# The Anchor-Grade Salary Figure Standard, and What Fails It

*Grade any single pay figure you found as pass or fail against fixed criteria, so a second reader would reach the same verdict on whether to rely on it.*

- Canonical URL: https://www.refolk.ai/candidates/guides/anchor-grade-salary-figure-standard
- Pillar: Reading the market
- Format: Standard
- Published: 2026-08-10
- Last reviewed: 2026-08-10
- Reading time: 16 min

## Key takeaways

- Glassdoor's own 2024 analysis found employer-posted ranges contain holder pay only 67% of the time, and for 22% of listings the actual pay fell below the posted range.
- Crowdsourced medians are biased in one direction, not randomly noisy: validated benchmarks put Glassdoor as much as 20,000 pounds below the true senior-level median.
- Title is the wrong key; level and tier are the join. The same Senior SWE title spans roughly 2x total comp, and medians for one title fanned out across a $135,000 gap over seven sources.
- Any single salary number above the roughly $250K-$300K FAANG base cap is almost certainly total comp, which makes the cap a free classifier for the base-versus-total question.
- BLS OEWS carries a roughly ten-month publication lag and blends six panels over three years, so it is a scope and floor reference, not a live market rate for a fast-moving role.
- In Refolk's index there are 2,948 current Compensation Analysts in the US versus 39 in the UK, a roughly 75x gap that signals how much thinner non-US pay data usually is.

You have a salary number in front of you - from a job board, a crowdsourced site, a government table, a friend of a friend - and you are about to build your expectations and your negotiation on it. This guide is for that moment. It gives you a fixed pass/fail rubric so you can grade that single figure as anchor-grade or not, and so a second person grading the same figure would reach the same verdict.

Most pay guides either decode what a number means or explain how HR builds bands. This one does something narrower and more useful mid-negotiation: it tells you whether the specific number you found is trustworthy enough to cite, before you cite it. The advice "use multiple sources" is true and useless without a rule for what makes any one source count.

## What "anchor-grade" means

Anchor-grade is a fixed bar: a figure is anchor-grade when it passes all five criteria below, and it is not anchor-grade when it fails any one of them. There is no partial credit, because a number you half-trust in a negotiation behaves like a number you fully trust.

An anchor-grade figure is one you would be willing to say out loud to a hiring manager and defend if challenged. Everything below the bar is still allowed to inform you - it just cannot be the number you plant your feet on.

The five criteria, in the order the procedure applies them:

| Criterion | Passes when | Fails when |
|---|---|---|
| Datable | You have a collection month and year | Only a page publish date exists, or none |
| Labeled | Base, total cash, or total comp is known | The figure is unlabeled and above cap |
| Level-matched | Your level and tier map to the sample | You matched on title alone |
| Disclosed | n and 25/50/75 percentiles are visible | A single number with no distribution |
| Cross-checked | A second method lands within tolerance | No second source, or an unexplained gap |

Read the criteria as gates, not a score. A figure with a perfect date, a clean label, and a matched level still fails if it is one unverified data point with no sample size behind it. The failure of any gate demotes the number from "anchor" to "directional".

> **Rule:** One failed gate demotes the figure
>
> The rubric is conjunctive. A figure must pass all five criteria to be anchor-grade. If it fails one, you may still use it as a directional reference, but you may not cite it as your anchor.

## Why one figure is not enough on its own

The core problem is that two sources can both be "accurate" and still disagree by six figures, because they sampled different populations. Accuracy within a source tells you nothing about whether that source answers your question.

Glassdoor's 2024 analysis found employer-posted ranges contain the reported pay of people holding those jobs about 67% of the time. That sounds reassuring until you read the other half: for 22% of listings, the pay reported by holders fell below the posted range, and reported salaries skew toward the lower half of the published band. So even an "accurate" range routinely sits above where people actually land.

