# The Salary Figure Standard, Seven Checks to Cite, Adjust, or Discard

*Run any salary figure through seven pass or fail checks and reach the same Cite, Adjust, or Discard verdict a second grader would, with the adjustment written down.*

- Canonical URL: https://www.refolk.ai/candidates/guides/salary-figure-cite-adjust-discard
- Pillar: Reading the market
- Format: Standard
- Published: 2026-11-20
- Last reviewed: 2026-11-20
- Reading time: 21 min
- Keywords: which salary website is accurate, how to check if salary data is reliable, glassdoor salary vs actual offer, using BLS wage data for salary negotiation, H-1B salary data to benchmark offer, conflicting salary data which to trust

## Key takeaways

- A visa wage filing (LCA) only has to clear a prevailing-wage floor set at the 17th, 34th, 50th or 67th percentile, so a median taken without splitting by wage level measures floors, not the market.
- The federal occupational wage survey (OEWS) is published about 10 to 11 months after its May reference date, and each estimate blends six panels collected over three years, so its real age is older than its publication date.
- In Glassdoor's matched study, 67% of posted ranges contained employee-reported pay, but more than 60% of in-range salaries sat below the range midpoint, so a posted midpoint is not a target.
- Self-reported salaries ran about 7.3% below administrative payroll records in a linked German study and cluster toward the middle, so crowd medians can hold up while crowd 10th and 90th percentiles are the weakest numbers you have.
- In Refolk's index, 530,715 US profiles carry the title Software Engineer or Senior Software Engineer, and in a 25-person sample only 20% said Senior, so a job title alone does not tell you the level.
- A figure gets Cite as is only when all seven criteria pass, Cite after adjustment when every failure can be fixed with a written sentence another person could repeat, and Discard when any failure cannot be fixed.

This standard is for a job seeker with a pile of salary figures that disagree by tens of thousands, who needs to decide which ones can support the number they're about to name. You grade one figure at a time against seven pass or fail criteria, and each figure gets one of three verdicts: Cite as is, Cite after a stated adjustment, or Discard. It comes before the Market Pay Band Run and the Posted-Pay Conversion Standard. Those two build a band and convert it, and this one decides which numbers are allowed into the band in the first place.

## What does this standard grade, and what are the three verdicts?

It grades a single salary figure, meaning one number from one source in one unit. The result is a verdict that a second person grading the same figure would match. You don't grade sources in general, and you don't average them. The usual advice to "cross-reference multiple sources" breaks down because the sources measure different things. A survey of straight-time pay, a legal wage floor, and a crowd-reported total cash figure will disagree even when all three are correct.

The three verdicts:

- **Cite as is.** All seven criteria pass. You can put the number in front of an employer exactly as it appears.
- **Cite after a stated adjustment.** One or more criteria fail, but each failure can be fixed. You write a sentence next to the figure saying what you changed and how, specific enough that someone else could repeat it and get the same result.
- **Discard.** At least one criterion fails in a way you can't fix. The figure can still help you decide what to ask about, but it never appears in your band or in a negotiation.

Terms used throughout:

- **Pay component**: which part of pay a figure covers. That can be straight-time gross pay, base salary, total cash (base plus bonus and other cash), or total compensation.
- **Reference date**: the date the pay was actually measured. This is often not the date the figure was published.
- **Relative standard error (RSE)**: a survey's estimate of its own sampling error, expressed as a percentage of the estimate.

> **Rule:** Unwritten adjustments do not exist
>
> If you adjusted a figure, the adjustment sentence sits next to it in your notes. A figure with an adjustment that isn't written down is graded as if nothing was adjusted.

## Which salary source is accurate for what?

None of the public sources is accurate in general. Each one measures a specific pay component, at a specific age, with a specific way of showing (or hiding) its sample. So the honest answer to "which salary website is accurate" is a description of what each source can and can't tell you.

| Source | What it measures | Typical age at use | Sample disclosure |
|---|---|---|---|
| OEWS (federal survey) | Straight-time gross pay, 10th to 90th percentiles | About 10 months at release; panels up to 3 yrs older | RSE shown; suppression reason withheld |
| LCA (visa filing) | Offered wage plus a floor at one of 4 levels | Quarterly files | None; about 1.62 positions per LCA |
| Glassdoor (crowd) | Self-reported total pay, split into base and additional | Dated submissions | n shown (10 to 9,808 seen) |
| Posted range | Employer's good-faith estimate | Posting date | Not applicable |

### The federal occupational wage survey (OEWS)

OEWS is the Bureau of Labor Statistics survey of wages by occupation and area. It is the largest source here: about 1.1 million sampled establishments, with the May 2024 sample covering about 55% of national employment and a 65.7% establishment response rate. It publishes the 10th, 25th, 75th and 90th percentile wages and the median. Annual figures assume 2,080 hours of work a year.

