or EUR 99 a year billed annually instead of EUR 261, net of VAT, for as long as that subscription runs; the sentence states when the reduced price ends, that on an annual Starter plan it applies instead of the November 2026 yearly offer, and what it cannot be combined with. The wording in force before is kept at /terms/archive/2026-11-12.
  • Black Friday deal, billed yearly, until 30 November 2026 23:59 Athens time: Starter annual at €99 a year instead of €261, every year for as long as you keep the plan (code BLACKFRIDAY99, applied automatically at checkout). It takes the place of the first year offer for Starter; Growth, Signal and Scale keep theirs. GET /public/launch-offer carries it as black_friday.annual; the monthly fields of black_friday are unchanged.
  • The 402 record_quota_exhausted answer names the plan that fits. It now carries an upgrade object: upgrade.demand (the records your account asked for in the last 30 days, refused calls included) and upgrade.options (the smallest plan that covers that demand first, then the next plan up, each with its allowance, price as on /pricing, price_text, vat and its own checkout_url). upgrade_plan and upgrade_url keep their meaning and name the first option, so for most free keys upgrade_plan is now starter (it was growth). Every other field is unchanged; the message is now one fact per line.
  • Black Friday deal, until 30 November 2026 23:59 Athens time: Starter monthly at half price, every month for as long as you keep the plan (code BLACKFRIDAY, applied automatically at checkout). GET /public/launch-offer carries it as black_friday while it runs; on a 402 the Starter option carries it as offer. A key whose Starter subscription started on the deal changes plan through support (409 black_friday_switch_by_hand on an in-place change; the 402 then carries upgrade.by_hand: true and no upgrade_url).
  • The 402 no longer carries the one time 25 per cent upgrade offer: the top level offer object (kind, percent, text, never documented) and its sentence are gone, and the offer is not applied at checkout. The Black Friday deal takes its place.

