The Pay Figure Admission Standard, Six Criteria Before a Number Counts

Every pay number that reaches you during a search is a claim about the market, not the market itself. That covers a salary site estimate, a posted range, a visa filing, a friend's offer, and a chatbot answer. This standard is for job seekers deciding whether one incoming figure is solid enough to move their target or their counter. It gives six pass/fail criteria, an aging rule for stale numbers, and a log row that lets two people grade the same figure and reach the same verdict.

What makes a pay figure admissible?

A pay figure is admissible when it passes all six criteria: provenance you can trace, a named pay component, a level match, a geography match, a known reference date aged to today, and a recorded sample depth. If a figure fails any one of them, it stays in your log as context but does not move your target or your counter.

"Admissible" here means allowed to change a decision. Building a target range from public data is a separate job. This standard is the gate each number passes before it enters that range. It is also the check you repeat every time a recruiter, a forum thread, or an AI assistant quotes you a figure.

Every figure gets exactly one provenance class:

• Government survey. BLS OEWS percentiles. OEWS (Occupational Employment and Wage Statistics) is the federal survey of employer-reported wages by occupation and area.
• Visa filing. An LCA record. An LCA (Labor Condition Application) is the wage attestation an employer files for a visa worker. These records are published by OFLC, the Labor Department's Office of Foreign Labor Certification.
• Posted ask. A range on a job listing.
• Self-report aggregate. A salary site figure built from user submissions.
• Personal offer. An offer letter, yours or someone else's.
• AI. A chatbot answer.

Some questions have no published threshold. Where that happens, I set one so that grading can be repeated:

• Single observation. One offer, one filing, or one posting has a sample depth of 1. It passes depth only when you use it about that same employer. Used as a market figure, a count of one fails.
• Unknown reference date. If you cannot find the date a figure describes, it fails on age. You cannot age a number without a start date.
• Level comes from the source, not from you. You may not assign a level to a figure that arrived without one. The source must state the level, or the criterion fails.

What does each pay source actually measure?

Each source counts a different slice of pay, has a different level field, and arrives with a different lag. That means the source's own definition settles most component and level verdicts before you even look at the number. Use the table below as the reference. The notes under it cover what each source proves and how it misleads.

Government survey percentiles (OEWS)

The BLS technical notes define exactly what OEWS counts. Counted:

• base rate
• cost-of-living allowances
• guaranteed pay
• hazardous-duty pay
• incentive pay, including commissions and production bonuses
• tips

Not counted:

• overtime
• severance
• shift differentials
• nonproduction bonuses, which include year-end bonuses
• the employer's cost of supplementary benefits
• tuition reimbursement

The BLS concepts text does not mention equity, and secondary sources say stock is excluded. Some sites repeat the claim that OEWS excludes all bonuses. That is wrong: commissions and production bonuses are counted.

What it proves: breadth. OEWS covers about 830 occupations in about 530 areas, drawn from roughly 1.1 million establishments.

How it lies: it carries no work level. Its percentiles are also top-coded, meaning that above a ceiling the survey reports the ceiling rather than the real value. One secondary source puts the ceiling at $239,200 a year, or $115 an hour, for one release. I could not confirm that from the BLS page, and the ceiling can change between releases, so read the current technical notes. A percentile sitting at the top code is a floor, not a value.

If you want to use BLS wage data in a salary negotiation, this is the practical consequence. OEWS passes component, geography, and depth easily, but on its own it always fails level. It can bound your number. It cannot set your counter.

Visa filings (LCA)

Each LCA record lists three things: the offered wage, the prevailing wage, and a PW_WAGE_LEVEL field set to I, II, III, IV, or N/A. The records reflect certified job offers, not actual hires. The prevailing wage is the minimum wage the employer must attest it will pay.

What it proves: an explicit level. Among the common public sources, it is the only one with a level field.

How it lies: it is an offer and a legal floor, not accepted pay. H-1B salary data is widely described as base pay only, but I could not confirm that from a primary source. Grade the component as "base, unconfirmed" and compare the figure only against base pay. Rates of miscoded records are not publicly established.

Posted ranges

Glassdoor's economic research team matched job listings to user-reported salaries for the same job, employer, and city. The results:

• 67% of ranges had reported salaries falling inside them.
• In 22% of listings, actual pay sat below the posted range.
• In 11%, actual pay sat above it.
• More than 60% of the in-range salaries sat below the midpoint.
• Ranges in states with pay transparency laws were no more accurate.

Indeed Hiring Lab put the typical spread of a posted range at about 17%.

What it proves: what one named employer is asking, right now.

How it lies: people read the top of the range as the market. Is a posted salary range accurate? As an ask, yes. As a record of what people are actually paid, it leans high.