Using BLS wage data in a salary negotiation runs into two structural limits:

- **It excludes overtime and bonuses.** It measures straight-time gross pay.
- **It is older than its label.** Releases come once a year, usually late March or early April, and report May of the prior year. Each estimate combines six semiannual panels collected over three years, and the earlier panels are aged forward using Employment Cost Index factors. The May 2024 estimates include panels going back to November 2021.

Some cells are not published. BLS doesn't publish the exact cutoff for suppression and says it cannot give the specific reason any one estimate was withheld.

### Visa wage filings (H-1B Labor Condition Applications)

A Labor Condition Application (LCA) is the wage filing an employer submits before an H-1B petition. Two groups of fields matter:

- **The offered wage**: `WAGE_RATE_OF_PAY_FROM` and `WAGE_RATE_OF_PAY_TO`, read together with `WAGE_UNIT_OF_PAY`.
- **The floor**: `PREVAILING_WAGE`, read with `PW_UNIT_OF_PAY` and `PW_WAGE_LEVEL`. The Department of Labor's record layout defines the prevailing wage as the wage "for the job being requested for First Worksite Location."

The four wage levels sit at the 17th, 34th, 50th and 67th percentiles, a scale set in 2005. The Department of Labor has proposed raising them to the 34th, 52nd, 70th and 88th percentiles. I have not verified whether that proposal is final, so check the current rule before you read a level as a percentile.

### Crowd self-reports

Self-reported pay is close to right in aggregate and noisy for any one person. One study found that industry-average Glassdoor wages correlated with Census Bureau data at 0.87 (Karabarbounis and Pinto). That is a match between industries, not between individuals. In a German study that linked survey answers to social-security records, people underreported their administrative wages by about 7.3%. Their answers bunched toward the middle, with fewer very high or very low reports, and the documented reason was rounding. Nobody has published error rates for the other mistakes you'd expect, such as base mixed with total cash, part-time pay entered as full-time, take-home entered as gross, or inflated titles.

### Posted ranges

A posted range is the employer's estimate. Washington, DC, for example, requires the range the employer "in good faith" believes it will pay. One practitioner source says no US statute sets a maximum range width. Ontario caps the spread at $50,000. A New Jersey cap of 60% of the minimum is described as proposed by one source and as in force by another, so check the New Jersey Administrative Code before relying on it.

### Hearsay

Hearsay means a figure from a friend, a recruiter, or a forum post. It has no reference date, no pay component, and no sample size. It is its own source type here because by default it fails the most criteria.

## What are the seven criteria, and what does each one prove?

Each criterion asks one yes or no question about the figure. A pass proves the figure describes the same job, pay, place, and time as the number you'll compare it with. Each one below says what a pass proves and what the figure looks like when it misleads you.

### 1. Source type

- **Pass:** the figure carries exactly one of the five labels.
- **Proves:** you know which of the known weaknesses above apply to it.
- **Misleads you when:** an aggregator page mixes LCA rows, self-reports, and postings under one heading. If you can't tell which source produced the number, it fails.

### 2. Pay component

- **Pass:** you've written down the component, and it matches the component of the number you'll compare it to.
- **Proves:** you're not comparing base pay with total cash.
- **Misleads you when:** OEWS (no bonus) is compared with an offer that includes a bonus, which makes the market look lower than it is. Or a Glassdoor headline (total pay) is compared with a base-only offer, which makes your offer look low. Whether Glassdoor estimates are total or base is disputed. Its pages headline total pay with a base and additional split, while one third party says the estimates are base only. Settle it for the specific figure in front of you.

