We Analysed 3,910 Real Job Postings. Sharing a Job Title Gets You a Quarter of the Way.

Two companies advertise the same job title. How much of the second one’s requirements does the first one’s list already cover? We measured it across 3,910 live postings and 6,298 same-title comparisons. The median answer is 25%. This page shows the number, the method, the control that had to pass before we published it, and the four earlier versions of this analysis that were wrong.

The findings in one place

The headline: a job title predicts about a quarter of the list

The question worth asking is not “how similar are two job postings” in the abstract. It is the one a job seeker actually faces: I tailored my resume to the last posting. How much of the next one does it already cover?

So that is what we measured. For every pair of postings that shared a normalised job title but came from different companies, we computed the share of posting B’s named skills that posting A also named — in both directions. Across 21 title groups and 6,298 measurements, the median was 25.0% and the mean 28.0%.

Put plainly: a job title tells you roughly a quarter of what a given employer will ask for. The remaining three quarters are specific to the posting in front of you.

Coverage by job title

Only titles with at least 8 postings and at least 20 cross-company measurements are shown. “Skills named” is the median count of tracked skills per posting in that group — it matters, because coverage measured over very short lists is noisier.

That last row is not a typo, and it is the most useful line in the table. For enterprise customer success manager, the median pair of postings from different companies shared none of the tracked skills. Same title, same seniority, same industry — and no named requirement in common.

How we checked the number was real before publishing it

A coverage figure on its own proves nothing. If the method counted ordinary English as a requirement, any two documents would “overlap” and 25% would be an artefact. So the analysis carries a control it has to pass:

The same measurement was run on pairs of postings with different titles at different companies — a Software Engineer against a Commercial Account Executive. If a shared job title means anything, same-title pairs must overlap clearly more than those. The pre-set threshold was 1.15×.

Every comparison in both rows crosses company boundaries. That matters more than it sounds: companies reuse a posting template, so two postings from the same employer share wording for reasons that have nothing to do with the job. An earlier version of this analysis failed to exclude them and produced a control that beat the finding — more on that below.

What employers actually name

Document frequency: the share of postings that name a term at least once. A posting saying “SQL” five times counts once.

The top of that table is worth sitting with. The single most-requested thing in 3,910 technology job postings is not a technology. “Cross-functional” appears in 44.3% of them — more than twice as often as AWS, and more than three times as often as SQL. “Roadmap” outranks every programming language in the corpus.

The things people are told are essential

This is where the corpus is most useful, because it contradicts a lot of advice. Each of these is widely described as a must-have, and each appears in under 4% of postings:

Two readings of this are both correct, and it is worth being clear about which one this study supports. It does not show that Scrum certification is worthless — a term can matter enormously in the minority of postings that name it. What it shows is that you cannot know which of these matters without reading the posting. Adding all of them to a resume in the hope of covering the field is how resumes become unreadable, and it still misses the specific list any one employer wrote.

By job family

Percentages are the share of postings within that family. Families are inferred from the job title.

Product management is the sharpest result here: 78% of product postings name a roadmap. If you are applying to product roles and the word does not appear on your resume, that is a gap worth closing before any of the technology gaps.

Method

Corpus. 3,910 postings collected on 6 August 2026 from the public Greenhouse job-board API (boards-api.greenhouse.io/v1/boards/{board}/jobs?content=true) — the documented, unauthenticated endpoint that powers each company’s own careers page. No HTML scraping, no rate evasion, no access controls bypassed. Postings under 400 characters were dropped as too short to contain requirements.

Companies. Affirm, Airbnb, Asana, Brex, Cloudflare, Datadog, Discord, Duolingo, Elastic, Figma, GitLab, Instacart, MongoDB, Reddit, Robinhood, Stripe, Twilio, Vercel. Counts range from 549 (Stripe) to 48 (Discord).

Vocabulary. 174 named skills, tools, methods and standards, published in full at the bottom of this page. Proper nouns are matched case-sensitively. Four terms that are also ordinary English — Go, Excel, React, Spark — additionally require a technical context word within 60 characters. “Java” is matched so that “JavaScript” cannot satisfy it.

