Tony Wang13 min readAI Tool Mentions in Job Postings Are Accelerating — Even at the Same Employers
Claude, ChatGPT and OpenAI mentions run 2–5x higher in the newest postings than an older slice of the same employers. Not a startup-mix or text-length artifact.
Search a jobs dataset for “ChatGPT” or “Copilot” and you’ll hit a wall fast: past a few thousand matches, the API can’t tell you how many more there are — Elasticsearch stops counting past a 10,000-result window, the same limit that trips up every naive keyword search in this series. Worse, plenty of the “AI” hits aren’t about AI at all: “Copilot” catches aviation co-pilot listings, “LLM” catches Master-of-Laws postings, and bare “AI” now shows up in boilerplate hiring-disclosure paragraphs regardless of the role.
Fix the counting method — a documented sample instead of a capped total, word-boundary matching with explicit disambiguation instead of raw substrings — and a real signal survives: the newest job postings mention AI tools substantially more than older postings from the very same employers. Not because trendy startups post more often. Not because recent listings just have longer text. The effect holds up inside a single ATS provider’s company population, and it’s strongest exactly where you’d expect — Ashby, the ATS startups and AI labs actually use.
The obvious approach breaks first
Ask "how many job postings mention GitHub Copilot?" and the natural move is search?q=GitHub Copilot and read total. That number comes back 10,000+, capped — as does Perplexity, Midjourney, "AI agent," bare "LLM," and nearly every other term we tried except a handful of distinctive single brand names.
The cap isn’t the only problem. Ambiguous terms drown in false positives long before they’d hit it. We tested this directly: in an 8,000-posting sample, the raw word "Cursor" matched 26 postings — almost entirely the mouse-cursor sense in UI/QA job descriptions, not the AI code editor. Requiring real AI context nearby ("AI," "code editor," "IDE") cut that to 1. "Copilot" alone matched 32; requiring "GitHub," "Microsoft," or "AI" nearby brought it to 25 — a smaller but still real correction, since aviation and generic-idiom "copilot" postings exist too.
| Term | Naive word-boundary match | AI-context required | What the naive count was catching |
|---|---|---|---|
| Cursor | 26 (0.33%) | 1 (0.01%) | Mouse-cursor UI/QA language, almost entirely |
| Copilot | 32 (0.40%) | 25 (0.31%) | Aviation co-pilot roles, generic idiom |
That’s why this study leans on a sample with disambiguation rules instead of a live search total — the total either can’t count past 10,000, or counts the wrong thing.
What’s mentioned across the whole open-postings stock
Restrict to the handful of AI brand names distinctive enough to stay under the search cap, and the full corpus — every open posting, not just recent ones — gives exact counts. ChatGPT leads by a wide margin.
As a share of all 1,859,828 open postings: ChatGPT 0.48%, OpenAI 0.38%, Anthropic 0.23%, Claude 0.19%, "Generative AI" 0.05%, Gemini 0.03%. Small percentages — but this is the stock of everything currently open, accumulated over however long each posting has sat live. It says nothing about direction. For that, you need the newest slice.
The newest postings tell a different story
Sort the same provider’s postings by "most recently posted" instead of by company name, and AI-tool mentions jump — consistently, across every AI-heavy provider we checked.
| Provider | Claude | ChatGPT | OpenAI | "Generative AI" |
|---|---|---|---|---|
| Ashby — recent | 7.8% | 4.8% | 4.4% | 3.6% |
| Ashby — control | 3.5% | 1.2% | 0.9% | 3.5% |
| Greenhouse — recent | 2.1% | 0.9% | 0.8% | 1.6% |
| Greenhouse — control | 2.0% | 1.1% | 0.2% | 0.5% |
| Workday — recent | 0.5% | 0.4% | 0.1% | 0.7% |
| Workday — control | 0% | 0.1% | 0% | 0.1% |
| Oracle — recent | 0.1% | 0.2% | 0% | 0.3% |
| Oracle — control | 0.1% | 0% | 0% | 0% |
| SmartRecruiters — recent | 0.1% | 0.1% | 0.1% | 0% |
| SmartRecruiters — control | 0% | 0% | 0% | 0% |
| Lever — recent | 0.4% | 0.1% | 0.1% | 0.6% |
| Lever — control | 0.3% | 0% | 0.2% | 0.7% |
Ashby is the clearest case. Claude mentions run 2.2x higher in the newest postings (7.8% vs 3.5% — 78 vs 35 hits, large enough samples to trust), ChatGPT 4x (4.8% vs 1.2%), OpenAI 4.9x (4.4% vs 0.9%). Ashby is the ATS most startups and AI labs use, so this is close to "AI companies talk about AI tools in their own job ads more than they used to" — not a shocking finding on its own, except that the same direction shows up at Greenhouse and even Workday, ATS providers with a much more conventional enterprise customer base. It's absent at SmartRecruiters and Lever, where AI mentions stay near zero regardless of recency — the effect isn't universal, it's concentrated where you'd expect AI-forward employers to sit.