2026-11-11

  • Salaries: for jobs in the United Kingdom, Ireland, the Netherlands and Denmark, a salary read from the advert page whose amount and period do not fit together is withheld (for example a monthly salary that the page's structured data labels per week, or an hourly rate labelled per month). Where the advert states the unit next to the amount, we use that unit. The test uses each country's statutory minimum wage where there is one (Denmark has none) and fixed ceilings. A salary that reaches JOA another way (the source's own job feed, or our reading of the advert text) is not yet checked this way. The job keeps every other field; the change reaches /v1/changes as a free correction (correction: true).
  • /v1/changes: an updated event is billed only when the employer changed the job (title, description, apply link, employment type, location text, or a remote type or salary they state). Changes from our own processing (company name and slug, structured location, category, seniority, a salary we parsed or normalised, posting date fixes, a description we fetched for a job that came without one or with a short stub) are still delivered as updated, now with correction: true, and cost no records.
  • Terms updated (11 November 2026): new paragraph "Choisir le service public and Emploi-Territorial" on /terms (/terms#french-public-service). Jobs from these two French public service job sites are re-used under the Licence Ouverte 2.0, with the source named, a link to the original offer page and the time we last read it; personal mailboxes and mobile numbers in their adverts are replaced with a note pointing to the original advert. The wording before it is kept at /terms/archive/2026-11-11-nhs.
  • New source: Choisir le service public (choisirleservicepublic.gouv.fr), the job board of the French public service run by the DGAFP, with offers of the State, hospital and territorial public services; provider choisirleservicepublic, source type public_agency. JOA reads the site's own offer lists and one offer page per new or changed offer, at most one request every 4.5 seconds. Offers the site relays from Emploi-Territorial are read at Emploi-Territorial. Each row links to its offer page on choisirleservicepublic.gouv.fr, or to the employer's application link the offer states, and carries attribution "Choisir le service public"; the job page shows "Source: Choisir le service public". The rows are re-used under the Licence Ouverte 2.0 (Etalab), which the site states for its texts: keep the credit, the link to the offer and the date of last update (last_verified_at) when you display them. Personal mailboxes and mobile numbers of contact persons are replaced by "[contact on the original advert]", as for Emploi-Territorial.
  • New source: Emploi-Territorial (www.emploi-territorial.fr), the job portal of the French territorial civil service (communes, departments, regions and their public bodies), run by the GIP Informatique des Centres de Gestion; provider emploiterritorial, source type public_agency. JOA reads the site's own RSS feed and one offer page per new or changed offer, at most one request every 4.5 seconds. Each row links to its offer page on www.emploi-territorial.fr, or to the application link the offer states, and carries attribution "Emploi-Territorial"; the job page shows "Source: Emploi-Territorial". The rows are re-used under the Licence Ouverte 2.0 (Etalab): keep the credit, the link to the offer and the date of last update (last_verified_at, the last time JOA read the offer at its source) when you display them. Personal mailboxes and mobile numbers of contact persons written in an advert are replaced by "[contact on the original advert]"; functional mailboxes such as recrutement@ and switchboard numbers are kept.
  • /public/providers names the open licence of a source next to its credit: new licence_name and licence_url, present only for a source that publishes under one (NHS Jobs, Teaching Vacancies: Open Government Licence v3.0; Emploi-Territorial and Choisir le service public: Licence Ouverte 2.0). The coverage page links the licence next to the credit.
  • Sponsor lookup: a name sent with its legal form now matches only the register entry that prints that name. Group, Holdings, International, UK and country words count, and plc is not Ltd: "Rolls-Royce Holdings plc" no longer matches "Rolls-Royce plc", and "Barclays Bank UK PLC" with town London no longer matches "Barclays Bank PLC". An entry that is only close is listed in possible_names with a hint, never matched, with or without a town. Two register entries are never merged into one hit.
  • Sponsor lookup: the exact printed legal name now always matches, also when the register has entries close to it ("ASML Netherlands B.V.", "Adyen N.V.", "Novo Nordisk A/S" and "KPMG LLP" used to be ambiguous). match_type has a second value: exact_legal_name when the name you sent is the printed legal name word for word, exact_normalised when it is equal once legal forms and words such as Group or UK are set aside (a name sent without its legal form). Accept the first without a check; show the legal_name or check the second.
  • Sponsor lookup: not_found now means that no register entry has the name and none contains it. A name that a register entry contains, or that differs from an entry only by a legal form or a word such as Group, is now ambiguous, with the entry in possible_names ("Monzo" used to be not_found with "Monzo Bank Ltd" in possible_names). hits can be empty on ambiguous. New suggested_name: when exactly one register name is, or contains, the name you sent without a legal form, it is given as a lead. Check it, then send it as the name in a second call. It is never a match and is never set for a close entry of a name sent with its legal form.
  • Sponsor lookup: one answer in every country. A franchise brand is excluded with or without a town ("McDonald's" was matched to one outlet on the Canadian list); every company of a staffing group is excluded alike; "J Sainsbury Public Limited Company" is "J Sainsbury plc"; a single word sent with no country is no longer matched through the Danish list alone; a name without a legal form that a register prints in several towns is one hit per town. possible_names now finds possessives ("Sainsbury's"), web names ("ASOS" lists "ASOS.com") and 3 letter names.
  • Sponsor lookup: a town you send is never overruled, and a generic name is not matched by its name alone. When the register prints your words as a whole name and also as other entries ("Red Lion": "THE RED LION" in Machynlleth and "RED LION (JUNCTION 16) LIMITED" in Daventry), or prints a name you sent with a legal form only without one while a company with a legal form of its own is beside it ("Phone Zone Ltd"), every one is a candidate: the town picks one or the answer is ambiguous. A word you wrote, such as Group, no longer chooses between entries that all print it ("Este Medical Group").
  • Sponsor lookup and jobs: every statement now names the register entry ("Rolls-Royce plc" appears on the official register ...). On jobs, a group_company decided by the name alone (the employer's name carries another country or legal form) now reads: A company with the same name in the United Kingdom, "Merz Aesthetics UK Limited", appears on the official register ... It may not be the employer named on this job. The statements for a posting that names another legal entity of the group, a site entity and a curated link are unchanged.
  • Sponsor lookup, fixed: sending a region for an employer outside the US (for example a UK employer) answered HTTP 500 for the whole call. A region is now accepted for any country; only the US registers print one, so it is ignored elsewhere.
  • Sponsor lookup: one bad employer no longer fails a POST. Every employer is checked on its own. One that cannot be looked up as sent (no name, a country that is not two letters, an unknown field) comes back in its place with the new status: invalid, the error code in reason and the problem in hint, while the others are looked up. Only a call in which no employer can be looked up is still a 422. The new status: error (reason: lookup_failed) means JOA could not look up that one employer this time. Neither is charged. An unknown field is now named in the error.
  • Visa information on jobs. Every job now carries visa_checks. For a job in the United Kingdom, Ireland, the Netherlands or Denmark it holds two factual checks of the advert against the official published rules of the country's main work route: whether the occupation is on the official list (occupation_check) and how the advertised pay compares with the official thresholds (salary_check), each with a plain statement, conditions and its sources and dates. These are facts and sources, not advice and not a decision on any person's case. [] for other countries. Jobs with checks carry the immigration_information legal note. sponsor_registers, filters and metering are unchanged.
  • Full boolean search on q, title, title_exclude and description_contains: AND (a space, AND or &), OR (OR or |), NOT (-word, !word or NOT word), parentheses, phrases in double or single quotes, prefix with * or :*, and followed-by with <-> and <N>. Precedence: parentheses, then NOT, then followed-by, then AND, then OR, so a | b & c is a | (b & c). Expressions written in the Fantastic.jobs notation, such as (react | vue) & (node | golang) & !aws, work as written.
  • No more automatic prefix. Until today a bare word of five or more letters inside a query that used any syntax also matched longer words, so react OR vue found "Reactor" and "Reactive". A bare word is now always the exact word; write react* for a prefix. Values that use no syntax are unchanged.
  • New company_name_contains: the same boolean search on the employer name, for example company_name_contains=!(Amazon | Meta).
  • Fantastic.jobs parameter names are accepted on /v1/jobs, /v1/jobs/closed, /v1/jobs/similar and /v1/export: title_advanced (and advanced_title_filter, title_filter) as title, description_advanced (and advanced_description_filter, description_filter) as description_contains, organization_advanced (and advanced_organization_filter) as company_name_contains. location_advanced is refused with a 400 that names country, city, state and region. A new guide lists every difference side by side.
  • Limits, refused before anything is read or charged: a boolean expression may hold up to 32 terms and 8 levels of parentheses (422 query_too_complex); plain values have no such limit, as before. Every text filter may now be up to 500 bytes (was 200). A single word with a dangling | or & (Philadelphia |) is still the plain search of that word. A malformed expression is a 422 query_syntax naming the character position.
  • Boolean queries are faster: the expression is resolved once per request and handed to the search index as a constant. A title query of three groups with an exclusion went from about 6 seconds to well under a second.
  • Sponsor lookup: a name that carries a country or a legal form no longer matches a register entry of another country or legal form. "AstraZeneca UK Limited" with country NL used to match "AstraZeneca B.V."; it is now not found, with that entry in possible_names and a hint saying why. Forms one country uses for the same kind of company still match each other (Ltd, Limited and plc in the UK; Inc and Corp in the US).
  • Sponsor lookup: names of 2 and 3 letters are looked up. "GSK plc" and "ARM Limited" now match; a bare short name such as "IBM" is ambiguous until you add the town or send the full legal name, and a short name that other companies also carry is never a pick.
  • Sponsor lookup: possible_names now finds your name as whole words anywhere in a register name ("Arup" lists "Ove Arup and Partners International Ltd", "Sainsbury" lists "J Sainsbury Plc"), closest first. Still suggestions, never matches. New hint field: one plain sentence on what to send to get a single answer, such as adding the town.
  • Jobs: each sponsor_registers entry now has match_scope and statement. exact_entity means the register entry fits the employer named on the job. group_company means the name is the same but the employer's name carries another country or legal form, such as "Merz Aesthetics GmbH" and the UK register's "Merz Aesthetics UK Limited": a company of the same group appears on the register, not the employer itself. 164 of about 12,300 entries are group_company. Nothing was removed, and the sponsor_register filter is unchanged.
  • Sponsor lookup: English names of Dutch universities, university medical centres, research institutes and public bodies ("Eindhoven University of Technology", "University of Amsterdam", "Erasmus MC") now find the Dutch name the IND register prints, from a curated table, only when that is exactly one organisation on the register. Such hits say matched_on: alias.
  • Jobs: sponsor_registers is more precise. A register entry whose name carries Holding(s) or Group no longer labels a company without those words; when several organisations share a name, only the one in the company's towns is used, and London, New York and other big cities do not decide alone; one-word names, schools, hospitals and other local names need a second piece of evidence or a town; postings by a franchisee, by a staffing firm working at the company's site, or by another employer that shares the company's name now read []; a posting that names another legal entity of the group (a Workday "Company:" field) or a site entity for another town reads group_company. About 14% fewer register entries are shown, measured precision rose from 92% to 99% on a fresh sample of 100. These changes are not signalled in /v1/jobs/changes: refetch the jobs you keep.
  • Legal note on immigration information. Every job now carries legal_note_code and legal_note: the code and text of the legal note that goes with the row, both null when none applies. Jobs with a non-empty sponsor_registers carry immigration_information, and every /v1/sponsors/lookup answer carries the same note at the top level. The note says the information is not immigration advice, legal advice or any other service to candidates, may be incomplete, inaccurate or out of date, and that a qualified, certified professional should be contacted first. /v1/help lists every note (legal_notes).
  • Terms and Privacy updated (11 November 2026): new section "Immigration and sponsor information" on /terms and "Information we publish about employers" on /privacy. Information about visa sponsor registers, salary thresholds or similar rules is information only, not immigration or legal advice and not a service to candidates, and may be incomplete, inaccurate or out of date. If you show it to jobseekers or other end users, you must show the legal note supplied with it (or a notice with the same meaning), unless you are an authorised immigration adviser presenting it within your own regulated service; if you show it without the note, you do so on your own responsibility, you indemnify JOA against claims arising from that use and, where the law that applies to you does not allow that in full, you reimburse JOA's reasonable legal costs to the extent that law allows. New final section "Governing law and courts" on /terms: Greek law, and the courts of Thessaloniki have exclusive jurisdiction, without taking away a consumer's mandatory home-country protection or right to sue there. The previous versions are kept unchanged at /terms/archive/2026-11-09 and /privacy/archive/2026-11-01.
  • The API's own contract now matches the API, and a few answers are more honest. /v1/help is rewritten from the code: the seniority values are Intern, Entry, Mid, Senior, Lead, Manager, Director, Executive, not_stated (junior was never accepted), remote is remote, hybrid, on_site, not_stated, the six filters sponsor_register, country_match, agency, engagement, contract_category and working_schedule are listed, /public/jobs, /public/jobs/{id}, /public/companies and /public/companies/{slug} are described as the website's own routes (no listing rows for anyone else), the error table lists every code a public endpoint answers (filter_too_long at 500 bytes, query_failed as a 500; missing_key, plan_required, record_allowance_spent, country_locked and bad_timestamp are gone because nothing answers them), and the MCP section names all eight tools. A test now fails the build when help, openapi.json and the code disagree. openapi.json carries info.version 2026-11-11 and an operationId on every operation. /v1/me states the limits of the plan actually enforced (it could show the key row's plan instead). A usable key on one of those four /public routes now gets 403 listings_need_v1 (401 before; an unusable key still gets 401, same code and message). A value the database cannot use (an impossible date, number, id or cursor) is a 422 bad_parameter_value, or 422 bad_cursor for a /v1/companies cursor, where it used to be a 500 query_failed; a time with year 0000 or a UTC offset of 16 hours or more is refused by its own parameter code. /v1/me/export-estimate adds remaining_unlimited (true when remaining_this_month is -1). Personal contact scrubbing also catches mailboxes made of one name of the contact named next to them. The 8 level parentheses cap now holds on every text value, plain ones too (the 32 term and prefix limits stay boolean only); literal parentheses such as Nurse (RN) (Night) are fine. GET /v1/jobs refuses include_closed with 400 unsupported_parameter instead of ignoring it: use GET /v1/jobs/closed, or status=any. GET /v1/jobs/{id} also puts the advert text inside data as data.description; the description beside data stays, with the same value, for compatibility.
  • Saved search alerts (beta) now cover jobs posted in the last 30 days, counted in whole UTC days. A job JOA finds late keeps the board's posting date and is no longer mailed as new when the board posted it more than 30 days ago; a job without a posting date is mailed as before. The digest says so.
  • New source: NHS Jobs, the recruitment service of the NHS in England and Wales (www.jobs.nhs.uk, run by the NHS Business Services Authority), provider nhs_jobs, source type public_agency. JOA reads the official NHS Jobs vacancy feed, which NHS Jobs offers to external recruitment platforms, in full every 6 hours: every live external advert of NHS trusts, health boards and the other employers that recruit through NHS Jobs. Each row links to its advert on www.jobs.nhs.uk, where the application is made, and carries attribution "NHS Jobs"; the job page shows "Source: NHS Jobs". The rows contain public sector information licensed under the Open Government Licence v3.0: keep the credit and the link and acknowledge the licence when you display them (see the terms). The description of these rows is the short summary the feed publishes (about 150 characters, cut by NHS Jobs), followed by the salary text, not the full advert; the job page labels it a summary and links to the full advert, and carries no JobPosting markup for these rows. Voluntary and honorary posts and recruitment agencies are not published. An advert is removed when two complete reads of the feed in a row no longer list it, or at once when its closing date has passed.
  • Legal note on jobs from NHS Jobs: every job from NHS Jobs (source nhs_jobs) carries the legal note nhs_jobs_source: "These jobs come from NHS Jobs under the Open Government Licence v3.0. Review the Terms, section NHS Jobs: https://jobopportunitiesapi.org/terms#nhs-jobs". New on every job: legal_notes, the list of every legal note that applies to the row as {code, text}, [] when none applies. legal_note_code and legal_note keep one note, the first in order (immigration_information, then nhs_jobs_source), so their value did not change for any row that carried a note before; a job from NHS Jobs whose employer is on a sponsor register carries immigration_information there and both notes in legal_notes. /v1/help lists both notes. The Terms paragraph on NHS Jobs now has its own anchor, /terms#nhs-jobs.

