Methodology & sources

What the summary cards mean

Verified job cuts is the main figure. It counts cuts with a filing or an independently reported source behind them. Explicitly AI-attributed is a subset of Verified job cuts. It counts only the cuts whose source explicitly names AI as a cause. Announced job cuts is a separate announcement-history figure: source-linked plans reported at the announcement stage. When we can confidently match a later filing or report, we link or merge it. An announcement we have not matched does not mean the cuts went unexecuted. We do not count announced cuts in the Verified or AI-attributed totals, so the cards do not double-count.

Geography in the cards. Country and US-state filters describe the documented location of affected jobs, not an employer’s headquarters or every place it operates. A national announcement with no source-supported job-location state stays state-unspecified. We never infer a state for it.

What this is. A continuously updated, source-linked database of publicly reported layoffs worldwide. It records the source, evidence quote, entry status and revision history, so anyone can check every figure independently. It does not claim complete coverage in every country.

Where the data comes from

Every entry always carries a source label:

• SEC filing Legal 8-K and 6-K filings that we pull from SEC EDGAR full-text search. These are the strongest evidence we hold. They cover US public companies and foreign private issuers that file with the SEC.
• WARN notice State government mass-layoff filings from 46 US states and DC.
• Press release Investor-relations and newsroom feeds that we have reviewed. This is the “Press release” source in the tracker’s filter.
• News Named reports that we find through GDELT and Google News. We keep one only when the record has usable evidence. Eurofound ERM is a separately labeled European announcement source with its own threshold.

How often it updates. News and SEC filings: once daily, at 6 PM EDT. WARN notices: daily, sweeping every covered state. An automated anomaly review runs daily. It flags statistically unusual entries for a person to inspect: very large single notices, the same company filing in several states, and weak source links. The review catches them before anyone else does. A monthly self-audit re-opens a random sample of published rows and re-checks each one against its own source.

How we extract and check an entry

We search for layoffs in many dialects. The word list covers “layoffs”, “job cuts”, “redundancies”, “retrenchment”, “dismissals”, “sackings” and more than thirty other phrasings. We run it across GDELT’s 65-language translated index, so a report that never uses the word “layoff” still surfaces. A model pulls the facts out of news stories and filings, and every core fact must appear in the source text. Counts parse conservatively: a range resolves to the lower bound. Countries and industries normalize through fixed vocabularies, and we throw out any value that is not plausible. Each new record carries an evidence confidence and a publication status. Exact fingerprints, same-company guards and cross-source comparison prevent double counting. A candidate we are unsure about stays provisional, and it never quietly inflates the verified totals. WARN filings skip the language model. They also stay exempt from fuzzy dedup, because one employer can legally file several distinct notices.

Reason tags

Each entry can carry one or more reason tags. We assign a tag only when the stored source text explicitly supports it. An entry whose source states no reason stays untagged, and we never guess one. Two tags in the fixed vocabulary cover AI. AI or automation means the employer names AI or automation, with a quote on file. AI press-linked means the press ties the cuts to AI without the employer saying it. The rest of the vocabulary is Revenue decline, Restructuring, Merger / acquisition, Offshoring, Product discontinued, Cost reduction, Macroeconomic, Plant / site closure, and Bankruptcy / insolvency. One more tag, Government / public sector, covers public-sector actions such as federal reductions in force. A private contractor losing government work does not qualify. You can filter by tag on the tracker, and the API returns them as reason_tags. We read the two AI tags from the source text. They are a different measure from the AI headline tiles, which count the AI attribution flags ai_explicit and ai_causation. A tag labels one entry’s stated reason, while a tile counts the jobs behind an attribution. Both draw the speaker line the same way. AI or automation and the strict tile each need the employer to name AI. AI press-linked and the broad measure each hold the cuts where only the press said it. The tracker labels the two apart, so no one reads one as the other.