**67% - How often posted ranges contain the reported pay of holders**

From Glassdoor's 2024 pay-range accuracy analysis. Accuracy is two-thirds, and the miss skews low.

The skew is not random noise. It is mechanical. On a wide band, anchoring on the midpoint means anchoring above where most holders actually are. That is why a "conservative-looking" crowdsourced figure can still be the wrong anchor - it may just be old, and old submissions bias one direction.

> A conservative-looking figure is often not conservative. It is stale, and staleness only points one way.

## The four source types, and what each is good for

Every salary figure comes from one of four collection methods, and each fails differently. Classify yours before you do anything else, because the failure mode you should worry about depends on the type.

| Source type | Collection | Verified? | Refresh and lag |
|---|---|---|---|
| Crowdsourced (Glassdoor) | Self-report | No | Blends multi-year submissions; skews stale and low |
| Employer survey (Mercer) | Participant payroll, 700k+ incumbents | Yes | Annual, aged forward with the ECI |
| Government (BLS OEWS) | Establishment survey, ~1.1M | Yes | Annual, ~10-month lag, 3-year panels |
| Job-posting aggregation | Scraped postings | Partial | Real-time postings, "good faith" ranges |

Crowdsourced data is self-reported and unverified, there is no job-level standardisation to ensure like-for-like comparisons, and the methodology behind the estimates is not disclosed. Its bias is directional: because old submissions are retained and equity refreshers are underreported, validated benchmarks have found Glassdoor figures sitting 20,000 pounds or more below the true market median at senior levels.

Employer surveys sit at the other end. Mercer's benchmark reported more than 700,000 incumbents globally, reports base, total cash, and total direct compensation, and ages data forward using the BLS Employment Cost Index. It is the strongest source for a level-matched market median, if you can access it.

Government OEWS is an establishment survey covering all full- and part-time wage and salary workers in nonfarm establishments. It excludes the self-employed, owners, household workers, and unpaid family workers, and it reports straight-time gross pay excluding overtime and nonproduction bonuses. It is precise and broad but slow, which matters for the next section.

#### How far a figure sits from the underlying reality

1. **The number you found** - A median or range on a page
2. **The sample behind it** - Who was surveyed, how many, when
3. **The definition** - Base, total cash, or total comp
4. **The population** - Which level, tier, and geography
5. **Actual pay** - What holders of your exact role earn now

*Each layer adds a way the number can drift from what someone is actually paid.*

## Why government data is precise but wrong for a live negotiation

BLS OEWS is the most rigorous source on this list and still the wrong anchor for a fast-moving role. It reads as authoritative, and that authority is exactly the trap.

The lag is structural. The process of collecting and compiling OEWS creates a roughly ten-month time-lag between the reference period and publication, the data is updated only once a year, and each release combines six panels collected over three years - the May 2022 estimates, for example, drew on responses from May 2022 back through November 2019. BLS itself states it does not encourage OEWS use for time-series comparison.

So OEWS answers "what is the broad base-pay floor and scope for this occupation" well, and answers "what will they pay me next month" badly. In roles where variable comp dominates - equity at Tier-1 tech, commission in sales - a BLS base median can understate real pay dramatically, because it captures none of the stock, bonus, or commission.

> **Watch out:** A government median is a floor, not a market rate
>
> Quoting a BLS base median in a tech or sales negotiation is a common self-inflicted wound. It excludes stock, bonuses, and commission, and it lags the market by nearly a year. Use it to sanity-check the low end, not to set your ask.

## Why title is the wrong key, and level is the right one

Two honest sources can disagree by six figures on the "same" job simply because a title spans multiple realities. The join key that makes a figure comparable is level and company tier, not the words in the title.