2026-11-10

  • Official sponsor registers. Every job now carries sponsor_registers: the official government registers its employer appears on in the job's own country (only on jobs located in, or open to, that country), such as the UK register of licensed sponsors, the IND public register of recognised sponsors (NL), the US Department of Labor LCA and PERM disclosure data and the Canadian positive LMIA employers list, with the register's own date. It means the employer appears on the official register as of that date, not that it will sponsor a visa for a given job. [] when it is on none. Filter with sponsor_register=GB (or several countries, or any). On every plan, no change to metering.
  • New POST /v1/sponsors/lookup: name up to 100 employers (with optional country, town and US state) and learn whether each appears on one of those official registers, without pulling any jobs and for employers that have no jobs here. Each answer is matched, ambiguous (every candidate legal entity is listed, never a silent pick), not found, excluded (franchise brands, staffing firms and recruiters) or no register for that country, with the register, routes, legal and trading name, towns, the register's date, a source link, the match method and a confidence. GET /v1/sponsors/lookup?name= does one name, and the MCP server has a sponsor_lookup tool. Every plan, the free one included; 0.1 record per employer, rounded up per call.
  • Sponsor lookup: every hit now says how the name matched. matched_on is legal_name or trading_name (which register name your name equals), and match_type is exact_normalised: case, punctuation and company suffixes such as Ltd or Limited are ignored, and there is no fuzzy matching. not_found means the name is on no register JOA reads for that country; no_register means JOA reads no register for that country. possible_names are suggestions, never matches.
  • company now returns an employer's full set of jobs for any of its slugs. Where we hold one employer under several names, the slug of any of them selects all of them (?company=jpmorgan-chase returns the same 8,008 jobs as ?company=jpmorgan-chase-049c4e; before, it returned only the 7,058 under that one name), and a slug a company held before still works. Requests that used the slug on the job rows answer exactly as before.

2026-11-09

  • For copies you keep yourself. Every job now carries content_version and content_changed_at: the version changes only when the employer changed what the posting says (title, description, apply link, employment type, location text, a stated remote type or salary), never when JOA re-derives or repairs a field. Rows that have no version yet return null.
  • New POST /v1/jobs/status: send up to 1000 job ids and get each one's status (live, closed, withdrawn or unknown) with its closure or withdrawal details, content_version and last_verified_at, without re-reading the rows. Growth plan and above; 0.1 record per id checked, rounded up per call.
  • Time filters: since on /v1/jobs and /v1/jobs/closed returns next_since, a token for the next poll with no gap and no overlap. New listed_before, posted_before and verified_before make ranges. Every time filter (and since on the delta feeds) also accepts now, now-15m, now-2h, now-1d, today and yesterday (UTC). Existing requests answer exactly as before; CSV exports are unchanged.
  • The closure and withdrawal lists, /v1/jobs/expired and /v1/jobs/withdrawn, are now free on every plan, the free Explore plan included. Removing closed and withdrawn jobs from a copy is something every customer has to do, so it is no longer a paid feature. /v1/changes stays on Growth and above.
  • Terms updated (section "What you may do with the data"): every commercial use is allowed on every plan, including storing, displaying, summarising and reselling the records, within your plan's limits and with a visible credit. Removal works both ways: when an employer asks us to remove its data we remove it and tell you, and you remove it too; when an employer asks you, remove it and tell us, and we remove it here.

2026-11-08

  • Jobs open to several countries. Every job now carries countries, the list of every country it is open to with country first: ["DE"] for a one-country job, ["US", "CA"] for a role such as "Remote (United States | Canada)", [] when no country is known. country itself is unchanged on every row. The new opt-in country_match=any makes country, exclude_country and a locked key's country or continent match any listed country; without it (or with country_match=primary) filtering is exactly as before. Each job is still one record. CSV exports and sample files are unchanged; the JSON responses, NDJSON export and MCP results carry the new field.

2026-11-07

  • ESCO occupations are live. A job now carries esco_occupation_uri, esco_occupation_label and isco_code, with field_sources.occupation set to inferred, when its title equals a preferred label of exactly one occupation in the ESCO classification (v1.2.1), after we remove gender markers, brackets, a trailing segment and seniority words. Otherwise the keys are left out and field_sources.occupation is absent: unknown, never a guess. About 5.8% of served postings have one today (158,512 of 2,750,111). On 416 hand checked postings, 92.8% were exactly the right occupation and 97.1% were the right occupation or a neighbour in the same ISCO unit group. Matches are our automated inference from the title, not the employer's classification. This service uses the ESCO classification of the European Commission. There is no occupation filter yet. ESCO skills are still in work and not shipped: our checks so far are below our 0.9 precision bar. Thanks to Greg Turkawka of Kitsuno.ai for the idea.
  • USAJOBS is back: 8,187 federal jobs, almost all with pay. Provider usajobs, source type public_agency, read directly from the USAJOBS Search API (the U.S. Office of Personnel Management's federal job site) every 6 hours. Since the revision of its API Terms of Service on 9 October 2026, USAJOBS public job announcement data is under an open licence that allows redistribution. Only announcements open to the general public are published; those restricted to current federal employees or other groups are not. Each job links to its announcement on usajobs.gov, where the application is made, carries pay for about 99 in 100 jobs, and names the source: the job object's attribution is "USAJOBS (U.S. Office of Personnel Management)". Agency contact details are not published. An announcement is removed when USAJOBS stops listing it or its closing date passes. These jobs are an unofficial copy and are not operated, sponsored, approved or endorsed by OPM or any federal agency; see /terms.
  • Some sources state a pay amount but never its period or currency (a home-care feed that publishes "19" and nothing else). Where exactly one reading is possible, the missing unit is now filled and labelled: the currency only when the posting's country has one currency and the amount fits it, the period only where the magnitude allows one reading for that country and currency (19 USD in the US is hourly pay, 93,600 USD is yearly pay; anything in between stays absent). Amounts are never changed and never estimated, they remain the employer's own figures; field_sources.salary_period and field_sources.salary_currency read inferred on exactly these rows. About 11,000 served postings gain a pay unit and an annual EUR figure, so min_salary and max_salary can finally see them; a hand check of 122 filled rows against the postings found one error, and the rule that caused it now refuses that case.

2026-11-06

  • iCIMS employers that run their own branded career site on iCIMS Jibe (H&R Block among them) are now read directly at the source, tens of thousands of postings that are hard to reach through the iCIMS portals.
  • robots.txt is read per path, as RFC 9309 describes (groups, Allow and Disallow, wildcards, longest rule wins), in the fetchers and in the verifier, and we fetch only where it allows. Enforcement is on for iCIMS and NAV pages today and is being extended host by host.
  • A posting that reopens at its source now returns under its original id, and the events it causes on /v1/changes are corrections that cost no records.
  • New filter region on /v1/jobs and /public/jobs: ISO 3166-2 subdivision codes or English names (region=GB-DEV, region=Devon,Cornwall, region=US-OH, region=CA-ON, region=AU-WA), up to 60 values, 422 bad_region for an unknown value. It reads a new column that is filled for US states, Canadian provinces, Australian states and territories, UK nations and about 50 UK counties; a row whose subdivision is not established is never returned by a region filter and never excluded by other filters. The existing region field in the payload is unchanged.
  • Dataset files: the employment_type column of the monthly and weekly CSV files was empty on about three quarters of the rows, because the files were built from the database column alone while the API also shows the value read off the apply page. The files built on 1 and 2 November (182 files, 6.1 million rows) were repaired in place on 6 November: employment_type was filled on 1.84 million rows and no other column, row or snapshot date changed. The manifest of each product records the repair, and new builds use the same value as the API.
  • /public/coverage: definition change from 6 November 2026, 03:25 UTC. The page-verified share (verified_last_18h_pct, verified_last_24h_pct, page_verified) now also counts a confirmation by the official feed API for exactly two sources: NAV (Norway, nav_norway) and JobTech (Sweden, jobtech_sweden). Their application links point at employer systems that often disallow automated access or only render in a browser, so for them the feed is the confirmation and the application page is opened at most every 72 hours. Other sources are counted as before. The series has a break at that time; the effect on the day was +0.2 points and can reach about +1 point as the page checks lapse.
  • Work started on ESCO occupations and skills. Each posting will carry an ESCO occupation (with its ISCO code) and the ESCO skills it asks for, matched exactly or left unknown, never guessed. Thanks to Greg Turkawka of Kitsuno.ai for the idea.

