The Pay Figure Decoder: What Each Salary Number Proves

You are staring at a salary number - a posted range, a site's "average," a percentile band - and you need to know what it actually proves before you build your expectations on it. This guide is a source-neutral lookup for job seekers and anyone reading the market: jump to the figure type in front of you, read what it reliably proves, and read the two main ways it lies. Other guides help you build a target range or weigh a metro premium. This one decodes the raw numbers you meet on the way there.

What each pay figure proves and how it distorts

Every pay figure is a different measurement of a different thing, and each one is honest about something and silent about something else. The table below is the spine of this guide. Read the row for the figure in your hand, then read the sections that follow for the detail.

The single most useful habit this table builds is refusing to compare across rows without adjustment. A base-only government percentile and a total-compensation crowdsourced median are not the same currency. Treat each number as evidence about one slice of reality, and the distortions column tells you which slice to distrust.

The posted range: a filtering device, not a target

A posted salary range proves that an employer is willing to hire somewhere inside a band, and little more. Glassdoor's analysis of 147,568 job listings, matched to employee-reported pay by the same employer, city, and title, found employer-provided ranges are accurate about 67% of the time - meaning reported salaries fell inside the posted band in roughly two-thirds of cases.

That 67% sounds reassuring until you read what happens inside the ranges that hold. Over 60% of in-range salaries land below the median of the posted range. The band is wide on purpose. Employers post a broad range to attract applicants and preserve negotiating room, so the top of the band exists mostly to recruit, not to promise.

The misses matter too. In the same data, 22% of actual salaries fell below the posted range and 11% fell above it. So a posted range is not a guarantee even at its edges. The practical read: assume you land in the lower half unless you have specific leverage, and treat the ceiling as marketing.

The crowdsourced median: fresh, unverified, self-selected

A crowdsourced median proves rough central tendency from a large, recent sample, and its great virtue is freshness. Its great flaw is that nobody checked the entries. Self-reported data is plagued by self-selection bias, because people are more likely to report their salaries when they feel underpaid or massively overpaid, which inflates compensation expectations at both tails.

One vendor cites research indicating self-reported salary data diverges from verified compensation by 18 to 22% on average, with higher-level positions showing larger gaps because their pay structures are more complex. Treat that specific figure with care - the underlying study is unnamed, so it is vendor-asserted, not peer-reviewed. The directional mechanism is far better established than the exact percentage.

Two failure patterns hide inside a fresh-looking crowdsourced number. First, "updated today" can mean one new entry on top of years of stale ones. Second, tech total-compensation figures routinely miscount equity, with four-year grants entered as annual amounts, inflating the headline. Some platforms counter this: one tech-focused source advertises 245,000-plus data points and validates entries against offer letters and pay statements, which is a materially stronger position than an open, unverified box.

The survey mean and median: verified but stale

An employer-verified survey figure proves a defensible market rate collected from real payroll, and it is the number compensation teams cite when they must justify a band. The named traditional providers job seekers will encounter behind these figures are Mercer, Radford, and Willis Towers Watson. The cost of that rigor is time.

Companies self-report survey data once a year, it gets collated over several months, and by the time you receive it, it reflects market rates from at least six months ago. Vendors selling continuous data push a harder number: they argue the figures you apply today reflect market conditions from 12 to 18 months ago. The honest band is 6 to 18 months, and no neutral authority pins the exact figure, so this is a spot where sources disagree by design - the ones claiming the largest lag are usually selling the alternative.

Within a survey you will also meet both a mean and a median, and they encode opposite lies in skewed pay. Income is a classic right-skewed distribution, so the mean sits above the typical worker. A five-salary example ending in a single $1,000,000 outlier produces a mean of $238,600 that nobody earns. Use the median when you want the typical wage. Keep the mean only when you are estimating a total - to price a whole team, the mean times headcount is the right tool and the median underestimates the cost.

The government percentile: rigorous, pooled, base-only

A government percentile from the Occupational Employment and Wage Statistics program proves a suppression-tested population estimate for a specific occupation, area, and industry. The median, or 50th percentile, is the wage at which half of workers earn below and half above; OEWS also publishes the 10th, 25th, 75th, and 90th percentiles, and a mean computed by dividing summed wages by total employment.