Self-reported aggregates

Glassdoor's estimates count base pay only. In its own words they "do not include bonuses, commissions, tips, stock, benefits, or years of experience". Its "Low" badge means a small number of reports, data that is several years old, or both. No numeric confidence threshold is published. One critic says the site averages every salary ever submitted for a role. A Glassdoor salary page also mixes three different inputs: model estimates, ranges supplied by employers, and user reports.

So how accurate are salary estimate sites? You cannot tell from the badge alone, which is exactly why this standard asks for a count.

Personal offers

An offer letter has the strongest provenance there is for that employer, and a depth of one. Before you log a friend's offer, ask them three things: what it includes, what the title and level were, and when they received it.

Chatbot answers

Geiger et al. sent 98,800 prompts to each of four ChatGPT versions. The largest gaps came between model versions, and between prompts voiced as the employee and prompts voiced as the employer. Sorokovikova et al., in "Surface Fairness, Deep Bias", tested five models. When they compared the most advantaged and least advantaged personas head to head, the advantaged persona got the higher figure in 35 of 40 experiments (87.5%).

So can you trust a ChatGPT salary estimate? Under this standard, an AI figure fails provenance unless it cites a source you can open. If it does, grade that source instead.

How do you match level and geography when job titles differ by market?

Match on the title the local market uses, not the one you would use at home, and only accept a level the source states outright. If you look up a figure under a title borrowed from another market, it draws on a thin slice of the local pool and can fail both criteria at once.

Refolk's index of professional profiles shows how far apart two markets can sit on what is nominally the same job:

Three ratios fall out of those counts. The US has about 70 times as many Compensation Analysts as the UK. Within the UK, Reward Analysts outnumber Compensation Analysts about 4.2 to 1. And the US title pool is about 13.5 times the combined UK pool. The two UK titles also attract different employers. In a 25-profile sample from Refolk's index, UK Compensation Analyst employers included Google, Disney, and Network Rail, while UK Reward Analyst employers included HSBC, Railpen, and Ageas UK. A sample of 25 is a pointer, not a measurement. It is enough to show that the two titles are not interchangeable, but not enough to say how far their pay differs.

To grade geography:

1. Confirm that the metro or worksite matches yours.
2. Search local listings for the role and note which title dominates.
3. Fail the figure if its title is not the dominant local title, even if the metro matches.

To grade level, know which levels a source can give you. In LCA filings, the level field splits roughly as follows (secondary source, one fiscal year of filings):

• Level I: 18.8%
• Level II: 38.7%
• Level III: 19.3%
• Level IV: 15.3%
• No level: 7.9%

The 7.9% with no level always fail. A record at the right level still needs one more check: whether you are that level. The posting's band is written for a specific seniority, and one quick read is how well your history fits that posting's requirements. Refolk scores your fit against a specific posting from your own history. That tells you whether the level-stated band you are grading describes you, or the person one level above you.

How do you age a stale pay figure?

Find the date the figure describes (its reference date), count the months from then to today, and multiply by (1 + r) raised to the power months/12. Here r is the latest 12-month change in the Employment Cost Index (ECI) wages and salaries series, for the series closest to your employer and metro. The ECI is the BLS measure of change in what employers pay, and BLS itself uses it to age older OEWS survey panels.

The publication date is the trap. Survey figures pass through several stages before they reach you, and each stage adds months.

The figures in the table below come from one published 12-month ECI reading. The multipliers age a May-reference figure by 17 months. Pull the current reading from the BLS ECI news release before you use any of them, because the index updates each quarter.

Each multiplier is (1 + r) raised to the power 17/12. Take a May-reference OEWS figure read 17 months later and aged with the civilian series: every $1,000 becomes $1,046. Aged with the Boston series, it becomes $1,033 instead. That gap is why you should use a metro series when your metro has one, and the private series when your employer is private.

Reference dates by source:

• OEWS. The May named in the release. BLS has already aged the older panels to that month, so age forward from May. Do not age the panels a second time.
• LCA offered wage. The filing period. Files are published quarterly, about 6.5 weeks after the quarter closes.
• LCA prevailing wage. The reference month of the OEWS data it was built on. Each prevailing-wage year runs July to June and uses OEWS from more than a year before it starts, so by June the underlying survey is about two years old. OFLC does not age it.
• Posted range. The posting date. A live posting has zero months.
• Personal offer. The date the offer was made.
• Self-report aggregate. Usually unknowable. If the site averages across all time, the figure fails on age.

The release lag itself also moves. In one year a federal government shutdown pushed the OEWS release about six weeks later than the year before, so a figure's age at release varies from year to year. Always count from the reference month, never from an assumed lag.

The grading procedure, step by step

The procedure takes about 25 minutes per figure. It runs from capture to a logged verdict that a second grader can reproduce. No published source sets a seeker-side order for these checks. The sequence below is assembled from the agencies' own methods.