2026-11-05

  • Broken characters in advert text were repaired at the source. Stored postings carrying them fell from 9,368 to 2,150 and served postings from 1,055 to 435, out of about 2.65 million, and 119 dataset files were repaired in place. What remains is broken at the source.

2026-11-04

  • New statistics pages: job postings by country, by category, the share with stated pay by country, and remote, hybrid and on-site postings by seniority and employment type, each with a free CSV and the date of its snapshot.
  • Freshness hold: a posting whose last confirmation at the source (last_verified_at) is older than 17 hours 40 minutes is held out of the served results until it is confirmed again, so no served posting is older than about 18 hours. A held posting is not deleted.
  • API: description_contains now searches live postings only. With status=closed or status=any, and on /v1/jobs/closed and /v1/jobs/expired, it answers 422 description_contains_live_only instead of reading every stored advert; use title or q for closed postings. Live searches are unchanged.
  • Incident: from about 28 October a distributed scraper ran about 39,000 different filtered searches on the public job search of this website, which pushed its errors from about 37 a day to about 460 a day. No accounts, keys, private data or the paid API were touched. On 4 November it was blocked, keyless filtered searches now share a limited number of slots and a search that cannot run gets a fast "busy" answer instead of a timeout, and the errors went to zero.

2026-11-03

  • New source: NAV, the Norwegian Labour and Welfare Administration (arbeidsplassen.no), provider nav_norway, source type public_agency. Only ads whose application link is the employer's own application page or its applicant tracking system are published: ads that apply through NAV, by e-mail, through a job board or through a staffing agency are not. Applicants are sent straight to that link. Contact persons, e-mail addresses and telephone numbers are not published. NAV is named as the source on every one of these ads: the job object carries attribution ("NAV (arbeidsplassen.no)"), the job page shows it, and /public/providers and /public/coverage list it. An ad is removed within minutes of NAV marking it inactive or deleting it, and changes are applied as NAV makes them. /public/providers and /public/coverage provider_rows gain an optional attribution field for every source that has one.

2026-11-02

  • November yearly offer: until 30 November 2026 at 23:59 Athens time, the first year of any new annual plan costs the price of six months instead of nine (Starter €174, Growth €480, Signal €1,794, Scale €5,394, net of VAT), which is half of twelve months billed monthly. The plan renews at the normal annual price. The discount is applied at checkout, so there is nothing to type, and it is published without a code in the yearly object of GET /public/launch-offer while it runs. The free first month for monthly Starter and Growth is unchanged.

2026-11-01

  • The free 14-day Growth trial for company email addresses is retired. It is replaced by a free first month: sign up with a company email address and the first month of the monthly Starter or Growth plan is free, with the discount applied at checkout. A card is needed, the plan renews at the normal price after the first month, and you pay nothing if you cancel before the second month. One free month per company domain; annual plans, Signal and Scale are not included. Trials started before 1 November run to their end date. The free Explore key is unchanged and still needs no card.

2026-10-30

  • 31 more applicant tracking systems are now read at the source: 50skills, ApplicantPro, Bullhorn (non-agency portals only), ClearCompany, Coveto, Emply, Factorial, Flatchr, Fountain, Gem, GoHire, Harri, Hireology, Homerun, JobScore, Jobylon, Kula, Lumesse TalentLink, Phenom, Prescreen (onlyfy), Radancy, ReachMee, Taleez, TalentLyft, Talentsoft, Oracle Taleo, Trakstar Hire, Umantis, Varbi, Workstream and Zoho Recruit. About 15,000 employer boards were checked one by one before they were added, and boards run by staffing agencies were left out. Each source's first full read takes up to a day, so its roles appear over the next hours; ?provider= takes the new names.

2026-10-27

  • Unknown query parameters are refused on every endpoint now, not only on GET /v1/jobs and the export. GET /v1/me, GET /v1/me/usage, GET /v1/jobs/closed, the facet, coverage and statistics endpoints and /public/providers used to ignore a parameter they did not know, so the same typo that GET /v1/jobs answered with 400 unknown_parameter returned an unfiltered 200 elsewhere, and the caller had every reason to believe a filter applied. Each of them now names the offender and, when one is close, the parameter you meant. The 14-day request-log audit found no production integration sending unknown parameters to these routes, so this ships without a grace period.
  • /v1/jobs/closed accepts exactly the /v1/jobs filters plus its closure keys (closed_after, closed_before, closed_reason). GET /v1/me/usage accepts days and metric; its signed-in session twin accepts days only, because it does not validate a metric and an ignored metric is the miss this change exists to stop.
  • The OpenAPI specification now declares the 400 on all 21 operations (it declared 4 before), so a generated client has a branch for the error a typo produces. /docs/api/errors already listed the 400 status and the unknown_parameter code; the specification was the surface that disagreed.

2026-10-26

  • City search now understands a place qualifier written into the value. A two-letter term that is a US state code (the vocabulary ?state= takes) or a country (the ISO 3166-1 codes ?country= takes, plus UK for GB) qualifies the city terms instead of being matched as another city: ?city=Exeter, RI served every Exeter in the ledger before today, because the RI matched no city and was dropped without a word (the bare ?city=exeter set was measured at 451 rows before this change and the qualifier work leaves it alone), and now it serves only Exeters with region RI. ?city=Exeter on its own is unchanged. The US state reading wins a collision, so ?city=Delhi, IN is Indiana rather than India. A qualifier applies only while at least one city term remains: ?city=GB on its own still matches the city column as a value, because the ledger publishes 14 country-coded pseudo-city values (city=GB, city=AU, ...). A two-letter term that is neither a state nor a country is still a city term, exactly as before.
  • Two new fields on ?city= searches say how far the result spreads: location_countries breaks the matched set down by stored country, largest first with the unknown country last, and location_ambiguous is true only when that set holds rows from more than one known country. Both appear only when the request carries ?city= (a search we cannot match answers location_countries: [] with location_ambiguous: false rather than silence), both describe the whole matched set rather than the page in front of you, exactly as include_total counts it, and both are additive: a request without ?city= gets byte-identical payloads, and a breakdown that fails in its own time budget is omitted while the page is served anyway. A bare city search looks like a single-locality answer while it can span countries, and saying nothing about that was the defect. Documented with the response fields and in the place parameters.
  • field_sources.country gained a new value, unknown: the row has a country but nothing records where it came from, so the payload says that rather than dressing a guess as a fact. Every stored country now records its origin internally (producer, geo, repair or unknown) and the field follows that record: a recorded source origin reads published, a rule or a repair reads inferred, an origin nothing recorded reads unknown, and an llm_inferred marker still wins first. This rekeys published, which now follows that recorded origin instead of the absence of a marker, so a country our rules derived reads inferred where it used to read published. country_confidence is unchanged, and a row with no country still reads absent.

2026-10-25

  • Query grammar: parentheses group, and OR binds looser than adjacency. a b OR c d is now (a b) OR (c d); until today OR bound tighter and the same query read as a AND (b OR c) AND d. This is the change a customer asked for after sending title="change" "lead" OR "change" "manager" OR "training" "lead" OR "training" "manager", meaning four word pairs: it now returns the union of the four pairs, 88 GB titles, where it returned 1 (2 on the day they ran it). Parentheses are grouping now: the same query with a parenthesis around each pair also returns 88, where until today parentheses were dropped as punctuation, dropping them broke the pairing, and that version returned 0 rows with no word about it. If you wrote a workaround against the old rule, re-read it. title="change" OR "training" "lead" OR "manager", the form we recommended then, read as (change OR training) AND (lead OR manager) (88 rows) and now means change OR (training AND lead) OR manager (12,854 rows). Everything else is unchanged: the phrase "change manager" is still 29 GB titles, "change" "manager" still 47, the four quoted phrases "change lead" OR ... still 50. Every published example was re-measured against the same ledger before and after, and the two columns are in the query grammar.
  • Nothing in the query grammar is dropped silently any more. An unmatched parenthesis, an empty group, an OR with no term on one side and a bare AND or NOT keyword are now refused with HTTP 422 query_syntax naming the character and the spelling that works, checked before anything is read or charged. Until today several of them were dropped, which is how a correct query could return 0 rows and look like a thin corpus. a AND b is not an operator here: write a b for AND and -word for NOT, and quote "AND" to search for the word. Plain values are untouched: title=nurse (rn) still searches nurse rn, punctuation between words is still a word separator, and a value with no quotes, OR, star or attached dash still goes to the text index as written.
  • New GET /public/help, and GET /v1/help for callers with a key: the API describes itself in one machine readable document. Every endpoint and every parameter, the query grammar exactly as the parser implements it, the limits and quotas, the error codes and what to do about each, and one worked example per endpoint. It is versioned (help_version for the shape of the document, grammar_version for the grammar), additive within a version so a cached copy stays usable, and keyless on /public so an agent can read the contract before it has a key. A test walks the route table and fails the build on any endpoint the document does not describe, so it cannot drift away from the API.
  • Sign-in links now last 30 minutes, up from 20. Someone who asks for a link at lunch and opens the email at 3pm still gets in. 30 minutes is a deliberate ceiling, not a starting point: the link is the only credential (one working email address, no password), so it stays live long enough to read an email and click it, and no longer. The sentences that state the lifetime (the email and the confirmation on the site) say 30 minutes too, and the two cookies that ride along with it (where the visitor was heading, and which page sent them) now outlast the link instead of expiring before it.
  • The register page is now the sign-in form, the same one /login shows, instead of three steps of prose that pointed at /login. Every Get a free key button lands on a working email field: enter the address, click the link, and that click creates the account and signs it in. The free Explore key is still minted on /dashboard, one press after arriving.
  • The grant panel (free, time-limited keys for students, nonprofits and research, granted by hand) is back, this time on /pricing in its own block apart from the plans, and the plans page of the docs names it in its own note. It went away with the /register prose when that page became the sign-in form. The sign-in form footer link for visitors with no key now goes to /pricing instead of looping back to the page it sits on.