How the AI tag works

We sort each mention of AI into five classes: a primary cause, a contributing cause, a selection or operations tool, background context, and an explicit denial. Only a primary or contributing cause can earn the AI tag. Each one must carry an exact supporting quote from the source text. AI investment, projections about future automation, and AI used to select workers do not qualify on their own. The strict AI tag also requires the employer to have attributed the cuts to AI. A news report counts when it quotes or reports the employer naming AI, because the words are still the employer’s. A journalist’s own characterisation does not count, even when it is unambiguous. Cuts made to hire people with stronger AI skills do not earn the strict tag either. The test is whether the work went away or the required skill changed. If a system now does the work, that is an AI layoff. If the same work still needs doing by different people, that is restructuring. Next to the strict tag we keep a second, separately labeled measure: AI-linked, broad. The two tiers differ on both counts, speaker and strength. The strict tier holds cuts the employer itself blamed on AI, with the exact quote on file. The broad tier holds every other AI link, so a press characterisation the employer never made belongs there. It also counts cuts made while a company funds an AI pivot, and AI-driven market disruption. The API reports the strict tag in ai_explicit and the broad measure in ai_broad_jobs. We never mix the broad measure into the strict verified-AI totals, so the two are never added together. We publish every strict AI attribution with its quote on the AI layoffs, in the employer's own words page.

How “Roles most impacted” works

Some sources name which teams lost jobs. A source might say it is “laying off customer-support and recruiting staff.” A model reads that stored text and maps it to a fixed set of role categories. A second, independent pass must agree, and a supporting quote must be there, before we store the category. We infer nothing from a company’s industry, and we guess nothing. Each bar shows the total job cuts we tie to that team. The orange segment and the 🤖 figure show how many of those were AI-linked. A bar with no orange means job cuts we could not tie to AI, not an error. This chart covers only the minority of records whose source actually named the teams affected. So it is a sample of where cuts land, never a breakdown of the full total.

Coverage and honest limitations

Our US depth is greatest, because of the WARN and SEC sources. Europe has structured coverage of large announcements through Eurofound ERM. Outside those live collectors, country-level coverage comes from worldwide news discovery and any company newsroom feed we have reviewed by hand. Named filing systems such as SEDAR+, RNS, ASX, TDnet and HKEXnews are research candidates. We do not quietly treat them as live feeds. WARN and ERM each set their own thresholds and geography rules, so no one should sum them as a complete national census. Multi-state and multi-country entries can overlap, and the entry and source fields disclose that risk. An entry dated in the future is announced or filed, but not yet complete. Filtering the table by a country also includes cuts by employers headquartered there whose layoff spanned multiple countries. Each such row keeps its true “Multiple countries” label, and we never recount it as that country alone. The headline totals stay on the stricter job-location basis, so a global figure never inflates them.

The two counting bases, side by side

Both numbers below are correct. They answer different questions. We publish the stricter one and disclose the other, so anyone comparing us with an outside estimate can compare like with like.

Figures update automatically as records are verified. Compare like with like: quoting our job-location headline against a survey that counts by employer will understate us by the difference above, and the reverse will overstate us.

What we leave out. Rumored or unsourced layoffs. Layoffs with no stated job count. Forward-looking projections, such as “could cost X jobs by 2050,” rather than announced or executed cuts. Look-back summary articles that would double-count entries we already track.

We preserve the source links

Sources rot: states roll WARN notices into annual archives, and outlets move or delete articles. So we send every cited source URL to the Internet Archive (Wayback Machine) on an automatic schedule. Each entry shows its archived copy next to the original link. An entry whose source has no snapshot yet says so on the page, along with the date of its next automatic check. You can always see what is there and what is still missing.

Source-link preservation, measured live: 21,997 of 25,293 distinct source links (87.0%) have a permanent Internet Archive (Wayback Machine) snapshot. Of the rest, 0 are queued for the next daily archiving run, 90 have a capture requested and are retried every 3 days, and 3,358 are not in the Internet Archive yet and are re-checked weekly, forever. Rows without a snapshot say so on the page, with the date of their next check.