#### Which pay component a figure sits in

1. **Total compensation** - Cash plus equity and other pay lines; outside every source graded here
2. **Total cash** - Base plus bonus and other cash; Glassdoor headline total pay sits here
3. **Base salary** - LCA offered wage; Glassdoor base split
4. **Straight-time gross** - OEWS, excluding overtime and bonuses, annualized at 2,080 hours

*A figure can only be compared with another figure from the same layer, so write down the layer before comparing anything.*

### 3. Occupation code or title match

- **Pass:** for OEWS and LCA figures, the SOC code (Standard Occupational Classification, the federal list of occupation codes) covers the duties of the job you're targeting. For other sources, the title matches exactly.
- **Proves:** the figure is for the same work.
- **Misleads you when:** a broad code pools specialties that pay differently, or a crowd title such as "Analyst" covers several unrelated jobs.

### 4. Level match

- **Pass:** the figure's level is stated and matches yours. For an LCA that's `PW_WAGE_LEVEL`. For OEWS it's the specific percentile you chose. For a self-report it's years of experience.
- **Proves:** the figure describes someone at your career stage.
- **Misleads you when:** the title implies a level the person doesn't hold. Refolk's index shows how little a title says.

**530,715 - US profiles titled Software Engineer or Senior Software Engineer in Refolk's index**

In a 25-person sample of them, 76% carried the plain title and 20% carried Senior, so one title spans many levels and does not identify yours.

| Refolk index query | Total | Title split in a 25-person sample |
|---|---|---|
| Compensation titles, US | 5,454 | Manager 13, Analyst 8 |
| Compensation titles, UK | 139 | Analyst 11, Manager 9, Specialist 5 |
| SWE or Senior SWE, US | 530,715 | SWE 19 (76%), Senior 5 (20%) |

The splits come from 25-person samples and only show direction. What they do show is that title words don't line up cleanly with seniority, so match on years and duties. If you aren't sure which level an employer would place you at, [Refolk](/candidates) scores how well your history fits a specific posting. That's a quicker way to tell whether you'd be read as the level a figure describes than guessing from the title.

### 5. Geography

- **Pass:** the figure's metropolitan statistical area (MSA), county, or city is the same as the job's.
- **Proves:** local labor market pricing applies.
- **Misleads you when:** an LCA's prevailing wage is tied to the first worksite location, and that isn't where your job would be. Or a crowd figure gives a company name with no city.

### 6. Reference date

- **Pass:** you've stated the figure's age in months, and for OEWS you've also stated the panel span.
- **Proves:** you know how much the market could have moved since the measurement.
- **Misleads you when:** you date an OEWS figure by its release. The May 2024 estimate still contains the November 2021 panel, so dating by release understates its age.

### 7. Sample size or error

- **Pass:** you've recorded an RSE or a displayed count, and no footnote (top code, suppression, or a low-reporting flag) affects the percentile you're citing.
- **Proves:** you know how much weight the figure can bear.
- **Misleads you when:** a percentile built from a small count looks precise. A Glassdoor title page can show 25th to 90th percentiles from 10 salaries. LCA files show no sample information, and one filing covers about 1.62 positions on average (910,968 certified positions from 561,505 LCAs in one fiscal year).

## When is a failed criterion fixable, and when is it fatal?

A failure is fixable if you can write one sentence that someone else could follow to get the same corrected figure. It's fatal if you can't state the missing fact at all. The verdict follows mechanically: any fatal failure means Discard, any fixable failure with no fatal ones means Cite after adjustment, and seven passes mean Cite as is.

**Fixable failures, with the kind of sentence to write:**

- **Unit mismatch.** "Annualized from hourly at 2,080 hours."
- **Component mismatch with a known split.** "Base portion only cited; additional pay excluded to match a base-only offer."
- **Geography mismatch where the same source covers your area.** "Replaced with the same source's figure for the target MSA."
- **Level mismatch where the same source has your level.** "LCA rows filtered to Level III only."
- **Age.** "Figure is N months past its reference date; cited as a floor, not a target."
- **Posted range width.** "Range spread checked against Ontario's $50,000 cap as a yardstick; midpoint not used as target."

**Fatal failures:**

- The pay component can't be determined.
- The level can't be determined, as with a level-blind LCA median that you can't re-split.
- The occupation doesn't map to your duties, and no closer code or title exists in that source.
- No reference date exists.
- The figure is hearsay and at least one of the seven criteria can't be stated.
- The percentile you need is suppressed, or top-coded in a way you can't adjust (see the failure modes below).

