How the numbers are made, and how wrong they are.
Every number on this site comes from rules that read job ads. Here is what they read, what they decide, what they never keep, how often they get it wrong, and the data itself to download and check.
- 01Read205 company career pages and job boards, once a day, through their public job feeds.
- 02KeepOnly ads whose title names an AI role, and only technical ones: no sales, marketing, recruiting or legal titles.
- 03Count onceAn ad carried by a job board and the company's own page counts once: 11 reposts dropped in the latest run.
- 04Read the adRole and level from the title; skills, tasks, tools, stack and years from the requirements, with benefits and boilerplate set aside.
- 05Keep only the answersThe ad text is read and dropped. What stays is the role, level, skills, tasks, tools and dates.
- 06Check by handEvery month a sample is labelled by hand, blind to what the rules decided, and the error rates below are recomputed.
How often the rules are wrong.
A sample of 200 ads is being labelled by hand (0 done so far): 80% from the ads the rules kept, spread across roles, and 20% from tech ads they left out. The rates appear here as soon as it is scored, with every month after it.
Until then, read the numbers on this site as what the rules found, not yet as measured fact.
What decides each field.
The title must name AI work (AI, ML, LLM, agents, RAG, NLP, computer vision, forward deployed, research scientist and similar), name a technical job (engineer, architect, scientist, researcher, developer) and not be a sales, marketing, recruiting, legal, finance or management title. The description alone is never enough.
Matched on the title in a fixed order, first match wins: forward deployed, AI-native architect, solutions architect, agentic, platform, research, AI-native developer, applied AI engineer. A title that matches none is "other AI role".
From words in the title: intern, junior, head, principal, staff, lead, senior. A title that says none of them counts as mid level.
The ad is split at its headings. Tasks come from what you will do, skills and required tools from what is asked of you, preferred tools from the nice-to-haves, and benefits, company pitch and legal text are ignored. Years are the smallest "N+ years of experience" in the requirements.
- Role, level, remote or not, skills, tasks, tools, stack and years asked
- The platform it came from and the dates it was published, first seen and last seen
- A normalised job title, for counting titles
- The ad text: it is read and dropped
- Links to ads, and company names
- Ads from people looking for work, which can carry a name
- Mostly English-language ads from companies whose career pages run on the platforms read: tech companies and the US are over-represented.
- Role and level come from the title alone, so a vague title gives a vague answer.
- Keyword rules miss skills that are described without the usual words, and can count a word used in passing.
- Tracking began on 7 November 2026: earlier ads count only if they were still listed then. Some platforms give an old ad its original date, years back.
Download it, check it, cite it.
One row per AI job ad, derived fields only: 2,488 ads in version 2026-11-13, updated on the 1st of every month. Free to use with credit, under CC BY 4.0.
Faraji, V. (2026). Applied AI job ads: what employers ask for (version 2026-11-13) [Data set]. https://vfaraji89.github.io/applied-ai-simplified/market/method/
Licence: CC BY 4.0, for the dataset only. The code and the site are not open source.