The Entry-Level Index: method

Fixed 2026-10-03, before any capture had run on the full board registry. Changes are listed at the bottom with dates and reasons.

Question

How much prior experience do US business-side job postings ask for, and in particular postings whose title presents the job as early-career?

Data

The job engine (~/claude/job-scrape) sweeps about 4,000 employer job boards every morning. Since 2026-10-03 its research capture records every posting it reads that passes the business-role title screen, with the experience gate's verdict on the full description, the sentence that set the floor, the posted pay range, the function (from the title) and the metro (from the location).

Population

every posting is read in full. Other vendors are fetched only in part, which would bias the sample, so they are left out.

• Boards on Greenhouse, Lever and Ashby only. Their listings include the full description, so

US-remote screen.

• Located in the US: the location names a US place, or the posting passed the engine's

titles), and not the "Technical & engineering" function bucket.

• Business-side titles: the engine's title screen (no senior, engineering, sales-quota or intern

rejects outright are excluded, as are postings the gate judged only by pay (a high range with no experience signal), because no experience floor was computed for them.

• Postings whose description could not be read (under 150 characters) or whose title the gate

Unit of analysis

A distinct role: one company and one title. Large employers post the same title many times (one civil engineering firm listed more than 30 "Project Manager" requisitions in a single sweep); counting postings would let a handful of employers set the result. A role takes the reading of its most recently seen posting. An employer-weighted version (every company counts once) is reported as a robustness check.

Measures

Each role's experience floor comes from the gate's reading of the full description: "no experience needed" (says so, or a range from 0), "not stated", or a number of years (the gate takes the low end of a range and ignores years that are not about the candidate).

Early-career title. The title contains one of: junior (or jr), entry level, associate, new grad, new graduate, recent grad, graduate, early career, early talent, trainee, apprentice, rotational, campus, university, "analyst program", "development program", "summer analyst", "class of 20xx"; or a level I or 1 at the end of the title or before a comma, dash or bracket ("Analyst I", "Accountant 1 - Payroll"). "Data entry" does not count.

Change before any outcome was examined: the first plan used only the strictest entry-level words (junior, entry level, new grad, trainee, apprentice, rotational). A capture test on 200 live boards on 2026-10-03 found those titles on about 0.3% of postings, too few for a weekly figure, so the headline population was widened to early-career titles as defined above. Only title counts were looked at when making this change, never experience floors.

Headline

The share of early-career-titled roles whose full description asks for 2 or more years of experience, with a 95% interval from a bootstrap that resamples companies (roles from one company move together).

Reported beside it: the share asking for 3 or more years (FirstRung's cutoff), and the share that explicitly says no experience is needed.

Chapters

1. How floors are distributed, early-career titles against all business roles.
2. The headline by function, metro and company size (open postings on the company's board: under 25, 25 to 99, 100 to 499, 500 or more).
3. Pay: the median posted midpoint by experience floor, and an ordinary least squares regression of the log midpoint on the floor, with function and metro fixed effects and heteroskedasticity-robust standard errors. Posted ranges are what employers advertise, not what anyone is paid, so this is descriptive, not causal.
4. Pay transparency: the share of roles with a posted range, and the range width (top divided by bottom) by metro.
5. The weekly series: the headline for roles first seen in each week.
6. Twelve examples of early-career titles that ask for 5 or more years, quoting the sentence, drawn with a fixed seed.
7. How accurate the gate is, from the rules-vs-LLM benchmark (~/claude/rules-vs-llm).

Publishing

analyze.py reads the capture database and writes site/data/index.json, summary.json and the downloadable roles.csv.gz (one row per distinct role: company, title, function, metro, floor, pay, first and last seen, a link to the posting). Version 1 follows 7 daily sweeps; after that the site is rebuilt weekly. Every number on the site is read from index.json.

Changes after this point

2026-10-20, before edition 1 was published. The analysis had already been run once on the full capture (18 sweep days, 2026-10-03 to 2026-10-20) when these changes were made, so they are not blind. Each is reported beside the fixed version.

1. Quoted examples (chapter 6). Reading the first draw showed that the title rule also matches words that name the work, a rank or the employer rather than the job's level: "University Recruiter" (hires at universities), "Leadership Development Program Manager" (runs the program), "Associate Partner" (a senior rank), "… at University of Louisville" (the employer). Examples are now drawn, with the same seed, only from titles that still count as early-career once those phrases are removed: a university or campus recruiter; a development program manager, director, lead, specialist or coordinator; an associate partner, principal, director, vice president or general counsel; "university of" and "college of" (NOT_LEVEL in analyze.py).
2. Sensitivity check. The headline is also computed without those titles. It moves from 31.8% to 31.3% (13 of 1,092 roles removed). The headline keeps the fixed definition.
3. Exploratory: which word. The headline split by the word that made the title early-career, taken in this order: entry level, new grad or early career; junior; level I or 1; associate; university or campus. Groups under 30 roles are not shown. This was not in the plan and is labelled exploratory wherever it appears.
4. Weekly series (chapter 5). Shown with a company-clustered interval for weeks with at least 30 roles. The first week holds every role that was already open when capture began, so it is not comparable with the weeks after it.
5. Gate accuracy (chapter 7) waits for the rules-vs-LLM benchmark, which needs hand labels; edition 1 says so instead of citing a figure.
6. Publishing. Edition 1 follows 18 daily sweeps rather than 7. The weekly rebuild is not scheduled yet; until it is, each edition is published by hand and dated.

This page is METHOD.md from the study's repository, rendered as written. Back to the findings.