Self-mentions excluded. Datadog’s own postings say “Datadog”; MongoDB’s say “MongoDB”. A vendor naming its own product is not an employer asking a candidate for a skill, so those mentions are dropped. Left in, they put two corpus vendors into the top twelve most-demanded skills.

Title normalisation. Seniority prefixes (senior, staff, principal, junior, II, III), parenthetical suffixes and trailing locations are stripped, so “Senior Software Engineer II (Remote, US)” and “Software Engineer” group together. Without this, groups fragment to two or three members and the overlap figure is noise.

Coverage. |A ∩ B| / |B| — the share of B’s named skills that A also names. Computed in both directions for every cross-company pair. Deliberately not Jaccard: Jaccard answers “how alike are these documents”, which is not the question a job seeker has.

Four versions of this analysis were wrong. Here is how.

Any study can be made to produce a striking number. These are the ones this one produced before it was correct, published because a method you cannot inspect is a method you should not trust.

• Ordinary English counted as technology. The first pass scored the Go programming language at 31.8% of postings. It was counting “go above and beyond” and “go-to-market”. Case-sensitive matching plus a context requirement brought it to 10.2%. Excel, React and Spark had the same defect.
• Sets too small to measure. The second pass asked the right question over a 40-term list, which gave a median of two terms per posting. Overlap on two-element sets is nearly binary — several title groups returned a median of 1.000 alongside a third of pairs sharing nothing at all. Arithmetically correct, analytically worthless.
• HTML counted as skills. Switching to a corpus-derived vocabulary put div, span, href and class into the requirement list at 30–36% document frequency. The collector stripped HTML tags before decoding entities, so any posting containing escaped markup had it turned back into real markup afterwards. Decode first, then strip, and repeat until stable.
• A control confounded by company. That same pass reported that same-title postings overlapped less than random pairs — the opposite of what any valid measure would show. Control pairs had been drawn from postings adjacent in an ID sort; IDs begin with the company name, so the “random” pairs were nearly all from the same employer, sharing a template. Requiring every compared pair to cross companies fixed it, and is why the published control is trustworthy.

Limits — read these before citing the study

• This is not “the job market”. It is 18 US technology companies that publish through Greenhouse. It over-represents engineering (1,475 of 3,910 postings) and under-represents healthcare, education, government, trades, retail and small employers — which is to say, most jobs.
• A vocabulary of 174 terms cannot capture everything a posting asks for. The median posting named 4 of them; 115 named none. Coverage figures describe overlap in named, trackable skills, not in everything an employer wants.
• Some terms carry more than one meaning and their figures should be read as upper bounds. “Pipeline” covers both sales pipelines and data pipelines. “Discovery” covers sales discovery and product discovery. “Audit” covers financial and security audits. “Accessibility” may be inflated by accommodation statements in posting boilerplate rather than by genuine WCAG requirements.
• A snapshot, not a trend. Everything here is one collection date. It cannot say whether any requirement is rising or falling.
• Title groups are uneven. The smallest published group has 9 postings. Small groups move a lot on a handful of documents; the number to lean on is the pooled median across all 6,298 measurements, not any single row.

What to do with this

The practical conclusion is narrow and, we think, well supported: a resume written against a job title is written against about a quarter of what any given employer asked for. Not because the writer was lazy — because the title genuinely does not contain the rest of the information.

The fix is not a longer resume or more keywords. It is reading the specific posting and checking your resume against that, once per application. That takes a few minutes by hand, which is why most people do not do it, and it is the entire job of the free tool on this site: paste a posting and your resume, get back the list of terms the posting uses that your resume does not. No score, no model deciding for you, no account.

Cite or reuse this

The figures on this page may be quoted, charted or rebuilt freely, including commercially, with attribution and a link. If you want to check the work rather than take it on trust, the collection and analysis are both plain Python of about 200 lines each and the API is public — write to us and we will send both scripts and the posting IDs.