2026-10-24

  • Salary filters now agree with the payload about what a published salary is. has_salary=true (and its alias structured), require_fields=salary, the published salary facet chip counts a salary our verification pass read straight off the source's own page (field_sources.salary already said published), not only the listing column's own structured stamp. The served-row count with a published salary moves from 134,954 to 320,143: the payload had been publishing those figures all along, so this corrects an understatement rather than adding data. has_salary=any gained the same figures so it stays a superset of true. has_salary=any and the country with_salary figures (both count any figure the payload can show, published or read out of the advert) gained the same extract figures: the country total moves from 405,564 to 579,834 rows. A salary with no mark anywhere is still never published and never counted (270 rows were unpublished today for exactly that reason), and two tests now pin both directions so the payload and the aggregates cannot drift apart again.
  • Country: 309,726 rows gained the ISO country of the posting, in two passes that use only what the record itself states: the source's own country field or ISO code, a country named in the location, a province or state paired with its city, a place that is unique in the gazetteer, or a country code in the domain (253,088 rows); then, where an employer's other located rows agree on one country at 99% of 50 or more rows, that country (56,638 more, and only where the row's own location does not contradict it). The share of served rows carrying a country rose from 87.1% to 93.73%. Where the source states a country it is published as before; where we derive it from the place, field_sources.country says inferred. A country stays empty when the location text does not name a unique one (multi location rows, town names shared between countries); we do not guess.
  • Verification honesty: a page whose posting appears only through JavaScript, with nothing about the vacancy in the page itself, no longer counts as proof the posting is live. Enforcement reached the remaining large self-serve platforms (BambooHR, Workable, Paycom, Dayforce) after ADP, Oracle, iCIMS and Workday. Such rows are now unknown: not verified, not deleted, checked again on the normal cadence. The page-verified share drops on purpose, because a confirmation we cannot stand behind no longer counts.
  • Duplicate postings on Workday: an application URL written with a trailing /apply and the same URL without it are now recognised as one posting (seen on tenants such as Humana and Truist), so one of the pair can be held as a duplicate instead of both being served. This pass runs on the normal duplicate sweep.
  • Result flooding: collapse=company returns at most one row per employer per request and adds company_more_roles, the number of further roles of that employer that the request read and hid behind this row; max_per_company=N (1 to 200) caps one employer at N rows per page. Both are opt-in, the published sort order is unchanged, hidden rows are never deferred (the cursor resumes after the last row the request decided on), counts still count matching rows, and only rows returned are charged. The browse pages now use it, so one employer no longer fills the first page.
  • Ashby verification label: a row whose confirmation came from a read of Ashby's own job board now carries last_verified_via: board_api. Nothing about counting changed: such a row is page-verified as before, and the coverage report says plainly that an Ashby check reads Ashby's system of record rather than the individual posting URL.
  • One employer, one key: postings we used to show under two company names (a firm and its shared applicant-tracking host, or two spellings of the same firm) now appear as one employer everywhere: the company page, filters, exports, the changes feed and every employer count. 1,936 duplicate company keys fold into their canonical employer, covering 68,953 served rows; those rows are republished as corrections in the changes feed and cost no records, and existing company links keep resolving.