Ruling out the boring explanations
Two objections are obvious enough to check before trusting this.
Is "recent" just secretly "more Ashby/Greenhouse, less Workday"? A recency-sorted pull across all providers does have a skewed mix — but not in the direction that would explain the effect:
| Provider | Share of recent-sorted sample | True corpus share | Ratio |
|---|---|---|---|
| SmartRecruiters | 44.3% | 13.7% | 3.24x over-represented |
| Oracle | 15.2% | 11.2% | 1.35x over-represented |
| Ashby | 2.7% | 2.6% | 1.03x, roughly matched |
| Greenhouse | 8.1% | 9.4% | 0.86x under-represented |
| Lever | 2.1% | 3.2% | 0.68x under-represented |
| Workday | 13.2% | 46.6% | 0.28x under-represented |
A recency-sorted pull is dramatically over-weighted toward SmartRecruiters (3.24x) — a provider with near-zero AI mentions regardless of recency — and under-weighted toward Workday (0.28x) and Greenhouse (0.86x). If anything, this mix shift should have diluted the naive topline AI-mention rate, not inflated it. The within-provider comparison above, which holds the provider fixed, is the real test — and it survives.
Is "recent" just secretly "longer, more fully-scraped description text"? More text mechanically increases the odds any given keyword appears somewhere in it, independent of actual content.
| Provider | Recent — avg chars | Control — avg chars | Difference |
|---|---|---|---|
| Ashby | 5,372 | 5,308 | +1.2% |
| Greenhouse | 5,531 | 5,021 | +10.2% |
| Workday | 4,464 | 5,658 | −21.1% |
Ashby is essentially flat (+1.2%) despite a 2–5x AI-mention gap. Greenhouse's +10.2% is real but far too small to explain a 3–4x rate difference on its own. Workday's recent postings are 21% shorter, not longer — the wrong direction for a length artifact — yet still mention AI tools more. Neither confound survives contact with the data.
Skills, for reference
The same sampling method sidesteps the 10,000-result ceiling for general skill terms too — every one of 17 common skills we checked via a raw search total hit that cap, useless for ranking anything. In the 8,000-posting sample, word-boundary matching gives real numbers:
| Skill | Postings | % of sample |
|---|---|---|
| Excel | 722 | 9.03% |
| Python | 433 | 5.41% |
| Project Management | 385 | 4.81% |
| AWS | 276 | 3.45% |
| SQL | 269 | 3.36% |
| Salesforce | 216 | 2.70% |
| Machine Learning | 192 | 2.40% |
| Azure | 164 | 2.05% |
| Kubernetes | 155 | 1.94% |
| Java | 104 | 1.30% |
| TypeScript | 101 | 1.26% |
| React | 101 | 1.26% |
| Power BI | 95 | 1.19% |
| JavaScript | 91 | 1.14% |
| Docker | 89 | 1.11% |
| Google Cloud | 68 | 0.85% |
| Git | 63 | 0.79% |
| Tableau | 56 | 0.70% |
| Node.js | 36 | 0.45% |
Excel outranking Python is a useful check on itself: this dataset spans retail, healthcare, and field-services hiring alongside tech (the job postings dataset breaks the same volume out by department label), so a skills ranking dominated by office-suite basics is exactly what an honest, unfiltered sample should look like.
