Read this first
What this edition is, and what it is not
Edition 1 is corpus weighted. Every figure below is a property of job postings, read from 2,151,213 active postings measured on 2026-08-11. No figure describes a person, and this edition carries no aggregate of assessment results at all.
The reason is worth stating plainly rather than leaving to be worked out. An aggregate of assessment results publishes only when enough people sit inside each bucket that nobody in it can be identified, and that floor is 5. Early on, almost nothing clears it. An index that quietly dropped the thin buckets and published the rest would be reporting whichever slice happened to be large, which is a different measurement wearing the same name.
So this edition reads the posting corpus instead, says so on its face, and publishes the floor anyway, so the next edition can be checked against this one rather than believed.
https://olive.is/benchmarks · Edition 2026 Q3The floor, and what it withheld
No bucket publishes with fewer than 5 people in it.
Buckets withheld by that floor in this edition: unknown, not zero. The site build holds no read path to assessment results; aggregates are served server-side by the engine, which applies the floor before anything leaves it.
Recording an unmeasured thing as a zero is how a later edition ends up reporting growth from a baseline nobody took. Unknown stays unknown until something reads it.
2026-02 to 2026-07
What the postings say
1.4% to 8.1%
of active postings carried an explicit AI-usage requirement in 2026-07, read two ways over one corpus of 2,151,213 postings
3.0x and 2.1x
growth in the requirement rate from 2026-02 to 2026-07, on the extracted-skill and raw-text instruments
79.7%
of the postings carrying an AI-usage requirement sit in Technology, 27,697 of 34,743
every figure carries its limit. Both instruments measured on 2026-08-11
The series
The requirement rate by month, both instruments
| Month | Postings | Extracted skills | Raw text |
|---|---|---|---|
| 2026-02 | 445,637 | 0.5% | 3.8% |
| 2026-03 | 2,559,476 | 0.5% | 3.2% |
| 2026-04 | 993,788 | 0.7% | 4.5% |
| 2026-05 | 1,429,361 | 0.8% | 6.8% |
| 2026-06 | 1,023,700 | 1.1% | 8.9% |
| 2026-07 | 998,166 | 1.4% | 8.1% |
The partial month after 2026-07 is excluded from this table and from every figure on this page. It is the largest reading in the series and it sits on a fraction of a normal month of volume, so publishing it would be quoting a collection artifact as a trend.
Findings
Five readings, each with the reason it might be wrong
The bracket
1.4% to 8.1% of active postings carried an explicit AI-usage requirement in 2026-07, read two ways over one corpus of 2,151,213 postings.
The two instruments disagree by roughly six-fold and both ends publish. The lower reads extracted skills against a frozen term list and understates by an unknown amount; the higher scans the whole posting text and catches wording that is not always a requirement. Neither end is presented as the number.
The climb
3.0x and 2.1x growth in the requirement rate from 2026-02 to 2026-07, on the extracted-skill and raw-text instruments.
A posting-prevalence trend moves when employers change their minds and also when which employers are hiring changes. The direction carries; the level does not. The partial month after 2026-07 is held back rather than quoted, because it is the largest figure in the series and sits on a fraction of a normal month of volume.
The concentration
79.7% of the postings carrying an AI-usage requirement sit in Technology, 27,697 of 34,743.
Concentration inside the cohort, not the rate inside the industry, which is a different and lower number. The two disagree: Financial Services carries the highest rate at 7.3% of its own postings while holding a much smaller share of the cohort.
The seniority gap
72.9% of the AI-usage cohort asks for five or more years of experience, against 47.8% of all postings; the junior end runs 9.2% against 32.0%.
Read off a minimum-years field that the source system documents as inflated upward, and that is filled on about half of all postings. The gap between the two columns is the finding; neither column is a population estimate.
The missing field
0.0% of active postings populate the structured field for an AI-materials policy, which is the field an applicant would need to know what is permitted.
This is a fact about the data source, not about employers. The field is scaffolded in the posting schema and unused, so a posting that states a policy in its prose still reads as empty here. What it establishes is narrow and worth stating: the structured signal does not exist, so nothing downstream can be filtering on it.
The method, and both term lists
2,151,213 active postings, measured 2026-08-11. Series runs 2026-02 to 2026-07. The partial month at the end of the series is excluded from every figure.
Two instruments, not one
A posting is counted two ways, and the two disagree by roughly six-fold. Both ends publish and neither is presented as the number. The first reads a frozen term list against the skills a posting declares, and understates by an unknown amount because a posting the extractor did not fire on cannot be counted. The second scans the whole posting text, which catches wording the first misses along with some that is not a requirement at all.
Instrument one: the extracted-skill term list
A posting counts when its extracted skills intersect this set:
- ai fluency
- ai tools
- generative ai tools
- ai assisted development
- ai assisted development tools
- ai enabled tools
- ai tooling
- prompt engineering
- github copilot
- copilot
- chatgpt
- claude
- claude code
- ai adoption
- ai integration
Excluded on purpose, as build-the-model rather than use-the-model. This is the most contestable part of the definition, which is exactly why it publishes beside the inclusions instead of being left implicit:
- machine learning
- artificial intelligence
- ai ml
- llm
- agentic ai
- ai agents
Instrument two: the raw-text phrases
A phrase match over the full posting description, bypassing the skill extractor, which is what makes it an independent reading rather than a second look at the same field:
- AI tools
- generative AI
- ChatGPT
- Copilot
- AI fluency
- prompt engineering
- AI-assisted
The aggregation rule
No bucket publishes with fewer than 5 people in it, and the number of buckets withheld publishes beside the ones that survive, so an absence is visible instead of silent. In this edition no assessment aggregate was read at all, so the withheld count is unknown rather than zero.
What this corpus cannot see
A posting-prevalence trend moves when employers change their minds and also when which employers are hiring changes, so the direction carries and the level does not. Role tags are free text and only about half of the cohort maps to an occupation code; the rest publishes as an unclassified bucket rather than being dropped. And a requirement stated in prose but not in a structured field reads as absent here, which is a fact about the data source rather than about employers.
Cadence, and how to check this
The Index publishes quarterly. An edition is named for the quarter of the data it reports rather than the quarter it goes up, so a late edition never re-labels a measurement. This one reads 2026 Q3.
Every figure above can be reproduced from the two term lists and the corpus definition on this page. If you disagree with a definition, the disagreement is about a list you can read rather than about a number you cannot.
Journalists working on hiring and AI can reach the desk at press@olive.is. The corpus pages behind this edition are benchmarks and the evidence ledger; the role hubs are at profiles by role.
2,151,213 postings, measured 2026-08-11