2026-10-23

  • Employer names: a posting that states its own employer (its schema.org hiringOrganization) now shows that employer as company, instead of the name of the ATS tenant it was fetched under. 13,788 served rows changed on this basis (for example the shared Workday tenants of Loblaws, Hibbett and Finish Line, Siemens Healthineers, Stanley Black & Decker). Tenant tokens that used to stand in as an employer ('Eqid', 'Myview', 'Jdgroupnam') are replaced where the employer's own pages name the company: 11 tenants named from the employer's career-site twin (Cummins, Chubb, EXL, Grupo Coppel and others, about 5,900 served rows), 79 UKG tenants named from the job board's own configuration (6,037 rows; the 606 postings shown as one TV station's news brand are now Gray Media), and 3 tenants from the employer's own site. These are corrections to our own data: the /v1/changes events carry correction: true and cost no records. Where a tenant has no evidence of its employer yet, the name is unchanged; we do not guess one.
  • Staffing and recruitment agencies: 15,434 served rows from 743 agency companies are no longer served (they are corrections and cost no records on /v1/changes). Holds on 2,349 agency and board companies were deleted by a faulty nightly job on 21 October; the 512 held on strong rule evidence and 87 more that independent evidence confirms are back (the other 66 rule matches now have a positive employer verdict and stay served), and 144 companies matched by a sourced agency brand list (109 brands) or a legal-form word (Personaldienstleistungen, Zeitarbeit, uitzendbureau, bemanning and others) are held. A held company is held on every spelling of its name, new companies are checked when they appear, and the nightly job can no longer delete a hold without a positive employer verdict. A random sample of 300 served rows on 23 October found 4 agencies (1.3%, 95% interval 0.5 to 3.4%) and 8 agency or job-board rows in total (2.7%, 1.4 to 5.2%).
  • Search is case-insensitive in every script. The database folded only the 26 ASCII letters, so a lower-case Greek, Cyrillic or accented-capital query missed every posting that capitalised the word: q=λογιστής found nothing in Greece while q=Λογιστής found 3, and title=ärztin found 131 German titles against 32 for title=Ärztin. q, title and title_exclude now fold case and accents the same way on both sides (λογιστής, Λογιστής and ΛΟΓΙΣΤΗΣ are one query); description_contains and city are case-insensitive in every script. Queries in plain ASCII return the same rows as before.
  • employment_type: a contract-duration word on its own (Permanent, Regular, CDI, Employee) is no longer shown as Full-time; it says nothing about hours. 3,193 served rows whose source states nothing else now show no employment type. JOIN rows keep Full-time, which JOIN's own structured data states. At the same time 8,345 rows gained the type their source publishes (Recruitee and Personio codes, Greenhouse and BambooHR fields, and variants such as Full Time Employee or Fixed Term Contract), so employment_type=full_time matches 757,829 rows (751,343 before) and 39.3% of served rows carry a value. These corrections are marked on /v1/changes and not billed.
  • posted_at: 95,318 served rows that had no date now carry the posting date the source itself states (the page's schema.org datePosted, or the ATS's own published date, as on Recruitee, Pinpoint and Softgarden), with posted_at_precision: date. Dates we could not trust stay empty: Greenhouse's updated_at, the placeholder date exactly two years before today that iCIMS portals publish, and dates after we first saw the posting. 79.4% of served rows now have a posted date (75.9% before). These corrections are marked on /v1/changes and not billed.
  • /public/coverage: the page-verified share now counts every time we opened a vacancy's own application page (or its ATS record of that vacancy) and found it live in the last 18 hours, even when a later read of the employer's openings list came after it. Until today that later list read replaced the record of the page open, so those rows were not counted and the figure read about 6 points low (78.9% at 16:12 UTC on the old count, 85.0% at 16:22 UTC on the corrected one). The definition is unchanged: a confirmation from an openings list alone still does not count.
  • WordPress plugin 1.0.3: incremental imports now ask for jobs that are new to the API since the last run (listed_after) instead of jobs posted within the last day. A job is imported as soon as it reaches the API, even when the employer posted it days or weeks earlier or gave it no posted date. The first run after upgrading looks one day back and de-duplicates what you already have.
  • /v1/changes: corrections are no longer billed. When we correct our own data in bulk (an employer name we fixed, a field we back-filled, a row a quality rule removes), the events it causes carry correction: true and cost no records, so receiving our fix can never use up your allowance. A row that is new to you (created) is billed as usual.
  • /v1/jobs: a query that cannot match (a word no listing contains, a city we do not have) is answered in the same shape as any other page. With include_total=true it now says total_estimate: 0 (exact); until today it gave no count at all and added limit/offset keys no other page has.
  • /v1/changes restated itself once, free of charge: every row we serve as updated, every open row we do not serve as withdrawn, and every closure since 22 September as delisted (withdrawn where the row may no longer be shown). The earlier feed could leave rows in a mirror that we no longer serve (it delivered held and gated rows as created or updated, and a hold, an opt-out or an employer change produced no event), and it could skip a re-admission or a closure. Applying the restatement brings any mirror built from it back to exactly what we serve. A country-filtered /v1/changes now reads an index of the feed itself, so it no longer slows down as the rest of the world changes.
  • New listed_after on /v1/jobs and /v1/jobs/closed: rows that entered the API after the instant you give (new rows, and rows back after a withdrawal or a closure, each counted from its latest entry). Use it to import new jobs on any plan. posted_after is the employer's own date and runs a median 31 hours, and often weeks, behind the moment a job reaches the API, so a posted_after window misses most new jobs. The n8n node and the WordPress plugin move to listed_after in their next releases.
  • The error reference lists the codes the API returns for an unparseable timestamp: bad_listed_after, bad_posted_after and bad_verified_after (422), and bad_since (400; 422 for a cursor we did not issue). It named a bad_timestamp code that the public API never returned.
  • /v1/changes no longer misses rows. A new row was keyed on when our harvester first saw the posting, which is up to hours before it entered the API, so a poller whose cursor had already passed that time never received it; changes written by long-running jobs could be skipped the same way, and a row that entered or left the served set through a hold, an opt-out or its employer record produced no event at all. The feed now records every change to the served set and publishes it, usually within a minute or two (removals first; a bulk correction of many rows can take a few minutes longer). Events are relative to the served set: rows we do not serve no longer arrive as created or updated, and a closure of a row that was never served is not an event (it is in /v1/jobs/closed). Rows first seen since 22 Oct 16:00 UTC that the old feed skipped are published again as updated.
  • WordPress plugin 1.0.2: a key with a monthly record allowance (the free Explore key included) is now paced so the allowance lasts the month. Scheduled runs share what is left evenly until the reset, a run started by hand may use the rest, and the settings page shows the allowance. Version 1.0.1 could spend up to 4,500 records a month on its defaults against a 1,000-record free allowance.
  • Licence: /v1/jobs/closed and /v1/changes returned closed rows from three discovery-only sources (EURES, the German Federal Employment Agency and France Travail), whose content we do not redistribute. They no longer do; /v1/jobs/expired no longer lists their ids.
  • /v1/jobs/expired and /v1/jobs/closed no longer show a closure until no transaction that could still commit an earlier one is open, so a poller's cursor cannot step past a closure that is still being written.
  • /v1/export now refuses parameters it cannot honour instead of ignoring them: limit, cursor, offset, sort, include_total, quality and include_poster_type answer 400 unsupported_parameter, and any unknown parameter answers 400 unknown_parameter, before anything is charged. Until today ?limit=1 was ignored and streamed every matching row, and a misspelled filter exported the unfiltered corpus. Our own documentation examples used limit=1 on the export; they now use a small country (country=LI).
  • include_total=true can no longer fail a request. The count now has its own 4-second budget; past it the page is still served and total_estimate_status says timed_out (otherwise exact, or at_least above 10,000). Before today a slow count turned the whole call into a 503. include_total and both fields are now in the OpenAPI document and on the parameters page.
  • Comma-separated lists are no longer cut short silently. A list longer than its maximum is a 422 too_many_values; until today the extra values were dropped and the narrower query was answered and billed as if it were the one asked for. country, exclude_country and state now take up to 60 values (every EEA country fits), seniority takes all 9 values, and company_domain / exclude_company_domain take up to 100 (one integration was sending 40 employer domains and silently searching only the first 25).
  • has_website, org_type and include_discovered are /v1/companies parameters; /v1/jobs used to accept them and do nothing. They are now a 400 unknown_parameter on /v1/jobs (the request log shows no customer sending them there).
  • MCP search_jobs: every argument the tool schema lists is forwarded to the API, and an argument it does not list is refused with a did-you-mean instead of being dropped. include_total was advertised and dropped; it now returns the total in the result and in the summary line. sort is advertised only when the deployment can rank.
  • Keys without a monthly record allowance now receive X-RateLimit-Records-Limit: unlimited on row-returning responses, as the rate-limit page documents; the header used to be absent.
  • /public/coverage and /public/site-stats: stale is now computed from measured_at, as documented (it used the freeze time), and both responses state their window in refresh and stale_after_hours.
  • Free keys: the panel shown when a key is created now says whether the free 14-day Growth trial started and, if it did not, why (personal webmail, school or disposable address, the company domain already had its trial, an earlier trial or a paid plan). It no longer calls a trial key an Explore key. It, the homepage, the dashboard and llms.txt no longer say Explore covers every endpoint: the delta feeds (/v1/changes, /v1/jobs/expired) and bulk export (/v1/export) are on Growth and above. POST /public/auth/keys returns the reason as trial_not_started.
  • Pricing, the register page and llms.txt state the trial rule exactly: a company email address, one trial per company domain, and the key continues on Explore when the trial ends.
  • Homepage: each ATS platform chip shows the value ?provider= takes next to its display name (d.vinci is dvinci, join.com is join), and the counts are dated with /public/providers' generated_at instead of being described as live.
  • Docs: Build a job board said no published plugin or connector existed; there are five (a WordPress plugin, an n8n template, a Make blueprint, an Activepieces piece and an Apify Actor). It now also states the two licence obligations from the terms: the visible credit “Data: Job Opportunities API” wherever listings are shown publicly, and passing on employer removals.
  • A category filter without a country (for example category=Healthcare&remote=remote) now reads a date-ordered index: 8.7 s to 0.15 s.
  • min_salary and max_salary: the docs now say they compare every normalisable figure, published or read from the advert text, and that has_salary=true limits them to published salaries.
  • Search is faster and no longer times out on common words. A q= page is now fetched by two equivalent plans at once, an index walk in date order and a match-set scan, and the first to finish answers; both return exactly the same rows in the same order. Measured on the live ledger for a 21-row page: q=engineer 10 s (and 503s) to about 0.1 s, q=nurse 10 s to 0.16 s, q=software engineer 4.4 s to 0.15 s; rare terms stay fast. One-value filters (country=US&category=Healthcare) now read the date-ordered index directly: 9.9 s (and 503s) to under 0.1 s.
  • country and exclude_country now accept only ISO 3166-1 alpha-2 codes (plus XK for Kosovo). An unknown code such as ZZ used to answer 200 with an empty page, indistinguishable from a market with no jobs; it is now a 422 bad_country (UK suggests GB, EL suggests GR).
  • Did-you-mean on unknown parameters covers the names real integrations send: salary_min suggests min_salary, count and total suggest include_total, keywords suggests q, page suggests cursor, and camelCase or upper-case names (postedAfter, HasSalary) suggest their snake_case form.
  • Documentation corrected to match the API: unknown parameters are refused (400), not ignored; limit outside 1 to 200 is a 422, and only a limit above the plan's page cap is reduced (visibly, X-JOA-Limit-Applied); the error table and the OpenAPI document now list the 400 responses.
  • A live job no longer carries an expiry date from an earlier closure. 100 rows (74 of them served) that had come back to life at their source still showed upstream_expired_at, and expires_at from it; the leftover date is cleared, the three revive paths now clear it, and a database rule stops a live row from carrying one.
  • /v1/jobs/{id}?include_closed=true now resolves every role that /v1/jobs/closed lists, under the same rules as that list. A role removed from the live set before it closed (its apply link failing, for example) used to be listed as closed and then answer 404 by id. Rows that were never a vacancy (an ATS vendor's test board, an article listing jobs) are no longer listed as closures, by id or on /v1/jobs/closed.
  • New closure reason apply_link_dead: a role whose own apply link kept failing at its source (HTTP 404/410, a page saying the position is closed, the ATS reporting it closed, or a host that no longer resolves) can now be closed, not only hidden: it is removed from the live set after two independent failing checks and closed when it still fails on three or more consecutive checks at least 14 days later. Roles closed this way appear in /v1/jobs/closed, /v1/jobs/expired and /v1/changes with closed_reason: apply_link_dead; until today such roles left the live set but were never reported as closed. The existing backlog (about 220,000 roles that had already left the live set) is being closed at up to 5,000 an hour from 06:28 UTC today, each with closed_at = the time it was closed. A role whose apply host stops resolving for 24 hours or more is now removed from the live set the same way.
  • last_verified_at now means only a positive confirmation at the source: the posting's own page or its ATS API record found live, or the posting listed in a fresh read of the employer's own job list; if nothing is newer, the time it was first harvested. Until today it was refreshed hourly for every row whose record still existed in our pipeline, without a request to the source, so a row could look freshly verified when nothing had looked at it. A new field, last_verified_via, says which kind of confirmation each timestamp is. verified_after, verification_stale, the export, MCP and the /v1/meta/freshness counts use the same definition.
  • One source for the published counts. Facet chips, the provider table, the per-country table and the coverage report are now counted in one database pass over exactly the rows /v1/jobs serves, and carry one timestamp. A facet chip's count is now what its filter returned at generated_at; until today the chips counted a looser set (about 18% more rows, and provider chips about 58% more), and the provider chips were written minutes after the rest under the same timestamp. /public/providers now reads that same pass (its generated_at equals /public/facets') instead of a separate 20-minute count.
  • Countries are ISO 3166-1 codes only (plus XK). 35 values that are not countries (TX, UN, SW, XX and others, on about 4,200 ledger rows, 37 of them served) were counted as countries and served as country values. Where the source's value was certain they now read the real country (XI, Northern Ireland, is GB; a bare US state code such as TX is US, with the state as region, and field_sources.country reads inferred); the rest now have no country. The published country count drops accordingly, from 239 to the true number of ISO codes with served rows.
  • providers now means one thing everywhere: providers with at least one served row. /public/providers lists those in data with a providers count, and the licensed sources with no served row in a new not_serving list (still accepted by ?provider=). /public/coverage adds providers, provider_rows (the per-provider counts, which add up to live_listings), providers_not_serving, generated_at and the definitions; /public/coverage/countries adds countries_total (its table lists only countries with 25 or more rows); /public/site-stats adds providers_with_live_roles.
  • Homepage: the ATS platform chips now come from the same daily snapshot as the figures at the top of the page, so they add up with them and carry the same date. Docs: a new section, What each public count counts, states the population and the timestamp of every public count.
  • Salaries: a published pay period is now always one of hour, day, week, month or year in lower case. Two ATS sources sent it upper-case (HOUR, YEAR, about 3,200 served rows), and a pay structure such as BASE + TIPS, OTHER or COMMISSION ONLY was published as if it were a period; those rows now keep the amount with no period. Because the upper-case period slipped past our plausibility rule, 180,000 USD per HOUR was annualised to 344,448,000 EUR and matched any min_salary; every salary now goes through one rule (the period is normalised first, then the plausibility bands, then the annual figure), and about 230 served rows whose figure cannot be pay (for example 70,000 USD an hour) no longer show a salary.
  • Salaries read from the job page (used when the listing itself has none) follow the same rule: about 11,500 served rows changed. Most lose a figure that cannot be pay (130,000,000 USD a month, 26.59 EUR a month, an annual salary labelled per week); a currency written as a symbol or a word (£, €, EURO) is now its ISO code, and a period written in Dutch, German, French, Italian or Spanish (p/maand, Monat) is now the English period.
  • Employer names: about 1,250 employers that were shown under an ATS account code (capiosverigeab, ymcaboston, Workday globalhr, Oracle ewwt) now carry their real name (Capio, YMCA of Greater Boston, Raytheon, Americold), on about 58,000 served rows. The name comes from the employer's own job pages, or from the copy of the same requisition on the employer's career site. The hourly company refresh can no longer replace an employer's name with an account code.