What this is not
- Not a dated time series. The public API has no date-range filter; "recent" vs "control" is a recency proxy from two sort orders, not a monthly or weekly trend curve.
- Not proof Claude has overtaken ChatGPT. ChatGPT still leads the full open-postings stock by a wide margin (8,854 vs 3,561). The recency effect is a growth signal on top of that stock, not a reversal of it — yet.
- Not a "% of jobs require AI skills" claim. Bare "AI" matches 17.06% of the 8,000-posting sample, but our boilerplate-stripping list only caught 0.03% of postings — meaning that number is very likely still inflated by EEO/AI-hiring-disclosure paragraphs many ATS templates append regardless of role, and we don't have a complete-enough boilerplate library to clean it. We report it, flagged, rather than pretend it's clean.
- Not a causal claim. More mentions of AI tools in job ads is not the same as more AI adoption, more AI hiring, or higher AI skill requirements — it's what employers chose to write down.
Who this helps
Analysts / data eng: if you're building any kind of keyword-prevalence metric from full-text job data, check your denominator against the search API's result-window cap before trusting a total, and build disambiguation rules for any term that collides with unrelated common usage — this post's whole method exists because "Cursor" alone would have been garbage.
Recruiting / employer-brand teams: the AI-mention acceleration is real but provider-dependent — it tracks which ATS an employer's category of company tends to use, not a universal shift.
Researchers citing the series: the recency effect is strongest and most reliable at Ashby; treat the Workday/Oracle numbers as directionally suggestive, not independently conclusive, given the small hit counts.
Query AI-tool and skill mentions across 14 ATS providers yourself
Search open roles, filter by provider and department, and pull the same postings this study samples — same REST surface, 2,000 free credits a month, no card.
Frequently asked questions
Has Claude overtaken ChatGPT in job postings?
Not in the full open-postings stock — ChatGPT still leads with 8,854 mentions versus Claude's 3,561 (snapshot 2026-07-30). But in the newest postings on Ashby, Claude is mentioned in 7.8% of listings versus 3.5% in an older cross-section of the same employers, a 2.2x recency effect that's larger for ChatGPT (4x) and OpenAI (4.9x) at the same provider. It's a growth signal on top of the stock, not proof of a reversal.
How do you know the recency effect isn't just more startups posting more often?
We checked. A recency-sorted pull is actually under-weighted toward AI-heavy Ashby (1.03x its true share) and Greenhouse (0.86x) versus SmartRecruiters (3.24x over-weighted, near-zero AI mentions) and Workday (0.28x under-weighted). If provider mix explained the effect, it would work in the wrong direction. The within-provider comparison — same employers, recent vs. an older cross-section — holds the mix fixed and the effect survives.
Could recent postings just have longer, more complete description text?
We checked that too. Ashby's recent and older samples have nearly identical average description length (5,372 vs 5,308 characters). Workday's recent postings are actually 21% shorter than the older sample, yet still mention AI tools more often — the opposite of what a text-length artifact would predict.
Why not just search for these terms and count the total?
Every common AI term we tested (GitHub Copilot, Perplexity, Midjourney, bare 'AI agent', 'LLM') hit the Jobs search API's 10,000-result ceiling, making a raw total useless for ranking. Ambiguous single words are worse: 'Cursor' matched 26 postings in an 8,000-posting sample, but 25 of those were mouse-cursor UI language, not the AI code editor — requiring real AI context nearby cut it to 1.
What percentage of jobs require AI skills?
This study can't answer that honestly. Bare 'AI' matches 17.06% of a sampled 8,000 postings, but known EEO/AI-hiring-disclosure boilerplate that many ATS templates append regardless of role inflates that number, and our boilerplate-stripping list only caught 0.03% of postings — nowhere near complete. We report the number flagged as unreliable rather than presenting it as a real skills-demand figure.
What are the most-mentioned general skills in job postings?
In the same 8,000-posting sample: Excel 9.03%, Python 5.41%, Project Management 4.81%, AWS 3.45%, SQL 3.36%, Salesforce 2.70%, Machine Learning 2.40%. Excel outranking Python reflects that this dataset spans retail, healthcare and field-services hiring alongside tech, not just software roles.