Write every grade into the same row format. Then a second person, or you a week later, can check the verdict without re-reading the source.

Figure as given:
Source and link:
Date received:
Period stated (hourly / annual):
Scope stated:
1 Provenance class (survey / visa / posted / self-report / offer / AI): PASS / FAIL - reason
2 Component (base / cash total / total comp): PASS / FAIL - reason
3 Level (as stated by source): PASS / FAIL - reason
4 Geography (metro + dominant local title): PASS / FAIL - reason
5 Age (reference date / ECI series / months / multiplier / aged figure): PASS / FAIL - reason
6 Sample depth (count or badge): PASS / FAIL - reason
Verdict: ADMIT / CONTEXT ONLY
Failed on:
Second grader and their verdict:

How the standard goes wrong

Most grading errors come from the same eight patterns. In each one, a figure looks as if it passed a criterion when it did not. Each pattern below names what it looks like and the check that catches it.

1. Blended provenance. A Glassdoor salary can come from model estimates, ranges supplied by employers, or user reports. If you cannot tell which of the three produced the number, provenance fails. Check: find the input type on the page itself, not in your memory of the site.
2. Component mismatch. An OEWS median looks like a full offer, but it leaves out year-end bonuses and, per secondary sources, equity. Compare it with a total-compensation offer and you will understate the market. Check: read the BLS concepts text and compare like with like.
3. Floor read as a level ask. An LCA wage is a legal minimum attested for one level. A Level I figure used as a senior counter fails, and so does any filing with no level. Check: read the PW_WAGE_LEVEL field before reading the dollar figure.
4. Title translated across markets. In Refolk's index, UK Reward Analysts outnumber Compensation Analysts about 4.2 to 1. Search the US title in the UK and you miss most of the market. Check: count which title dominates local listings before you grade geography.
5. Publication date taken as reference date. A government wage figure published in one year usually describes a May about 17 months before you read it. Check: write the reference month in the log before calculating anything.
6. Top of the posted range taken as the market. Twice as many listings had actual pay below the range (22%) as above it (11%). Employers post the band they could stretch to, not what they typically pay. Check: log the range as an ask and grade the midpoint, not the top.
7. Badge without a count. A "High confidence" badge can sit on top of old reports that have been adjusted for inflation. Check: if no count is shown, depth fails, whatever the badge says.
8. An unstable chatbot answer. The same question returns different numbers depending on persona, prompt voice, and model version. Check: rerun it neutrally twice in fresh sessions. If the answers diverge, the figure fails.

For failure mode 8, use a prompt that strips out persona and voice. Geiger et al. found the largest gaps between model versions and between employee-voiced and employer-voiced prompts, so the neutral version names neither side.

What is the typical annual base salary, excluding bonuses, commissions, and equity, for a [job title] at [level or years in role] in [metro area]?
Name the specific data source behind the figure and the date the data describes.
If you cannot name a source, say so instead of estimating.

A pay figure is evidence only once you can say what it counts, for whom, where, and as of when.

Where is the evidence behind this standard thin?

The criteria rest on agency definitions and published studies, but several inputs are secondary or unpublished. Where that is the case, the standard defaults to fail rather than guess.

• The OEWS top code. The $239,200 a year / $115 an hour ceiling comes from a secondary source, not the BLS page. Read the current technical notes before treating any high percentile as real.
• LCA component. The claim that it is base pay only is not confirmed from a primary source. Miscoding rates are not publicly established. Grade the component as "base, unconfirmed".
• Equity in OEWS. The BLS concepts text does not mention it. Only secondary sources say stock is excluded.
• Self-report thresholds. No numeric confidence threshold is published for salary site badges. A minimum report count does circulate, but it is unverified, so this standard ignores it.
• The order of steps. No published source documents a seeker-side order. Aging is the disputed step: BLS ages to a reference date before publishing, while OFLC reuses OEWS a year later without aging.
• Title counts. The Refolk index counts are counts, but the employer split comes from a 25-profile sample. Check local title usage yourself for any role outside this example.

How do you keep the log current and use a figure that passed?

Re-grade the age criterion whenever a new release lands, and re-grade the whole row whenever the role, level, or metro you are targeting changes. A figure that passed last quarter can fail today on age alone.

Re-check triggers:

• ECI: quarterly. Update the multiplier on every admitted row.
• OEWS: once a year, with a release date that varies. When a new May reference arrives, replace the old figure rather than aging it further.
• LCA disclosure files: quarterly, about 6.5 weeks after quarter close.
• Posted ranges: when a posting closes or is re-posted, record the new date.

Once a figure is admitted, it goes into your target range or your counter, and the counter goes back to the specific posting it came from. Applying under that counter is less work when the application already matches the band. Refolk tailors your resume to the posting and drafts the cover letter, so the level the figure assumed is the level the employer reads in your application.