#### Where a figure lands before you grade the details

Horizontal axis runs from Does not match your case to Matches your case. Vertical axis runs from Low disclosure to High disclosure.

| Quadrant | What it means |
| --- | --- |
| Documented, wrong case | Adjust only if the gap has a written, repeatable fix; otherwise Discard |
| Documented, your case | Cite as is, or after a minor stated fix |
| Undocumented, wrong case | Discard |
| Your case, undocumented | Hearsay sits here; Discard until all seven criteria can be stated |

*Disclosure decides whether a figure can be checked, and match decides whether it is about you; a figure needs both to be cited.*

This rule doesn't weigh one source above another. A self-report with a date, a stated component, a displayed count, and an exact title match can beat an OEWS cell for a broad code in the wrong metro area. Disclosure and match decide the verdict, not reputation.

## How to grade one salary figure, step by step

Grading one figure takes about 22 minutes the first time and much less once you've graded a few from the same source. No published source documents this exact sequence. I put it together from each source's documented field rules. Lookup tools tend to start from the employer, and visa-filing guidance tends to filter by case status and annualize first. The order below puts the cheapest checks, the ones that kill the most figures, at the front.

#### Grade one figure

1. **Tag the source type** - Label it government survey, visa filing, self-report, posted range, or hearsay. One label only.
2. **Write down the pay component** - Straight-time gross, base, total cash, or total compensation. OEWS excludes overtime and bonuses; an LCA is offered wage only; Glassdoor splits total pay into base and additional.
3. **Match the occupation** - Map your title to a SOC code for OEWS and LCA figures; check the exact title for everything else. Write the code plus a one-line match note.
4. **Match the level** - Use PW_WAGE_LEVEL for LCA rows, the specific percentile for OEWS, and years of experience for self-reports. State the level and compare it with yours.
5. **Check the geography** - Compare the MSA, county, or city with the job's location. Write a match or a named adjustment.
6. **Record the reference date** - OEWS May reference date and panel span, LCA fiscal year, or submission or posting date. State the age in months.
7. **Record sample size or error** - Note the RSE, the displayed count, and footnotes such as (5) and (8). Write a number or the word undisclosed.
8. **Issue the verdict** - Seven passes is Cite as is; fixable failures only is Cite after adjustment; any fatal failure is Discard. Write the adjustment sentence next to the figure.

Keep one record per figure. The template below is the whole record. If any line is blank when you reach the verdict, that criterion failed.

**Salary figure grading record**

```
Figure: $______ per ______ (hour / year)
1. Source type: government survey / visa filing / self-report / posted range / hearsay
2. Pay component: straight-time gross / base / total cash / total comp
   Compared against: ______ (must be the same component)
3. Occupation: SOC ______ | Title-match note: ______
4. Level: ______ (PW_WAGE_LEVEL / percentile / years) | My level: ______
5. Geography: figure ______ | job ______ | match / adjusted
6. Reference date: ______ | Panel span (OEWS): ______ | Age: ___ months
7. Sample or error: RSE ___ / n ___ / footnotes ___ / undisclosed
Fails (fixable): ______
Fails (fatal): ______
Verdict: Cite as is / Cite after adjustment / Discard
Adjustment sentence: ______
```

*One record per figure. Leave nothing blank; write "undisclosed" where the source shows nothing.*

## How does salary-figure grading go wrong?

Most bad salary arguments start with a figure that passed a quick look and should have failed one criterion. These are the documented ways that happens, each with the check that catches it.

### Reading an LCA median as a salary

A median across all wage levels, for a code dominated by Level I filings, gives you a low number that reflects the 17th percentile floor. In one analysis, three fifths of H-1B jobs in 2019 were certified at the two lowest wage levels. Before using H-1B salary data to benchmark an offer, also keep in mind:

- **Certification is not a hire.** Employers file more LCAs than they win visas. Certification happens before the lottery, and only 85,000 new cap-subject visas exist each year. The Department of Labor itself notes that the disclosure data doesn't show what the employer ultimately used the application for.
- **One filing can cover several workers.**
- **The required wage is the higher of the prevailing wage and the actual wage**, so the floor is a legal minimum, not a description of pay.