2026-10-20

  • Signup attribution: the free-key signup recorded, per key, how its account found us (the referring site and campaign tags captured on first landing, and an optional "How did you hear about us?" answer from the signup form). The referrer/campaign capture is consent-gated (off by default; turn it on under "Customise" in the cookie notice) and, once allowed, kept in your browser for up to 30 days before being attached to a key you create. Never required, never an IP address, never fingerprinting; see Privacy and Cookies.
  • New source: Teaching Vacancies (UK Department for Education), provider teaching_vacancies, prepared under the UK sources batch (not yet live). Rows carry the Open Government Licence v3.0 attribution and licence URL; the licence's one condition (no fee or commission for contacting, interviewing or hiring a candidate who responds to a Teaching Vacancies listing) is stated on /terms, Sourcing and refusals and Legal, cookies and crawlers. The /bot page wording on how we find an employer's careers page was also tightened: we read robots.txt and, where a site publishes one, the standard /sitemap.xml at the site's root; we don't guess any other addresses.

2026-10-18

  • Boolean flags (include_description, has_description, include_discovered, remote_confirmed, include_total) and has_salary now accept only true / false (any case). 1, 0, yes and no used to be accepted silently; they now answer 422 bad_boolean (bad_has_salary) and nothing is charged.
  • Employer names: 24 employers hiring through Ashby (among them 1Password, Cerebras and Ethon) were shown as 10x Team for about four hours after a name repair matched them by the shared ATS host; the repair now only trusts hosts that belong to one employer, and the names and company slugs are restored (the interim slugs keep redirecting).
  • City: about 275,000 live rows whose resolved city was shown but not stored now also match city= and the city facets; new rows are filled every 10 minutes. Store and mall names in brackets (for example (Assembly Row - Coach)) are no longer read as the city.
  • Salary parsed from the advert text when the employer states it explicitly: an amount with a currency in a pay sentence with a stated period (for example $106,900 - $147,000 per year, £15.50 per hour, EUR 50.000 - 60.000 brutto/Jahr). These rows carry salary_source: parsed_description and field_sources.salary = inferred; a single stated figure fills salary_min only. Bonuses, sign-on and referral amounts, commission and on-target earnings, "up to" ceilings and company revenue figures are never read as pay. has_salary=true still means a salary the source published in its own field; has_salary=any, min_salary and max_salary include the parsed ones.
  • seniority is now also read from level words in German, French, Dutch, Spanish, Portuguese, Italian, Polish and Swedish titles (for example Praktikant, Ausbildung, stagiaire, alternance, Teamleiter, responsable, chef de projet, gerente, medior, pleno) and from staff-engineer titles, by fixed keyword rules. A title that names no level still has no seniority. field_sources.seniority stays inferred.
  • Both rule sets were checked by hand on country-stratified samples before release: 98.8% of 260 parsed salaries and 98.4% of 309 title-based seniority values were correct. About 22,000 served rows gain a seniority today; about 1.7 million served rows without a salary are being read for one, newest first.
  • Existing served rows are filled in the background at most about 30,000 rows per day for both fields together; each filled row appears once on /v1/changes as an ordinary updated event while the backfill runs (about 30,000 updated events a day from this, on top of normal traffic).

2026-10-13

  • stale_posting: adverts whose own posted_at is more than 24 months old are now held back by default as a quality gate (rule stale_24m). 48,294 rows that were served left the default served set (about 2.5%) and rows with no posted_at are never gated. quality=all re-admits them on paid endpoints, each carrying its verdict and rule id.
  • Mirrors see the change on their next poll as about 48,000 ordinary withdrawn events, billed against records_per_month like any other event. If your quota runs out mid-stream nothing is lost: your cursor stays where it stopped and the remaining events arrive after the monthly reset.
  • New source: employers who publish their own openings on SuccessFactors career pages, read from the page's structured data (provider company_site, tag sfpub-2026-10). 3,944 served rows on day one, including Nestle, DSV, Zurich Insurance, Vestas, Novo Nordisk, Swiss Re, Nordea, Danish Crown, Carlsberg, EPFL and ICRC.
  • About 79,000 rows whose location text ends in an explicit country name but carried an empty country now carry that country, so country= filters and the country facets match them (judged 150/150; new rows are filled hourly). Rows whose location does not end in a country name stay empty.
  • 35,901 distinct postings that had been grouped away by intake de-duplication are now projected and served: the de-dup groups identical postings (same company, title, country and city) instead of whole job families. Delta-feed consumers receive them as new rows.

2026-10-12

  • q now searches each posting in its own language. The index tokenises a row's title, company and location with the snowball dictionary of the language the row is labelled with (German, French, Spanish, Dutch, Italian, Portuguese, English; anything unlabelled or outside the seven stays on the plain dictionary), so a German search for the singular finds titles carrying the plural: maschine finds Maschinen, the way entwickler* already finds entwicklerin. The query side tries every dictionary at once, so a single query reaches rows in all seven languages, accent-insensitively in both directions: an accented query finds rows stored without the accent, and the reverse.
  • Postings with no language label keep exactly the behaviour they had: indexed and served through the plain dictionary, accent-folded but not stemmed. They lose nothing: they simply do not gain stem folding. (Roughly one in six served postings is unlabelled today; the label improves as detection re-runs, and each row picks it up as soon as it arrives.)
  • Scope, stated plainly: title, title_exclude and description_contains keep their current semantics in this release. The per-language advert-text search ships as its own step rather than as a silent half-measure.