What qualifies as a record, by jurisdiction

Filing rules differ from place to place. Different laws and different thresholds trigger a US state WARN notice, an SEC filing, a Eurofound ERM announcement and a press report. So what enters this tracker for one place is not the same thing as what enters it for another. The table below states, for each jurisdiction, exactly which register we read and what its own rules admit. We build it from the collectors’ own configuration, so it cannot drift from what actually runs. When a source does not document a threshold, the table says UNKNOWN instead of filling one in.

Because these definitions and thresholds differ, per-jurisdiction totals document different things and are not directly comparable with each other. Thresholds shown are the ones the source itself documents; anything not encoded in the collectors is marked UNKNOWN rather than filled in.

What a notice register does not tell you

A notice register records what an employer told the authority it intends to do. It does not record what happened next. Several of our sources are registers of this kind. Each publisher states the limitation itself, and we pass it on rather than smoothing it over.

Quebec. The ministry publishes the collective-dismissal notices it received each month. It says four things about that list. The notices are an intention to dismiss, not a finished layoff. A dismissal may fall outside the month its notice arrived in. A layoff later cancelled stays on the list. And each monthly list is a snapshot, so the ministry does not revise it afterwards. A Quebec total therefore runs high against dismissals actually carried out. We publish these rows as filed notices, and every row says so.

Poland, Mazovia. The regional labour office prints two numbers in the same paragraph. One counts people who lost work that month. The other counts people named in that month’s new notifications. We read the second and ignore the first. They measure different things, and adding them would count the same job twice.

US state WARN. The same gap exists here, with one difference that matters. Most states mark a notice rescinded or cancelled, and we drop those rows. Quebec publishes no such marker, so we cannot make the same correction there.

WARN notice periods, measured

The federal WARN Act (29 U.S.C. 2102(a)) requires covered employers to give 60 days’ written notice before a qualifying mass layoff or plant closing. Many state WARN records carry both the official notice date and the effective date, so we can measure the recorded notice period directly. The figures below are pure date arithmetic on those two recorded fields.

Across 15,817 US WARN notices that record an official notice date and a later, distinct effective date, the median recorded notice period is 61 days, and 31.5% (4,986 notices) record a gap shorter than the federal 60-day period.

Excluded and counted, never guessed: 21,497 notices missing one of the two dates; 5,308 whose stored notice and effective dates are identical (several states publish a single date, which the importer stores in both fields, so a zero gap cannot be told apart from a genuine same-day notice). This makes the shorter-than-60 share conservative: real same-day notices, if any, are excluded rather than counted against employers.

States with fewer than 25 datable notices are included in the overall figures but not listed separately. Figures recompute automatically as notices arrive.

What a short gap does and does not mean. The statute itself allows shorter notice under three exceptions: a faltering company, unforeseeable business circumstances, and a natural disaster (29 U.S.C. 2102(b); 20 C.F.R. 639.9). An employer may also pay wages in place of part of the period. Only a court may decide whether an exception applies (29 U.S.C. 2104). So we report a gap shorter than 60 days as exactly that: a recorded gap shorter than 60 days. It is a timing observation, not a statement that any employer failed to comply with anything. Some states also run their own notice laws with longer periods. The comparison here is against the federal 60-day period only.

Why our totals differ from other headline numbers

Three kinds of trackers measure three different things. Government statistics (BLS) count every separation in the economy, millions per month, with no event-level detail. Announcement surveys count corporate intentions. When a CEO announces “20,000 cuts over the next two years,” the full 20,000 lands in their total that day. Much of that may come through attrition, get scaled back, or never produce a single filing. This tracker counts only what has a verifiable document or quoted primary source behind it. That means the WARN notices and SEC filings that appear as those 20,000 cuts actually execute, plus reported cuts with a named-outlet source.