Two properties of OEWS trip people up. First, it pools multiple survey panels over three years - OEWS is a semiannual survey of roughly 200,000 establishments, and estimates combine six panels - so it is stable and rigorous but not current. Second, the wage definition is straight-time gross pay: it includes base, commissions, and production bonuses, but excludes overtime and non-production bonuses. Comparing an OEWS figure against a crowdsourced tech total-compensation number that includes equity is comparing two different quantities.

Who produces these numbers, and why the pool is thin

The benchmarks you read are shaped by a small, concentrated profession, and that concentration amplifies single-source bias. In Refolk's index of professional profiles, 2,950 people hold the current title Compensation Analyst in the United States, against just 39 in the United Kingdom - roughly a 76-to-1 gap.

That thinness matters when you read a UK or small-market benchmark: a handful of practitioners are cutting the numbers, so a single house methodology can dominate. The seniority split is also worth knowing when you decide whose figure to trust. Refolk's index shows the analyst and manager bands are close in size, so plenty of the published work comes from senior owners rather than junior data-pullers.

These counts are indicative, not a census - the title lists are not exhaustive, so read them as the shape of the pool, not its exact size. If you want to pressure-test a benchmark, the fastest path is talking to someone who has built one. Refolk writes your resume from your own history and tailors it to each posting, and I can also find the people who own these methods.

How to vet a pay figure before you anchor

This is the procedure a careful analyst runs before trusting any number. It takes about an hour end to end for a single figure, and most of the value is in steps four and seven, where segmentation and triangulation live.

How this goes wrong: the failure modes

Most bad anchoring traces to one of eight predictable errors. Each has a false positive - the plausible-looking wrong conclusion - and a check that catches it.

• Anchoring on the range midpoint. The false positive: "the range is $120-160k, so I'll get about $140k." The check: over 60% of in-range pay lands below the median, so assume the lower half unless you have leverage.
• Reading a mean as typical. The false positive: quoting a $238,600 mean when the median worker earns far less. The check: if mean and median are far apart, use the median for a typical wage.
• Matching on title only. The false positive: a defensible-looking number that is simply the wrong cut, like the Senior Software Engineer role that produced $145,000 and $175,000, a roughly 21% gap. The check: re-pull constrained by level, industry, size, and location.
• Trusting a fresh-looking crowdsourced number. The false positive: an inflated tech total-comp figure where four-year equity grants were entered as annual. The check: filter by date and read aggregates, not the top outlier.
• Ignoring suppression in government data. The false positive: assuming an occupation "doesn't exist" locally. The check: confirm whether the cell was suppressed by the three-firm rule versus genuinely unpopulated.
• Treating survey data as current. The false positive: pricing a new AI-engineering role off a survey collected two years earlier. The check: find the collection date and add the 6-to-18-month lag before trusting it.
• Confusing base with total comp. The false positive: comparing a base-only OEWS figure against a total-compensation crowdsourced figure that includes equity. The check: confirm each number's inclusions before comparing.
• Using median to estimate a budget total. The false positive: median times headcount, which underestimates real cost in a skewed pool. The check: use the mean for totals.

A copy-paste vetting note and a final checklist

When you record a figure in your search tracker, capture the same four fields every time so future-you can trust it. Paste the template below into your notes for each number you plan to anchor on.

FIGURE: $158,000 (median)
SOURCE TYPE: crowdsourced platform, offer-letter verified
STATISTIC: median (not mean)
AS-OF DATE: submissions filtered to last 12 months
SEGMENTATION: L5 / SaaS / 500-1000 employees / Austin metro
SAMPLE / SUPPRESSION: n = 240 filtered entries; base+equity included
CROSS-CHECK: OEWS P50 base = $131k (base only); survey mean = $149k (collected ~9 mo ago)
MY READ: anchor mid-$150s for total comp; base-only sources understate because they exclude equity

Before you let a number set your expectations, run it against this list.

Keeping this current

The figures decay in predictable ways, so re-check them on a schedule rather than a date. Survey benchmarks age 6 to 18 months by design, so anything you pulled more than two quarters ago deserves a fresh cut before an offer conversation. Posted-range behavior is stable enough to treat the 67%-accurate, 60%-below-median findings as a working rule, but re-pull the specific range for any live posting, because the band is set per requisition. Government percentiles update on the OEWS release cycle and pool three years, so treat movement between releases as slow signal, not noise. When a market moves fast, trust the freshest verified source you can find and downweight anything you cannot date.