How we score, in full.

Each match gets six sub-scores between 0 and 1. Their weighted sum is multiplied by the interview-likelihood multiplier (held between 0.3 and 1.5), and the result times 100 is the score on the card.

Base = 0.25 O + 0.15 H + 0.20 C + 0.15 P + 0.10 G + 0.15 F

A job sits in Realistic now when its score is at least 60 (you can lower the line), freshness is at least 0.5, pay fit is at least 0.6 when the posting states a range, no requirement rules you out, and the posting does not require a degree you have not listed unless it accepts equivalent experience. Everything else is under Show more with the first reason it fell below the line, and you can apply anyway.

When a degree requirement is the reason, the card quotes the posting's own sentence and this finding: employers that filter on degree requirements screen out qualified people; 88% say their systems do (Harvard Business School and Accenture, Hidden Workers: Untapped Talent, 2021).

Signals that predict whether an application gets a reply. Each one multiplies the base score within its range; the product of all of them is held between 0.3 and 1.5. The Why tab shows each signal that applied to a job with its evidence.

A posting can stay up after the job is filled. We score how likely that is from facts we can show with a date and a source, and we never use any word but these labels: Likely stale, No longer listed, Pipeline posting, Low response history.

At a score of 70 or more a posting is marked Likely stale and left out of automatic applications unless you include it; from 40 to 69 it needs one tap from you first; under 40 it is treated as live. You can include any posting.

Whatever the score, these facts set a posting aside:

• The employer's system no longer lists it on two checks in a row.
• Its application deadline has passed.
• The posting says there is no vacancy.
• It is seen only on an aggregator while the employer's own board is live.

Before every application we check the employer's own system once more, and a job that is gone is skipped, never sent. Layoff notices count only through public WARN filings.

Next to each job, a range for the chance of a first reply. It is an annotation: it never changes which jobs you see or their order. It starts from the reply rate of people who applied through us to similar jobs, shrunk toward the occupation's rate with weight n / (n + 50), and before we have that, from the ranges published research reports:

If you choose to tell us about yourself on your account page, a second range appears: “with what you told us”. Only factors you declared apply. Each one multiplies the chance by its published ratio, weighted by how well the evidence fits; the combined multiplier is held between 0.5 and 1.5, the chance between 1% and 60%, and the range is the central 80% interval, with room for how much employers differ (0.15, or 0.25 when an employer has no reply history with us). We never print a single number.

The words follow how far the multiplier is from 1: under 10% reads as about the same; 10% to 30% reads as somewhat lower or higher; over 30% reads as noticeably lower or higher.

Audit studies measure the first contact with fictitious applicants, not hires. Employers differ a lot: in the largest US study some companies showed no gap at all (Kline and others, 2022).

The ranking, the cut-off, the learner, the applications we send and anything an employer sees take no protected trait. Nothing is inferred from your name, photo, graduation dates or work history. What you choose to tell us changes only the chance range you see:

These answers change only the chance-of-a-reply estimate you see. They never change which jobs we show you or their order and never go on an application. We use them only to compute your estimate and, in aggregate with others' and never per person, to publish our annual fairness audit. If you choose to answer an employer's voluntary self-identification survey, that is a separate choice you make on your answers page, and we show you the employer's wording first. Remove these any time.

Answers you give to an employer's own voluntary self-identification survey are a separate choice, kept apart, and sent only to that survey, in the employer's wording, when you have chosen an answer for that question.

What the research found about race in hiring

Race, ethnicity and national origin are never a factor in any estimate, and we do not ask for them. The research is shown as context, because it is real and you should be able to read it:

• Bertrand and others, 2004: Are Emily and Greg More Employable Than Lakisha and Jamal? A Field Experiment on Labor Market Discrimination. White-sounding names 9.65% callbacks vs African-American-sounding 6.45% (3.2 pp, 50%).
• Quillian and others, 2017: Meta-analysis of field experiments shows no change in racial discrimination in hiring over time. Since 1989 whites received 36% more callbacks than African Americans and 24% more than Latinos; no change over 25 years for African Americans.
• Kline and others, 2022: Systemic Discrimination Among Large U.S. Employers. Distinctively Black names received 2.1 pp fewer contacts on average, concentrated in a minority of firms; the average gender gap is near zero with firms differing in both directions (between-firm SD 2.7 pp).
• Oreopoulos, 2011: Why Do Skilled Immigrants Struggle in the Labor Market? A Field Experiment with Thirteen Thousand Resumes. English-sounding names about 16% callbacks vs about 11% for Chinese, Indian or Pakistani names with identical Canadian credentials.
• Agan and others, 2018: Ban the Box, Criminal Records, and Racial Discrimination: A Field Experiment. Employers that asked about records called no-record applicants 63% more often; after ban-the-box removed the question, the white-Black callback gap at those employers grew from 7% to 43%.

The gap is concentrated in a minority of firms, and some show none.

Rules we chose, not thresholds a statistics agency publishes. Each is labeled where it appears.

• Status is our rule: Hiring above +3% over a year, Cooling below -3%, Mixed between, from each sector's postings index against its value a year earlier.
• The Realistic now line starts at 60.
• Likely stale starts at a staleness score of 70, and one tap is asked from 40.
• One application per employer every 30 days, and never two roles at the same employer on the same day.
• No account sends more than 50 applications in a day.

A fairness check runs on every change to the code: it ranks a test profile with declared traits and voluntary survey answers present and absent, and fails if any job moves or changes tier. It also reads the code itself and fails if any module other than the chance estimate and the survey-answer filler can reach those two kinds of answers. Once a year the same check runs over real recommendations in aggregate, never per person, and the result is published on this page.

• Logging into or driving your LinkedIn, Indeed, Workday, iCIMS, Taleo, Harri or Snagajob accounts.
• Storing any platform password or session cookie.
• Creating or pooling accounts.
• Solving or bypassing CAPTCHAs.
• Spoofing fingerprints or user agents.
• Answering voluntary self-identification questions without an explicit saved choice of yours.
• Background-check, credit or drug-screen consents.
• Non-compete, arbitration and policy acknowledgments.
• E-signatures and attestations.
• Assessments, one-way videos, phone screens and live interview assistance.
• Identity verification.
• Third parties' personal data.
• Application fees.
• Cold email at scale to recruiters or hiring managers.
• Outreach sent by the product. Notes are drafted; you send them.
• Any invented experience, education, title, date, metric or skill.

Every study the Why tab, the chance range and the tips cite. Most were read through their published summaries; the last column says which were checked against the full paper. Where the numbers come from