The spread is enormous. One recruiter analysis found medians for a single software-engineer title fanning out across a $135,000 gap over seven sources in the same year, with Glassdoor at $128,400 and ZipRecruiter at $110,140. By tier, the same "Senior Software Engineer" title spans roughly 2x total comp: at Google, Meta, or Amazon a Senior SWE (L5) earns $250K to $400K total comp, while at a mid-size enterprise or Series B startup the same title earns $150K to $220K. Glassdoor itself warns that user-reported salaries at the title level may over- or underestimate true pay when a title spans multiple seniority levels.

| Level | Tier-1 tech total comp | Non-tech / startup | Spread |
|---|---|---|---|
| L3 / entry | ~$170K-$190K | ~$95K-$110K base | ~1.8x |
| L5 / Senior SWE | $250K-$400K | $150K-$220K | ~2x |
| L6 / Staff | ~$457K median | thin or no stock | large |

The practical consequence for cross-checking: if you compare two sources without matching level, you manufacture false disagreement. The sources are not wrong; you asked them different questions. Match the level first, then the delta between two sources means something.

When you need level-matched comparison points and the public medians are too coarse, the fastest fix is to find the actual people at your target level and company tier. [Refolk](/candidates) writes your resume from your own history and tailors it per posting, and its people search lets you build a small level-matched panel instead of trusting a title-level median.

Ask me this: `Find senior software engineers at Google in the Seattle area to gather level-matched comparison points for an L5 offer.` - [run the search](https://www.refolk.ai/start?q=Find%20senior%20software%20engineers%20at%20Google%20in%20the%20Seattle%20area%20to%20gather%20level-matched%20comparison%20points%20for%20an%20L5%20offer.).

*Returns current people at a named level and tier so you can compare against a matched population rather than a title-level average.*

## The FAANG base cap is a free classifier

When a figure is unlabeled, one number does most of the classification work for you: base salaries in big tech cap out somewhere between $250,000 and $300,000, so any single number above that is almost certainly total comp. This is the cheapest test in the whole standard.

Base salary is the guaranteed cash portion. Total compensation adds equity, bonuses, retirement contributions, and benefits on top. The distinction is load-bearing because comparing base salary alone understates the compensation gap between tiers by 40% to 60%, and at Tier-1 tech equity can be 40% to 60% of total comp on its own.

Signals a figure is total comp: it exceeds the FAANG base cap, or the source is a tech-leveling aggregator. Signals it is base only: it matches a BLS figure, which captures no stock, bonuses, or other compensation. If the source is silent, run the cap test first, then look at the source type.

> **Tip:** Run the cap test before you trust any big number
>
> If an unlabeled figure sits above roughly $250K-$300K, treat it as total comp until proven otherwise. Reading a total-comp number as base is how candidates convince themselves an offer is a downgrade when it is not.

## Grade a single figure, step by step

This is the procedure. Each step is a few minutes; the whole grade takes about half an hour for a figure you care about. Work in order, because a figure that fails an early gate does not need the later ones.

#### The anchor-grade grading procedure

1. **Identify the source type** - Classify the figure as crowdsourced self-report, employer survey, government survey, or job-posting aggregation. Done when you can name the collection method in one sentence.
2. **Date the figure** - Find the collection or reference date, not the page's publish date. Done when you have a month and year; fail the figure if it is undatable.
3. **Check the base-versus-total label** - Confirm whether the number is base salary, total cash, or total comp. Done when the label is explicit or inferred from the FAANG base cap and source-type signals.
4. **Confirm the level and scope match** - Map your target level and job scope to the figure's population, not just the shared title. Done when you can state which level and company tier the number represents.
5. **Check disclosure depth** - Look for sample size, contributing-employer count, and the 25th/50th/75th percentiles. Done when you know n and the percentile spread, or you know they are missing.
6. **Cross-check against a second source type** - Compare the figure to at least one different method, such as a government median against a crowdsourced median at the same level. Done when you have a delta and can explain it.
7. **Grade pass or fail** - Apply the fixed thresholds - recent date, known label, matched level, disclosure present, cross-check within tolerance. Done when you have a written verdict a second reader would reproduce.

