Everybody keeps saying the Generative AI job market is on fire. Fine — but almost nobody tells you what those jobs actually ask for once you get past the buzzwords. So I stopped guessing and pulled 95 live "Generative AI" postings from Google Jobs (US), then counted what really shows up in the titles and the full descriptions. The interesting part wasn't the individual skills — it was which ones keep showing up together. That's the bit that should change how you prep.
Here's what I found, receipts included.
TL;DR
- A "Generative AI" job today is mostly an applied engineering job: building with existing models, not training new foundation models from scratch.
- Everything is converging on one thing: agentic RAG systems. RAG and agents barely show up apart anymore — 78% of the "agent" postings also mention RAG.
- The core stack — LLM + Python + RAG — appears together in ~36% of postings (34 of 95), the highest-leverage combo in the sample. Layer on agents and an orchestration framework and you cover most of the frequently recurring combinations.
- The money is real: among the 23 postings that disclosed pay, the median range-midpoint was about $187k, with individual midpoints spanning $121k–$274k.
- Reality check nobody hands you: only 13% of these roles are at Big Tech. Nearly half the market is a long tail of companies you've genuinely never heard of.
One caveat on scope: this is 95 US postings from Google Jobs, captured mid-2026. Treat it as a trend read, not a census. The US is the biggest, most mature GenAI labor market, so it's a decent leading indicator — but the numbers will look different in the EU, India, and elsewhere. Directional, not gospel.
And one on method: every percentage below measures what employers foreground in a description, not a verified hard requirement. A term can appear as "required," "a plus," or even "the legacy thing we're replacing." Read these as what the market talks about, not a checklist every employer enforces. (Full methodology at the end.)
The skill tiers: what to learn, what to skip
I split every skill by how often it shows up across the 95 postings. Read this as signal strength, not a set of hard gates:
| Tier | Skills | What it means for you |
|---|---|---|
| Core signals (>50%) | LLMs (75%), Python (61%), RAG (61%), Prompt engineering (54%), Agents (54%) | Appear in the majority of postings — the safest foundational bets |
| Common differentiators (15–40%) | Fine-tuning (38%), LangChain (34%), MLOps (22%), Vector DBs (22%), Multimodal (20%) | Frequently useful, less universal — these break ties in your favor |
| Narrower signals (<15%) | Role-specific tooling, niche domains | Relevant to specific segments, not the broad market |
One thing sits outside the skill tiers because it isn't a skill: a US security clearance appears in ~13% of postings. It's an eligibility/access requirement, not something you learn — but if you have one, it's a genuinely low-competition lane (more below).
If you're short on time, get the core-signal five solid first. They're common enough to be the highest-expected-value things to know — a deep grasp of fine-tuning won't help much if you can't stand up a RAG pipeline in Python.
The thing the raw keyword counts hide: RAG and agents are now one skill
This is the finding I most want you to sit with. Look at which skills land in the same posting:
- Of the 51 postings that mention "agent," 40 also mention RAG (78%).
- Of the 58 that mention RAG, that same 40 also mention agents (69%).
- 92% of agent postings also mention LLMs; 51% specifically name LangChain.
- 66% of RAG postings also mention prompt engineering.
So there's a hard core of ~40 postings — about 42% of the whole sample — that foreground RAG and agents together. The market isn't describing "a RAG person" or "an agent person." It's describing someone who ships agentic RAG — retrieval-grounded systems that then go do things through tool calls and multi-step planning. Build one project that does both and you're suddenly speaking the exact language most of these listings are written in.
The stack you should actually build
Frameworks matter, so let's name names instead of waving our hands at "AI tools":
- Orchestration: LangChain (34%) is the default. LlamaIndex (11%) is the RAG-specialist pick. 34% name at least one — so learn LangChain first, and pick up LlamaIndex too if the role is retrieval-heavy.
- Model providers: OpenAI (29%) leads; Anthropic/Claude (~12%) is the clear, growing #2. Honestly, employers care less about which SDK and more that you've dealt with the ugly parts — token limits, streaming, evals, latency, cost.
- Retrieval layer: vector DBs + embeddings show up in ~22%. Know how chunking, embedding models, and a vector store (pgvector, Pinecone, Weaviate, take your pick) actually fit together.
- Ship it: MLOps (22%), Docker (22%), Kubernetes (21%). A model sitting in a notebook doesn't get you hired. A deployed endpoint does.
- Languages: Python (61%) isn't optional. Java (14%) and TypeScript/React (~6–11%) turn up in the more product-facing roles.
The one portfolio project that covers most of this: a Python service that does RAG over a real corpus (chunk → embed → vector store → retrieve), wraps it in an agent with tool use via LangChain, calls OpenAI or Claude, has some basic evals, and ships in a Docker container. That single build hits the core-signal five plus three differentiators.
Pick a cloud — but it's not winner-take-all
Cloud isn't optional, and the order is clear: AWS (40%) > Azure (35%) > GCP (25%), with Vertex AI showing up in ~9%. Here's the twist though: 29% of postings mention 2+ clouds. Multi-cloud fluency is a differentiator all by itself. Starting from zero? AWS has the widest coverage here, so it's the safe default — but for Microsoft-heavy enterprise and consulting shops, Azure is often the smarter first pick. Then get comfortable in a second.