Suggested citation: CareerLift (2026). What 3,910 Real Job Postings Actually Ask For. Collected 6 August 2026 from the Greenhouse public job-board API. https://careerlift.agency/job-postings-study/

The full vocabulary

All 174 terms counted, grouped for readability only — the analysis treats them as one flat set. If a term you care about is missing, that is a real limit of the study, not a finding about the term.

Languages (18): Python, Java, JavaScript, TypeScript, Go, Ruby, PHP, Scala, Kotlin, Swift, Rust, C++, C#, SQL, R, MATLAB, Perl, Elixir

Cloud & infrastructure (20): AWS, GCP, Azure, Kubernetes, Docker, Terraform, Ansible, Jenkins, CI/CD, Linux, Kafka, Redis, GraphQL, REST API, microservices, serverless, Datadog, Splunk, Prometheus, Grafana

Data (28): Snowflake, Databricks, Spark, Airflow, dbt, Hadoop, BigQuery, Redshift, Postgres, MySQL, MongoDB, Elasticsearch, ETL, data warehouse, data pipeline, machine learning, deep learning, NLP, TensorFlow, PyTorch, scikit-learn, pandas, A/B testing, experimentation, cohort analysis, regression, forecasting, statistical modeling

Analytics & BI (10): Tableau, Power BI, Looker, Mode, Amplitude, Mixpanel, Google Analytics, Excel, dashboards, data visualization

Design & front-end (14): Figma, Sketch, React, Vue, Angular, Next.js, CSS, HTML, accessibility, design system, user research, usability testing, wireframes, prototyping

Product & delivery (18): roadmap, product strategy, user stories, Agile, Scrum, Kanban, Jira, Confluence, Asana, OKR, KPI, sprint planning, stakeholder management, cross-functional, go-to-market, product-led growth, discovery, prioritization

Sales & marketing (27): Salesforce, HubSpot, Outreach, Marketo, Pardot, SEO, SEM, content marketing, demand generation, lead generation, pipeline, quota, cold calling, prospecting, MEDDIC, MEDDPICC, Challenger, SPIN, account management, upsell, churn, customer retention, attribution, CRM, email marketing, paid acquisition, conversion rate

Finance & legal (14): GAAP, IFRS, FP&A, financial modeling, variance analysis, month-end close, revenue recognition, audit, NetSuite, QuickBooks, SOX, budgeting, contract negotiation, due diligence

Compliance & security (12): SOC 2, HIPAA, GDPR, PCI, ISO 27001, penetration testing, threat modeling, incident response, risk assessment, IAM, encryption, vulnerability management

Ops & support (13): Zendesk, Intercom, SLA, escalation, runbook, on-call, process improvement, vendor management, capacity planning, supply chain, logistics, Six Sigma, Lean

Frequently Asked Questions

What does “25% coverage” actually mean?

Take two postings for the same job title at two different companies. Count the skills posting B names, from our published list of 174. Then count how many of those posting A also names. The median result across 6,298 such comparisons is that A covers a quarter of B’s list. Practically: a resume perfectly tailored to one posting already matches about a quarter of the next employer’s named requirements, and misses about three quarters.

Does this prove job titles are meaningless?

No, and the control is what rules that out. Postings with different job titles overlapped 11.1%, so sharing a title roughly doubles the overlap. A title carries real information — it just carries far less than most people assume when they write one resume and send it everywhere.

Is 18 companies enough to conclude anything?

It is enough for what is claimed and not enough for more. 3,910 postings is a large sample of US technology hiring, and the effect measured is large and survives a control. It says nothing reliable about nursing, teaching, construction, retail or public-sector hiring, and this page does not claim otherwise.

Why is “cross-functional” the most common requirement?

Because it is cheap to write and hard to disagree with, and because collaboration across teams is genuinely most of what these jobs involve. For a resume it is a useful signal: at 44.3%, it is likelier to appear in a given posting than any technology, and it is one of the few terms worth carrying by default — provided you can point at an actual example when asked.

Can I get the raw data?

The postings came from a public API that anyone can call, and both scripts are short and unremarkable. Write to us and we will send the collector, the analysis and the list of posting IDs so you can reproduce or contradict the figures.