A note on tolerance in the cross-check step. There is no single universal number, because acceptable spread depends on how variable-comp-heavy the role is. As a working rule, two level-matched sources within about 15% of each other reinforce the anchor; a gap wider than that means you have not actually matched the populations, or one source is stale, and you must resolve why before you trust either.

Disclosure depth is where survey-grade data separates itself. Professional surveys report the 10th, 25th, 50th, 75th, and 90th percentiles, with the median and 75th most cited, and a 2023 SHRM survey found 87.6% of HR professionals use percentile data. They also disclose whether they report incumbent data (actual pay of current employees, which reflects tenure and merit) or hiring-rate data (pay offered to new hires, which reflects current external pressure). Crowdsourced confidence labels do not do this - they reflect sample volume, not the quality or consistency of what was submitted.

## How this goes wrong: the failure modes

The standard exists because these seven mistakes look like passes. Each one has a false positive - a signal that reads as reassuring but is not - so learn the false positive, not just the rule.

| Failure mode | The false positive that fools you | The check |
|---|---|---|
| Stale figure as current | A "very high confidence" label | Read the reference date, not the publish date |
| Base/total confusion | A number above the ~$250K-$300K cap read as base | Run the cap test and read the definition |
| Title without level | Matching "Senior" to a FAANG L5 | Match level and tier, not the title |
| Wide "good faith" range | Anchoring on the top of a $50K-$150K band | Check width against NJ's 60%-of-minimum bar |
| Government median as market rate | A BLS base median in a tech ask | Check whether variable comp dominates |
| High-variable-comp roles | A sales "total pay" range read as guaranteed | Check the fixed-versus-variable split |
| No sample size | A single number treated as a distribution | Require n and the 25/50/75 percentiles |

Two of these deserve extra weight because they are the most common.

The stale figure is dangerous precisely because crowdsourced platforms label confidence by volume. A role can carry hundreds of years-old submissions and show "high confidence" while trailing the real median by tens of thousands. The fix is boring and reliable: find the collection date. If a page will not tell you when the data was collected, the number fails the date gate regardless of how confident it looks.

The wide range that is not meaningful is the trap built into job postings. A $50,000 spread on a $100,000 job carries almost no information about your likely offer. New Jersey caps posted range width at 60% of the minimum salary, which is a useful external sanity bar - if a posted range is wider than that relative to its floor, treat it as a legal placeholder, not a signal, and expect to land in the lower half.

> **Watch out:** Confidence labels measure volume, not truth
>
> A crowdsourced "high confidence" tag tells you many people submitted, not that the submissions are recent, verified, or level-matched. A large sample of stale, mixed-level self-reports can be confidently wrong.

## When the data infrastructure itself is thin

Outside a few large markets, the pool of pay data is dramatically thinner, which means every figure deserves a stricter disclosure bar. This is easy to miss because a page in a small market looks identical to a page in a large one.

Use the depth of the local compensation profession as a proxy for how much data infrastructure exists to feed those pages. In Refolk's index of professional profiles, there are 2,948 current Compensation Analyst professionals in the United States versus 39 in the United Kingdom.

| Market | Current "Compensation Analyst" professionals | Share vs US |
|---|---|---|
| United States | 2,948 | 1.00 |
| United Kingdom | 39 | 0.013 (~75x fewer) |

A roughly 75x gap is a warning sign. It does not mean UK figures are worthless; it means a crowdsourced UK median is far more likely to rest on a handful of self-reports, so the disclosure gate - sample size and percentiles - should be enforced strictly rather than waived. For any market outside the US, raise your bar on the disclosure and cross-check steps, and lean harder on employer-survey and government sources where the sample is documented.

**~75x - More current Compensation Analysts in the US than the UK, in Refolk's index**

A proxy for how much thinner pay-data infrastructure is outside the largest markets. Non-US figures warrant stricter scrutiny.

