Two counts, one bulletin
The employment rate for people aged 16 to 64 was 75.1% in April to June 2026. Unemployment was 4.9%. Economic inactivity was 20.9%. Read together, those three describe a labour market that is doing roughly what it has been doing.
In the same publication: the number of payrolled employees fell by 78,000, or 0.3%, between June 2025 and June 2026. Between May and June it fell by 13,000, which the ONS rounds to 0.0%.
A reader could be forgiven for thinking one of those has to be wrong. Neither is. They are measuring different things, with different instruments, and this is a month where that distinction stopped being academic.
| Reading | What it counts | What it said |
|---|---|---|
| Employment rate, 16 to 64 | Share of the working-age population who did paid work in the reference week, from a household survey. Includes people who work for themselves and people nobody runs a payroll for. | 75.1% in April to June 2026 |
| Payrolled employees | Count of payroll records held by the tax authority. Not a survey and not an opinion: an administrative record of who was actually paid through a payroll. | Down 78,000 (0.3%) on the year to June 2026; down 13,000 (0.0%) on the month |
| Unemployment rate, 16 and over | Share of the labour force without work who are available for work and actively seeking it. Excludes anyone not looking, however much they might want a job. | 4.9% in April to June 2026 |
| Economic inactivity rate, 16 to 64 | Share of the working-age population who are neither in work nor looking for it. The category that holds everyone the unemployment rate leaves out. | 20.9% in April to June 2026 |
| Vacancies | Estimated number of unfilled positions employers are actively trying to fill. An early estimate, published ahead of the fuller figure. | 707,000 in May to July 2026, down 6,000 (0.8%) on the quarter |
One asks households, the other reads payslips
The employment rate comes from a survey. Interviewers ask people what they did in a particular week, and someone who did paid work is counted as employed - whoever they worked for, and whether or not any organisation runs a payroll on their behalf.
The payrolled-employee count comes from tax records. It is not asking anybody anything. It counts payroll records: people who were actually paid through an employer's payroll in that month.
So the two are not rival estimates of one quantity. "Is this person working?" and "is this person on a payroll?" are different questions, and a labour market can move the two apart without either measurement failing.
What this bulletin does not let us do is say how much of the gap is which. It reports no figure that decomposes the difference, so we are not going to offer you one. The honest version is that the payroll count fell while the survey-based employment rate held, and that the reason is not settled by this document.
Why the distinction reaches a job advert
Almost every job you can apply to on a job board is an offer of a payrolled job. Someone is proposing to put you on their payroll. That means the count that fell is the count that describes the thing a job advert is actually offering.
Vacancies sit alongside it. The early estimate for May to July was 707,000, a fall of 6,000, or 0.8%, on the previous quarter. Not a collapse. Not a recovery either.
This is worth stating plainly because a matching system - ours included - sees none of it. We can read a posting. We cannot read whether the payroll behind it is growing or shrinking, and neither can you from the advert.
The third number disagrees with itself
Pay grew 3.5% over the year for regular earnings, which exclude bonuses, and 4.1% for total earnings, which include them.
To turn that into what it bought, you divide by a price index. The ONS publishes two. Using CPIH, real growth was 0.5% for regular pay and 1.1% for total pay. Using CPI, the same pay growth becomes 0.7% and 1.3%.
That is not a rounding wobble. It is the difference between describing pay as barely ahead of prices and describing it as comfortably ahead, and it turns entirely on which index a writer picks. Anyone quoting a single real-pay number without naming the index has reported a choice as though it were a fact.
| Pay measure, and the index used | Real growth on the year |
|---|---|
| Regular pay, deflated by CPIH | 0.5% |
| Regular pay, deflated by CPI | 0.7% |
| Total pay, deflated by CPIH | 1.1% |
| Total pay, deflated by CPI | 1.3% |
What the ONS says about how much of this to carry
More than most statistical agencies, the ONS spends this bulletin telling readers where its own numbers are soft, and those warnings belong in any piece that quotes it.
On the survey: "Caution should be taken when drawing conclusions from short-term changes, and we advise users to focus on long term movements in the data." It adds that volatility remains, particularly for smaller breakdowns where sample sizes are smaller.
On the survey's collection: an operational issue in May 2026 led to temporary under-resourcing in the telephone collection operation.
On the most recent payroll reading: figures for July "should be treated as provisional estimates and are likely to be revised when more data are received next month". The claimant count for the latest month is provisional too.
None of that makes the bulletin unusable. It makes the twelve-month movements more load-bearing than the monthly ones, which is why the figure we have led with is a year-on-year change rather than a month-on-month one.
What we take from it
We run a matching product, so our interest here is narrow and worth admitting. When someone tells us the labour market is fine, or that it is falling apart, the first question worth asking is which instrument they read, because this month those instruments disagreed in public.
The second is whether the number is a level or a change. Most arguments about the job market are two people quoting different kinds of number at each other.
We publish how our own matching is tested, including where it falls short, for the same reason we print the ONS's caveats rather than the headline alone.
We test our own matching for bias and publish what the tests find, including the parts that do not flatter us. How we test our matching is the same discipline as printing a statistician's caveats instead of only their headline.