**Check:** split by `PW_WAGE_LEVEL` and confirm you're reading the `FROM` wage column.

> A visa filing records the least an employer may pay, not what anyone was paid.

### Mixing hourly LCA rows with annual ones

If you ignore `WAGE_UNIT_OF_PAY`, hourly rows look absurdly low next to annual ones and drag down any summary. **Check:** annualize every row before summarizing.

### Treating the OEWS top code as a ceiling

Through the May 2024 estimates, any percentile at or above the top bracket was footnoted "(5)" and shown as $115.00 per hour or $239,200 per year. That is a floor on the true value, not the value. The practice changed starting with the May 2025 estimates. For highly paid roles, older figures can't sit next to newer ones without a stated adjustment. **Check:** the footnote and the estimate year. If a top-coded percentile is the one you need, it fails as a fatal error.

### Using OEWS as total cash

OEWS excludes bonuses. If you compare it with a total cash offer, the market looks smaller than it is and the offer looks generous. **Check:** the component line on your record.

### Trusting thin crowd cells

Percentiles from a count of 10 look as precise as percentiles from thousands. Self-reports also bunch toward the middle because of rounding, so crowd medians are the most usable part and crowd 10th and 90th percentiles are the least. That is the core of the "Glassdoor salary vs actual offer" gap: the tails are where self-reports drift. A minimum of three reports per company and role is claimed only by a secondary source, so treat it as unverified. **Check:** the displayed count, and cite medians before tails.

### Using the posted midpoint as the target

| Posted range vs reported pay (Glassdoor study) | Share |
|---|---|
| Reported pay inside the range | 67% |
| Range above reported pay | 22% |
| Range below reported pay | 11% |
| In-range salaries below the midpoint | More than 60% |

The study matched full-time listings to employee-reported salaries by employer, city, and job title. It also found that ranges in states with transparency laws were as accurate as ranges in states without them. A range posted under a law gets no credit for that. **Check:** the range width, using Ontario's $50,000 cap or New Jersey's 60%-of-minimum rule (its adoption status is unverified) as a yardstick, and don't cite the midpoint as a target without an adjustment sentence.

### Grading the title, not the job

In Refolk's 25-person sample of US software engineers, 20% of titles carried "Senior." A title's seniority word is weak evidence of level. **Check:** years of experience and SOC duties.

### Letting hearsay in

A recruiter's figure is often the most specific number you have and the least documented one. **Check:** run it through the seven criteria. If any line stays blank, the verdict is Discard.

> **Watch out:** The precise-looking number is usually the weakest
>
> Single numbers with no unit, date, or count, like an LCA median, a crowd 90th percentile, or a recruiter's figure, invite trust because they look exact. Precision on the page is not disclosure. Grade the disclosure.

## How do you make a recruiter's or friend's number gradeable?

Ask the follow-up questions that fill the seven criteria, and grade the answers. If the person can't answer, the figure stays discarded. It isn't wasted, though: it tells you what to ask employers about directly.

**Follow-up questions for a salary figure you heard**

```
Thanks, that's really useful. Quick questions so I read it right:
1. Is that base salary, or base plus bonus?
2. What was the title, and roughly how many years in the role?
3. Which city or office was it for?
4. When was that, roughly what month and year?
5. Is that one person's pay, or a range across several people?
6. Hourly or annual?
```

*Send as written. Each answer fills one line of the grading record.*

If you'd rather ask a compensation professional than a friend, the pool depends on where you are. In Refolk's index there are 5,454 people in the US with Compensation Analyst, Manager or Specialist titles and 139 in the UK, a ratio of about 39 to 1. That's a count of people, not of salary data. It does suggest that a UK seeker will find far fewer people whose job is pricing roles, and should expect fewer second opinions of that kind.

## What to check before you name a number, and how to keep the grades current

Before a figure goes into your band or into a conversation, its record should be complete and its verdict should follow from the rule. Re-grade whenever a source publishes new data, because several of these criteria change on a schedule.

