Methodology

SHIFT is a labour-demand shift index. It measures changes over time in the number of job advertisements matching defined role searches. It does not count people, hires, vacancies across the whole economy or salaries, and it does not by itself establish why demand changed.

What it measures

For each kind of work, in each market, we count the job advertisements matching a list of job titles fixed before the counting — the same list, applied the same way, in every market. Edition 1 counts two windows of twelve complete months: the 12 months to June 2024 and the 12 months to June 2026, 4,241,295 and 9,323,023 advertisements respectively. What we report is not the count. It is how that count moved against the whole basket of work in the same country over the same two years. A kind of work that grew faster than the market it sits in reads differently from one that grew more slowly, even when both grew. That ratio is the reading — we call it the SHIFT Pull, and it is defined in full below. The counts themselves are the working.

How a figure is computed

Any index that combines markets has to choose how it combines them, and that choice can change which side of the market a kind of work lands on. The rule is therefore written out here in full rather than left implicit in the code.

Write Basket(m, year) for all the advertisements we counted in market m that year, across every kind of work. For an occupation o in a market m:

R(o, m) = [ Postings(o, m, now) ÷ Postings(o, m, base) ] ÷ [ Basket(m, now) ÷ Basket(m, base) ]

That is the occupation's growth relative to its own market, so anything that moved a whole market between the two years — including how much of it our provider indexed — cancels out of the numerator and the denominator together. The 24 market figures are then combined with weights fixed in advance:

w(m) = Basket(m, base) ÷ all base-window postings  ·  Ratio(o) = Σ w(m) · R(o, m) ÷ Σ w(m)

That combined figure is the SHIFT Pull. It is the one number this index reports for a kind of work: above 1 means the work pulled ahead of the market it sits in, below 1 means it fell behind, and 1 means it moved with its market. It is not a count, not a forecast, and not a statement about why demand moved.

The weights come from the base-window basket and are the same for every kind of work. They are deliberately not each occupation's own geography: weighting an occupation by where it happens to be advertised would let the shape of its coverage decide its answer. 5 cells in 552 have no base-window count for the occupation and therefore no ratio; they are excluded and the weights renormalised over the rest.

The figure is then tested against five alternatives

A single rule can flatter a conclusion. So each figure is recomputed under five other ways of combining markets, fixed before we looked at what they produced: the median of R(o, m); the unweighted mean; the geometric mean; the same weighted rule restricted to the markets that clear the floor; and a pooled ratio that sums postings before dividing.

There are five, and it is these five. They were settled on 20 August 2026, in advance, so that no test can be added later because it happens to change a label. If one is ever added, we will say when and why.

What the test decides is what we are willing to publish:

• All five agree — the direction is published.
• One disagrees — the direction is published, and the page says so.
• Two or more disagree — we publish the figure and no direction at all: no clear direction — reasonable aggregation methods produce different classifications.

In Edition 1 that leaves 17 of the 23 robust, 2 sensitive to the rule, and 4 with no published direction. We would rather say that five times than pick the rule that reads best. Where a direction is not published, the country-by-country picture behind it is unaffected.

Methodology revised before commercial launch following independent statistical review. All Edition 1 results were recalculated using the revised aggregation method. The earlier rule summed postings across markets before dividing, which allowed a market's change in coverage to enter the numerator of every kind of work present in it. The rule above removes that by construction rather than testing for it afterwards.

Why we never compare markets

Advertisement density, population, and how much hiring happens online all differ too much between countries for the raw counts to mean the same thing. So SHIFT never compares absolute levels across markets. What is comparable — and what the index plots — is the direction and pace of change within each market.

Source & cadence

The primary source is aggregated online job-advertisement data from a third-party provider, used under the applicable provider terms. Edition 1 is built from two twelve-month windows — the advertisements counted in the 12 months to June 2024 and in the 12 months to June 2026 — collected on one fixed schedule for every market and every kind of work. There is no monthly rollup and no moving average in this edition: each figure is a count, not a smoothed series. Both windows are counted on the same sources, which removes the largest way a market could move here for reasons of measurement rather than hiring. It does not remove every way: a source can index a country more heavily in one year than another, and where we have found that, the market is published without a direction. That discipline was not free: it is the reason the source list looks narrower than it could be, and what that costs us is set out below.