Treat our verified figure as a documented floor. It is smaller than the estimates, but every single number is clickable back to a legal filing or a named outlet. Since July 2026 we also track announcement-stage cuts as their own labeled tier. We tag them “Announced” in the table and show them as a separate headline number, so one page answers both questions. Unlike the announcement surveys, every announcement here links to its source too.

Where the uncounted cuts go. US reporting law leaves large, legal gaps that no receipts-based tracker can see. Naming them is part of being honest about what our floor is. The federal WARN Act requires a public notice only when a single site loses 50 or more people at an employer of 100 or more. So a company that spreads the same cuts across many smaller offices files nothing. An employer can also skip the public notice entirely by paying wages in place of the 60-day warning, and that layoff never reaches a state database. A global “reducing headcount by 10,000” announcement often resolves to far fewer US filings. Overseas cuts, natural attrition and voluntary buyouts, none of which file WARN, come out of the total first. Small businesses and contractor terminations rarely generate any filing or news at all. The economy-wide total for those uncounted cuts comes from statistical estimates, such as US Bureau of Labor Statistics surveys and weekly unemployment-claims data. Those estimates do not name an employer, and we do not restate them as tracker rows. A number we cannot trace to a source is exactly what this tracker exists not to publish.

Compare us with the public trackers category by category, and our number is not always the smaller one. We run higher than WARN-only aggregators, because we add SEC filings and named news on top of the same notices. We run at or above tech-event trackers by job volume. On our broad measure we run at or above the announcement AI surveys, with a quote on every entry. We run lower only on all-industry totals. That gap is receiptless cuts: federal-workforce reductions, buyouts, attrition, and small closings that file nothing. We do not claim them, because we cannot source them.

The tracker audits itself

Once a month, an automated audit draws a random sample of rows we have already published. It re-opens each row’s own cited source. It then checks that the source still supports that company, that count and that date. We write the result to the public health ledger, flattering or not. A mismatch goes to a person for review through the corrections process. We never silently edit it.

Latest audit result: 44/50 verifiable rows matched their source = 88.0% (33 unverifiable: dead link / bot wall) (checked Aug 1, 2026)

You need no permission to check our work. Every row links to its source, and the health page shows each collector’s live status.

Audit this tracker. An auditor’s pack indexes everything an outside reviewer needs. It covers the recall gold set and its protocol, the live data-integrity invariants, the monthly source-audit sampling, and the corrections log. It also lists the exact commands to re-run each measurement offline.

Who runs this

AskTheRecruiter.com builds and operates this tracker. The data is free to use with attribution (CC BY 4.0). There is no paid tier for the dataset and no paid placement in it. An employer cannot pay to be added, removed or reworded. The operator does not sell severance or outplacement services to the companies that appear here. Corrections reach us through the contact page, and we disclose every accepted fix in the public corrections log. The collection and checking code is public in the tracker’s repository.

Using the data

How to phrase a citation. Our totals cover what this methodology documents: a verifiable floor, not a census of every layoff that happened. The accurate phrasing is “According to AskTheRecruiter’s AI Layoff Tracker, N job cuts are documented for [period]” rather than “there were exactly N layoffs.” No tracker of any kind observes every layoff. Ours is the one where you can trace each counted cut back to its document.

Free with attribution to asktherecruiter.com (CC BY 4.0). The CSV and JSON buttons download exactly what your current filters show, or the full dataset when you set no filter. Each chart offers its own image or data download. For programmatic access, call GET /blog/wp-json/layoffs/v1/query. It is paginated, and its filter params match the page: years, quarters, months, industry, country, state, sources, reasons, q, from, to. Call GET /blog/wp-json/layoffs/v1/aggregate for totals and breakdowns. Corrections get priority via the contact page, and we disclose every fix in the corrections log on the tracker.