2026-10-10

  • Every job row can now tell you how fresh its verification is: rows whose last positive source confirmation (the existing last_verified_at field) is older than 18 hours now carry verification_stale: true. An honest label, never a removal: stale rows stay served, labelled. Fresh rows carry no new field at all.
  • /public/providers now publishes what the API actually serves. The per-provider live_listings counts rows through the served view (joa.v_served_v1, named in the payload note and dated in generated_at): the quality gates, dedup holds, strikes and opt-outs applied. The page previously carried the raw ledger number: 42% more than what the API returns. Providers whose served count is zero stay listed: the filter still accepts them, and a zero is disclosure, not silence.
  • The MCP tool descriptions and the parameter reference now state the valid values inline: seniority: Intern, Entry, Mid, Senior, Lead, Manager, Executive, Director, not_stated; employment_type: Full-time, Part-time, Contract, Internship, Temporary, not_stated; remote: remote, hybrid, on_site, not_stated; source_type: ats, career_site, public_agency, aggregator, agency (all verbatim from the live facets). No more guessing mid-level where mid is the value: shipped on an MCP power user report that LLMs guess parameter values until the lists sit in the tool descriptions themselves. That is where they belong.
  • The category vocabulary is now stated inline too: Healthcare, Engineering, Operations & Admin, Sales, uncategorised, Skilled Technician, Retail, Hospitality, Finance, Logistics & Transport, Customer Support, Marketing, Education, Manufacturing, HR & Recruiting, Data & Analytics, Consulting & Strategy, Construction & Trades, Legal & Compliance, Design, Product, Science & Research, Security, Procurement, Safety & Environment. It is in the MCP tool description and the parameter reference. The q parameter now says plainly: q is free-text search, category is a fixed list. And a second user report asked, so it is answered in the docs: facet filters match case-insensitively (values are lower-cased and whitespace-trimmed before matching), so lowercase lists work exactly as well.
  • Job titles read clean: the title field now displays titles with source-site noise (pipes, pipes with keywords, banner suffixes) stripped, the same clean form the site shows. The raw form remains the source of truth for matching.

2026-10-07

  • /v1/jobs now answers a non-zero offset with a **422** offset_not_supported instead of silently ignoring it and returning the first page again. The endpoint has always paged with cursor (pass back next_cursor); offset=0 is still accepted. /v1/companies keeps its documented offset.
  • A public product roadmap: what we have shipped, what we are building now and what comes next, on one page, with a form for proposing features. Proposals land in the queue we work from, and each item links to the changelog, the docs or the page where you can check it.
  • Four new contract fields on every job row: engagement (the employment relationship: employee | vendor | other), contract_category (permanent | fixed_term | project_based | seasonal | apprenticeship | internship | other), agency (direct | agency | umbrella) and working_schedule (full_time | part_time | flexible). Each is filterable (?engagement=, ?contract_category=, ?agency=, ?working_schedule=; comma = OR, unknown selects absence) and carried in field_sources. Values appear only when the posting or its source states them: nothing is inferred from silence, so most postings show unknown for the contract fields at launch, and contract_category=permanent returns only postings that state permanence. Hybrid national contract forms (for example a Polish umowa zlecenie or an umbrella arrangement) are reported as other, with the raw national term retained internally. Per-field coverage (stated by the source versus established) is published on /public/coverage. employment_type is unchanged and frozen. The per-country notes are in the parameters documentation.
  • q gained an extended search syntax, applied whenever it is used: quoted phrases ("customer support" matches the words adjacent and in order), trailing-star prefix matching (entwickler* also finds entwicklerin and the German gender-inclusive entwickler/in, which is a single token to the index), a OR b, and -term exclusions. Inside an extended query, terms of five or more letters are prefix-matched automatically (remote finds remotely, developer finds developers); quoted terms always match exactly. An accented query also tries its accent-folded form, so a search for the French word for developer also finds rows stored without the accent. Queries that use none of these keep the previous whole-word behaviour exactly. The same syntax now also applies to title, title_exclude and description_contains (the advert text), and the title and employment_type filters were reworked to stay fast under load.

2026-10-06

  • Advert text from GET /v1/jobs, GET /v1/jobs/{id} and the MCP tools no longer carries personal email addresses: an address that belongs to one person (for example a recruiter's own name-based address) is replaced with [personal contact removed]. Company mailboxes such as jobs@ stay in the text, phone numbers are unchanged, roles an employer listed with us directly keep their text as published, and apply_url is unchanged. The bulk export keeps its stricter rule (every email address and phone number removed).
  • Company open-roles history (/public/companies/{slug}/open-roles-history) is now snapshotted daily; it had been running weekly while the docs said daily. Real daily points start on 2026-10-06 (earlier snapshots exist only for 2026-09-28 and 2026-10-04). Each point now carries a method field (snapshot or reconstructed); no reconstructed points are published.
  • employment_type filter now matches every value the API shows. It previously matched only the harvested value and missed the employment type read from the posting page itself, so filtered counts rise sharply: on live listings, Full-time from about 117,000 to about 806,000, Part-time from about 10,000 to about 143,000, Contract, Temporary and Internship also up; not_stated falls from about 1.99M to about 1.15M. Displayed values are normalised to the five published values; a page-declared value outside them is no longer shown. full_time / part_time are accepted as spellings.
  • MCP server: get_job returns the complete advert text (previously cut at 4,000 characters). List tools keep a 4,000-character preview per row by default, mark shortened rows (description_truncated, description_total_chars) and accept description_max_chars up to 100,000. Text is cut on character boundaries, never mid-character.
  • New always-present job field description_truncated: true while JOA's stored copy of an advert is known to be cut short at its source. Adverts harvested under an old 4,000-character limit are being re-fetched in full.
  • Re-verification: the share of live postings verified in the last 24 hours is published on /public/coverage (verification), with a per-day history.

2026-10-05

  • Terms: data delivered on any plan, including the free Explore plan, may now be stored, displayed, resold and redistributed within your key's plan limits, with a visible "Data: Job Opportunities API" credit when shown publicly; the affiliate programme is the option for earning without reselling.

2026-09-28

  • Billing moved to Creem as our Merchant of Record: hosted checkout, webhook-driven activation, and a self-serve billing portal. The old Viva Wallet / Revolut payment code paths are retired.
  • Anonymous access to job and company rows now requires a free API key. Statistics endpoints (counts, coverage, freshness, providers) remain fully open with no key and no account.
  • Added the Starter plan: €29/month, 10,000 records/month.

2026-09-26 to 2026-09-27

  • Added an internal, HMAC-signed read API for partner integrations, and an internal market-signals rollup endpoint. Neither is part of the public API.
  • Indexed live listings for faster internal lookups, and extended the internal search path.
  • Restored the data-quality gate that determines which rows are safe to serve after a period where it had been unintentionally disabled.

2026-09-12 to 2026-09-17

  • Rebuilt the re-verification schedule so every served listing is re-checked on a deterministic cadence rather than an ad-hoc queue, with per-host rate limits and drift detection.
  • Improved employer-identity matching so postings from the same company are consolidated into one record instead of fragmenting across ATS tenants.
  • Hosts that stopped resolving are now excluded automatically rather than left queued indefinitely.

2026-08-27 to 2026-08-29

  • VAT numbers are now verified for real at checkout, so the total shown matches the total charged.
  • Every live API key now has a recorded owner, and signing out actually revokes the session.
  • Fixed two billing edge cases so a subscription payment could not mint a duplicate key.

2026-08-19 to 2026-08-23

  • Job listings with a confirmed-dead apply link are no longer served, and two data-completeness gaps were closed.
  • Company records with an empty sector list (previously ~402,000 of them) now carry real sector data.
  • /public/status and /public/stats were fixed to measure the live ledger correctly after both had been silently reporting stale or wrong numbers.
  • Added ?include_poster_type= (agencies and boards are excluded from results by default).
  • Page size became a per-plan limit rather than a fixed constant.

2026-08-14 to 2026-08-16

  • /public/stats and the sitemap were fixed after both incorrectly reported that the site had no jobs.
  • Stopped publishing France Travail (French public employment service) listings that its licence requires us to anonymise.
  • Added a country-level statistics endpoint so callers can ask what coverage exists for a given country before spending a record.

2026-08-05 to 2026-08-11

  • Issued the free Explore API key: no card required.
  • Made the single-country plan purchasable, and published the plan ladder with every number read live from the database rather than typed into a page.
  • Added a live demo on /api and published /coverage/countries so the coverage claim behind the pitch could be checked without a key.
  • Cached the coverage report so it stopped requiring a full ledger scan on every page load.

2026-08-01 to 2026-08-04

  • Added self-service account sign-up, and an admin console covering tasks that previously required an SSH session.
  • Added a self-service opt-out for employers who want their listings removed, and made it actually remove them from what the API serves.

2026-07-30 to 2026-07-31

  • Initial public launch of the jobopportunitiesapi.org ledger, API, checkout and accounts.

Compiled from the project's commit history. Something missing or wrong? Tell us.