Data retention & reproducibility

SHIFT retains the time-series observations, calculation outputs and audit records needed to reproduce the index, for as long as permitted by the applicable source-provider terms and required for methodological integrity. SHIFT does not claim a right to retain or republish third-party source content indefinitely.

Taxonomy & coverage

Edition 1 covers 23 kinds of work, chosen to put work with high AI exposure and work with low exposure side by side. Each is defined by a list of job titles in nine languages, fixed before the counting, — 744 search terms and 39 exclusions, fixed before the counting began, including terms that were tested and rejected. Each sits in one of seven groups derived from ISCO-08.

Coverage is 24 markets, and it is deliberately uneven. Hiring is advertised online to very different degrees from one country to another, so the number of markets in which a given kind of work is visible ranges from 7 to 24 — and one kind of work sits below our own eligibility threshold, which is disclosed below. Every page states how many markets it could see.

Three restrictions

The rules are fixed before the counting, and they are restrictive. These three do more to explain what a SHIFT figure is worth than any assurance we could give.

There is a floor, set in advance, and it governs publication rather than inclusion. A market is shown for a kind of work only once it clears 500 advertisements in the base window. Below that, a single cell is too small for us to be willing to report it on its own. The number is a minimum-volume rule we chose and fixed before counting — it is not derived from an estimate of sampling error, and we do not claim it is. It applies to every kind of work equally, so showing or withholding a market is never a judgement made under pressure.

Cells below the floor are not deleted from the aggregate. They stay in the market basket that every occupation is measured against, carrying their own small weight, so that each kind of work is compared against the same geography as every other. Dropping them would give each occupation a different denominator and make the occupations incomparable with one another. Across the index the sub-floor share of an occupation's numerator has a median of 0.82% and a maximum of 9.3%, and the largest single sub-floor cell holds 496 advertisements — so no small cell can drive a result.

One kind of work in Edition 1 sits below our own eligibility threshold. Our admission rule requires an occupation to clear the floor in at least eight markets; plumbers & pipefitters clears it in 7. The rule forbids changing a closed occupation list because of what the numbers turn out to say, so the pre-committed list is retained and the shortfall is disclosed instead. Removing that kind of work would change the reported basket trend by 0.08 percentage points. The admission rule has been clarified prospectively for future editions.

We never compare one country against another, for the reason set out above: the raw counts do not mean the same thing in two places.

We removed a large source from the count because it entered the index halfway through the period. A source that arrives mid-series makes growth that happened in our sight of the market look like growth that happened in the market. Leaving it in would have produced larger numbers and a false reading. Throwing data away is what protects the comparison, and it is the reason the source list is narrower than it could be.

What we can and cannot see

Almost all of what we count — 97.0% — comes from a single site. Two markets add a second one. A short source list is the price of counting both years the same way, and we would rather state it than have it found.

In 4 of the 24 published markets, at least a quarter of what we counted in the base window sits in occupation cells too small to publish on their own. Their direction is still published, because it is corroborated market by market against independent national sources — but in those, read the country as a direction and not as a measurement.

Two further markets were collected and are not published: under the source recipe this edition uses, our provider returns no advertisements for them at all — not few, none — so there is nothing to compare.

Independent work on AI and hiring points in more than one direction, and some of it complicates our own reading. We take that seriously rather than citing only what agrees with us. We are available to demonstrate how a particular figure was produced.

The AI lens — honestly

Artificial intelligence is an interpretive lens, not a statistical conclusion. Changes in advertised demand may reflect the economic cycle, hiring practices, platform coverage, occupational reclassification, automation or other factors. Any discussion of causation is clearly identified as commentary and is not presented as a result proved by the index. This index is called the Demand-for-Work Index because that is what it measures: where advertised demand moved. AI exposure is the lens we used to choose which kinds of work to put side by side — it is not a finding about cause. Independent work on AI and hiring points in more than one direction, and we say so above.