#### Before you name a number

- [ ] Every figure in the band has a completed grading record with no blank lines.
- [ ] Every Cite after adjustment figure has its adjustment sentence written next to it.
- [ ] Every figure being compared shares the same pay component as the offer or ask it is compared with.
- [ ] No LCA figure is a median across mixed wage levels.
- [ ] Every LCA row was annualized using WAGE_UNIT_OF_PAY before any summary.
- [ ] No OEWS percentile carrying the (5) top-code footnote is cited as a value.
- [ ] Every OEWS figure states its reference May and its three-year panel span.
- [ ] No posted-range midpoint is cited as a target without an adjustment sentence.
- [ ] Crowd figures cite the median before any 10th or 90th percentile, and the displayed count is recorded.
- [ ] Every hearsay figure is either fully graded or marked Discard.

**Re-check these mechanisms rather than relying on today's values:**

- **OEWS releases** come once a year, usually late March or early April. Move every OEWS record to the new reference May and re-read the footnotes, since the top-code practice changed once already.
- **LCA wage levels.** Check whether the proposal to move the levels to the 34th, 52nd, 70th and 88th percentiles has become final. If it has, re-map what each level means for filings made under the new rule.
- **Range-width rules.** Check the New Jersey Administrative Code for the status of the 60% cap before using it as a yardstick.
- **Crowd display minimums.** Check the site's own help pages for any stated minimum count rather than relying on secondary claims.

Once your figures are graded, the Market Pay Band Run turns the survivors into a floor, target, and stretch, and the Posted-Pay Conversion Standard turns a posted range into comparable terms. For visa filings specifically, the [visa-filing pay range guide](https://www.refolk.ai/guides/pay-range-from-visa-filings) covers the extraction in more depth. When you apply, every application compares your number with a specific posting. Refolk tailors your resume to each posting and scores your fit, so the level you claim in negotiation matches the level your application shows.

## Frequently asked questions

### Which salary website is accurate?

None of them is accurate in general, because each one measures something different. The federal wage survey measures straight-time gross pay and is about a year old when it comes out. Visa filings record offered wages and legal floors. Crowd sites collect self-reported pay, which runs a little low and bunches toward the middle. Posted ranges are employer estimates. Accuracy depends on whether a figure matches your occupation, level, place, pay component, and date, so grade each figure on its own instead of ranking whole sites.

### Can I use H-1B salary data to benchmark an offer?

Yes, but only after splitting rows by PW_WAGE_LEVEL and reading the WAGE_RATE_OF_PAY_FROM column annualized with its unit. The prevailing wage is a floor set at the 17th, 34th, 50th or 67th percentile, and a certified filing is not a hire. One filing can also cover several positions. A median across all levels mostly tells you where the floors are. A level-matched offered wage for your occupation and worksite can be cited with those caveats written beside it.

### Is Glassdoor salary data base pay or total pay?

Glassdoor's pages headline total pay and split it into base and additional pay, but at least one third party says its estimates are base only, so the point is disputed. Don't settle it in general. Settle it for the figure in front of you. Write down which part you are citing, base or additional, and compare it only with the same part of your offer. If you can't tell which part a number is, the pay-component criterion fails and the figure gets discarded.

### How old is BLS OEWS wage data when I use it?

Older than it looks. Each annual release reports the May of the prior year and comes out about 10 to 11 months later. Each estimate also blends six half-year panels collected over three years, with the older panels brought forward using Employment Cost Index factors. The May 2024 estimates, for example, include panels going back to November 2021. Record both the reference date and the panel span, not just the release date.

### Should I aim for the midpoint of a posted salary range?

Not without evidence. In Glassdoor's matched study, more than 60% of salaries that fell inside a posted range sat below its midpoint, and 22% of listings had ranges above what employees actually reported. Ranges posted in states with transparency laws were no more accurate than others. Treat the posted range as the employer's estimate, check how wide it is, and only cite the midpoint after an adjustment that you write down.

### What do I do with a salary number a recruiter or friend told me?

Discard it by default, unless you can state all seven criteria for it. A number you heard usually has no date, no pay component, and no sample size. Ask follow-up questions: is it base or total cash, what level, which city, how recent, and how many people it covers. If the answers fill every criterion, grade it like any other figure. If they don't, keep it as context for later questions. Don't cite it.

---

*From the Refolk guide library. I revise these guides rather than replacing them, so the current version is always at https://www.refolk.ai/candidates/guides/salary-figure-cite-adjust-discard*