Titles and seniority: there's room in the middle
- Engineer (34) is the default title, then Scientist/Research (15), Developer (10), Architect (8), and Manager/Lead (8).
- Only ~18% are explicitly "Senior." Most postings don't even bother specifying a level.
Translation: you don't need "Staff" on your résumé to get a foot in the door. Mid-level GenAI hiring is wide open right now.
Where the jobs actually are: clustering the employers
I bucketed all 95 companies into rough sectors by company identity. The buckets are approximate and not perfectly exclusive — a defense consultancy could arguably sit in two — but the shape is the expectations reset almost nobody gives you up front:
| Company cluster | Share | Examples |
|---|---|---|
| Other mid-market, specialist & less-recognized employers | ~47% | Mid-size firms, boutiques, contractors, subsidiaries |
| Big Tech / enterprise product | ~13% | Adobe, Oracle, NVIDIA, AT&T, DoorDash |
| Consulting & system integrators | ~9% | Deloitte, Booz Allen, c1advantage |
| AI-native companies / startups | ~9% | Kendia.AI, Innodata, EnthuZiastic |
| Finance & banking | ~5% | JPMorgan, capital-markets firms |
| Staffing / recruiting agencies | ~5% | US Tech Solutions, SVAM, Saransh |
| Semiconductors / hardware | ~4% | Sandisk, Applied Materials, Infinite Electronics |
| Academia | ~4% | Harvard and university labs |
| Defense primes | ~2% | Peraton (plus the clearance roles below) |
Two things jump out here:
- The long tail is the market. Around 47% of postings are at companies that aren't consumer-recognizable brands — some are large contractors or subsidiaries, just not names you'd rattle off. If you only apply to FAANG and the hot AI startups, you're fighting over roughly 22% of the openings — against everyone else doing the exact same thing. The real volume, and less competition, lives in the mid-market, the consultancies, and the contractors.
- Clearance is a parallel market. About 13% of roles want a US security clearance (that cuts across defense primes, consultancies, and gov contractors). Got one? It's your fastest, least-crowded way in. Don't have one? Filter those out and stop burning applications on them.
If your mental image of a "GenAI job" is a FAANG research lab, recalibrate. Write your résumé for delivery and business impact — that's what the consultancies and mid-market shops are actually buying — not just model benchmarks.
One more quiet signal from the listings: remote-first roles were uncommon in this sample — only ~10% prominently advertised remote. That doesn't automatically make the rest on-site (plenty are hybrid or just don't say), but if you're remote-only, you're fishing in a smaller pond here. Plan for it.
Your actual to-do list
- Build one agentic RAG app, end to end (Python + vector DB + LangChain + OpenAI/Claude + Docker). That single project matches the core-signal five and several differentiators.
- Add evals and cost/latency handling — the boring 20% that separates a demo from a hire.
- Pick a cloud that matches your target sector. AWS has the broadest coverage here (40%), but Azure is often the better first bet for Microsoft-heavy enterprise and consulting roles. Then get literate in a second.
- Apply to mid-level roles broadly. Don't self-filter for "Senior."
- Decide on the clearance lane. If you've got one, lead with it — it's your fastest, least-competitive path in.
The GenAI market rewards people who've actually built and shipped, not people who've read about transformers. Pick one end-to-end agentic-RAG project, make it real, and let these 95 job descriptions be your syllabus.
How I analyzed the postings
Because this whole post is built on original data, here's exactly what I did so you can weigh it:
- Source & date: Scraped from Google Jobs on 2026-07-10 via a scraper Actor. Query "Generative AI", location United States, country US — a fixed query, not personalized to my own location or search history.
- Sample: 95 postings returned. A mix of full-time, contract, and part-time. I did not de-duplicate repeat employers, so a few companies appear more than once (e.g. Innodata ×4, Deloitte ×3) — that reflects real posting volume but slightly weights active hirers.
- Keyword prevalence: case-insensitive substring match across each posting's title + full description. Variants group naturally — "agent" catches agents/agentic/AI agents; "fine-tun" catches fine-tune/fine-tuning. Each percentage is the share of the 95 postings whose text contains the term.
- Mention ≠ requirement. This measures what employers foreground in the text, not verified hard requirements. A term can appear as required, "a plus," or context. Treat the numbers as what the market talks about.
- Co-occurrence counts are raw intersections (e.g. 40 of 95 postings contain both "agent" and "RAG").
- Salary: only 23 of 95 postings disclosed a range. I took each range's midpoint; the reported figures are the median of those midpoints ($187.5k) and their min/max ($121.2k–$274.1k). This subset may be biased toward employers, states, or role types more likely to publish pay.
- Employer sectors: assigned by company identity; buckets are approximate and non-exclusive. The ~47% "other" bucket = companies that didn't match a named sector — not a claim that they're small.
- Known limits: single query, single day, US-only, one job board. Good for an exploratory trend read; not a census of the GenAI labor market.
Top comments (0)