## Verify before you cite

Run this checklist on the figure you are about to anchor on. If every box is checked, the number is anchor-grade and you can cite it and defend it. If any box is empty, demote it to a directional reference and find a figure that passes.

#### Before you cite this figure in a negotiation

- [ ] I can name the collection method (crowdsourced, employer survey, government, or posting).
- [ ] I have a collection month and year, not just a page publish date.
- [ ] I know whether the figure is base, total cash, or total comp.
- [ ] The figure sits below or above the FAANG base cap consistent with its stated label.
- [ ] I have mapped my level and company tier to the figure's population, not just the title.
- [ ] I can see the sample size and the 25/50/75 percentiles, or I have noted they are missing.
- [ ] I have a second source of a different method, matched on level, and can explain the delta.
- [ ] For a non-US market, I enforced the disclosure gate strictly rather than waiving it.
- [ ] I have written a one-line verdict a second reader could reproduce.

## Keeping the standard current

The rubric is stable; the inputs are not. The one number in this guide most likely to drift is the FAANG base cap, currently around $250,000 to $300,000. Re-check it against a couple of current big-tech levelling sources before you lean on it as a classifier, because if base caps rise, the "above cap means total comp" rule shifts with them.

The rest of the standard is mechanism, not value, and does not go out of date. Source types fail the same way regardless of the year: crowdsourced skews stale and low, government lags and excludes variable pay, employer surveys are strong but gated, postings publish wide good-faith ranges. Titles will keep spanning multiple levels, and the fix will keep being to match level and tier rather than words.

When you adopt this as a personal or team policy, write the verdict down every time. The value of a standard is that two people grade the same figure the same way, and that only holds if the grade is recorded in a form a second reader can check. A one-line note - source type, date, label, level, n, cross-check delta, pass or fail - is enough. That record is also what lets you defend the number if a hiring manager pushes back, which is the entire point of anchoring on a figure you graded first.

## Frequently asked questions

### Is this salary data reliable enough to negotiate on?

A single figure is reliable enough to anchor on only if it passes five fixed tests: it has a collection date within about eighteen months, a known base-versus-total label, a level and tier that match your role, disclosed sample size and percentiles, and a cross-check against a second source type that lands within tolerance. Miss any one and demote it to a directional reference rather than an anchor. That is the whole standard, and any two people running it should reach the same verdict.

### How accurate are Glassdoor salaries?

Glassdoor's own 2024 analysis found employer-posted ranges contain reported holder pay about 67% of the time, and for 22% of listings the actual pay fell below the posted range. Its crowdsourced medians are self-reported, unverified, and blend multi-year submissions, which biases them low and stale. Validated benchmarks have placed Glassdoor figures 20,000 pounds or more below the true median at senior levels. Treat it as a low-side directional signal, not an anchor.

### Are posted salary ranges accurate?

Posted ranges are accurate roughly two-thirds of the time by Glassdoor's measure, but accuracy is not the whole problem. Reported pay skews to the lower half of the published band, so anchoring on the midpoint of a wide range statistically aims high. Very wide 'good faith' ranges carry little information; New Jersey caps posted width at 60% of the minimum as a legal sanity bar. Expect to land below the middle.

### Which salary source should I trust most?

There is no single winner; each source is trustworthy for a different question. Employer surveys like Mercer, drawing on 700,000-plus incumbents, are strongest for level-matched market medians. Government OEWS is a reliable base-pay floor and scope reference but lags about ten months. Crowdsourced total-comp aggregators are useful for tech equity signals if the sample is disclosed. Match the source to your question, then cross-check across two methods at the same level.

### How do I check if a salary number is outdated?

Find the collection or reference date rather than the page's publish date, because a page can be republished while the underlying submissions are years old. Crowdsourced platforms retain old entries and rarely down-weight them, and government OEWS blends six panels over three years with a ten-month publication lag. If you cannot pin the figure to a month and year, fail it and find one you can date.

---

*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/anchor-grade-salary-figure-standard*